Visible light positioning deep learning fusion system based on multiple models

Abstract

How to reduce the error of visible light positioning is the technical problem focused in the current positioning technology, and reduction of the positioning error by adopting the neural network technology is an important scheme for solving the problem. The invention provides a deep learning fusion system for visible light positioning and a method thereof, and aims to solve the problems of unstable model training effect, insufficient positioning precision and poor generalization ability caused by insufficient data acquisition amount in the prior art, or the problems of long training time caused by overlarge data scale and the like. According to the method, the corresponding small model neural network or large model neural network is selected for training and prediction according to the scale of the collected data volume in different operation stages of the system, and a dynamic judgment and periodical comparison mechanism is combined, so that quick and stable positioning output is realized when the data is limited; when the data is sufficient, a high-precision and high-robustness positioning effect is realized, so that the adaptability and reliability of the system in practical application are improved.

Description

Technical Field

[0001] This invention originates from the needs of indoor positioning applications and relates to the field of visible light positioning technology, particularly to a method for constructing and positioning a multi-model-based visible light positioning deep learning fusion system. Background Technology

[0002] Visible Light Positioning (VLP) is a technology for determining spatial location based on the characteristics of wireless optical transmission. In recent years, it has received widespread attention in scenarios such as intelligent transportation, indoor navigation, and the Internet of Things (IoT). The basic principle of this technology is to utilize the visible light signal emitted by a light source, measure the intensity or angle of the light signal at the receiving end, and deduce the spatial relationship between the transmitting and receiving ends based on a specific algorithm. This process typically relies on the Lambertian distribution model to describe the radiation characteristics of the light source. However, in practical applications, visible light signals are easily affected by environmental interference and noise during propagation. The mapping relationship between the light signal and spatial location exhibits high nonlinearity and complexity, making it difficult for traditional analytical methods to accurately model. To overcome this problem, in recent years, deep learning methods, represented by neural networks, have been gradually introduced. By automatically learning complex input-output mapping relationships from large amounts of data, they aim to improve positioning accuracy and robustness.

[0003] Patent document CN115983325 A proposes a high-speed visible light positioning method based on convolutional neural networks (CNN). It uses a CMOS image sensor to capture LED light signal images and trains and classifies them using a batch normalized convolutional neural network (BN-CNN) to identify LED markings and determine the location of the receiving end. This method has high recognition speed and accuracy under high-speed motion.

[0004] Patent document CN115983325 A proposes a localization method based on graph neural networks. Guided by an attention mechanism, it simultaneously considers the topological relationships of multiple location lights and combines LSTM modeling of trajectory time dependencies, further improving localization accuracy and robustness. However, such large-scale models have a massive parameter scale and are highly dependent on training data, making them prone to overfitting when data is insufficient, resulting in unstable localization results. Furthermore, the training and inference resource costs are high, making them difficult to apply in the early stages when data is scarce or hardware is limited. Methods based on simple neural networks...

[0005] Patent document CN 113219408 A proposes an improved RBF neural network indoor positioning method. It determines the hidden layer structure through KPCA-K-means++ clustering and optimizes the parameters using GA-LMS, thereby improving performance and generalization ability.

[0006] Patent document CN 109615056 A describes a localization method based on particle swarm optimization extreme learning machine (PSO-ELM). This method utilizes the particle swarm algorithm to optimize input weights and thresholds, solving the instability problem caused by random initialization. The model is simpler, trains faster, and has higher accuracy.

[0007] Large-scale models have a massive parameter set and are highly dependent on training data. They are prone to overfitting and unstable localization results when data is insufficient. Furthermore, they incur high training and inference resource costs, making them difficult to apply when data is scarce or hardware is limited in the early stages. Small-scale models have a relatively simple neural network structure and limited expressive power, making it difficult to fully characterize nonlinear relationships in complex environments. They also rely primarily on RSSI features, making them susceptible to multipath effects and noise interference, resulting in poor stability when data is insufficient or the channel is complex. In real-world scenarios, RSS data acquisition is often incremental, meaning that only a small amount of data is initially obtained, with data gradually accumulating during subsequent operation. Existing methods often select a single model for training and use from the outset, failing to address the needs of different data scales. Specifically, using a simple model may allow initial training, but the accuracy will be insufficient; using a complex model, insufficient initial data will prevent stable training.

[0008] Therefore, a prominent contradiction exists in existing technologies: how to rationally select neural network models in real-world scenarios where data volume is gradually increasing, so that the system can maintain basic positioning capabilities when data is limited, and fully leverage the advantages of complex models as data becomes richer, thereby achieving high-precision, low-cost, and scalable visible light positioning. This invention proposes a visible light positioning method that selects a neural network model based on data volume. This method dynamically adopts an adapted neural network structure according to changes in the scale of collected data, thus balancing practicality and accuracy throughout the entire lifecycle of the positioning system. Summary of the Invention

[0009] The purpose of this invention is to provide a deep learning fusion system and method for visible light positioning, addressing the problems of unstable model training, insufficient positioning accuracy, and poor generalization ability caused by insufficient data acquisition, or long training times due to excessively large data scales in existing technologies. This invention selects appropriate small-model or large-model neural networks for training and prediction at different stages of system operation based on the scale of acquired data, and combines dynamic discrimination and periodic comparison mechanisms to ensure fast and stable positioning output when data is limited, and high-precision and robust positioning effects when data is abundant, thereby improving the system's adaptability and reliability in practical applications.

[0010] To achieve the above objectives, the basic idea of ​​this invention is as follows: Addressing the phased nature of data acquisition and the gradual accumulation of data volume over time in visible light positioning systems, a deep learning model selection mechanism based on data volume dependence is designed. In the initial stage with limited data, a small-scale neural network (such as a BP network) with simple structure and fast training convergence is used for training and prediction to ensure positioning stability and convergence efficiency under limited data conditions. As the data volume gradually increases and reaches a preset threshold, a large-scale neural network with more complex structure and stronger expressive power (such as a Transformer network) is used for training and prediction to fully explore the potential feature relationships in the large-scale data, thereby significantly improving positioning accuracy and generalization performance. To ensure real-time performance and stability during operation, this invention introduces a sub-model data discrimination system: by comparing the prediction mean and variance of the large and small models at the same prediction point, the model with the better output result is dynamically selected. Simultaneously, a fixed time interval is set during system operation to periodically compare and switch between the large and small models, ensuring that the optimal-performing model is always used at different stages. Thus, this invention achieves a dynamic transition from "fast and stable performance under small data" to "high-precision and robust performance under large data," overcoming the limitations of traditional single-model methods while ensuring the system's adaptability and practicality in different application scenarios.

[0011] The specific steps of this invention are as follows:

[0012] Step 1: Arrange N (N≥3) light-emitting diodes (LEDs) (101) as beacons on an indoor plane. The target terminal is equipped with a single photodetector (PD) (102). At a preset reference point u, the visible light signals from each LED are received sequentially, the RSS measurement value at that location is obtained, and the actual position coordinates (x, y, y) of the reference point are recorded simultaneously. u y u ), thereby constructing an RSS fingerprint database as input data and labels for the neural network

[1112] Fingerprint database: After all reference points are collected for the kth time, U-row data is formed:

[0013] Where RSS(u, n) represents the RSS value of the nth LED measured at reference point u;

[0014] Step Two: The data collection adopts a phased cumulative approach, dividing the collection process into two phases based on the number and scale of data collection:

[0015] Based on the number and scale of data collection, the data collection process is divided into two stages:

[0016] Phase 1 (Initial Acquisition Phase): Within the time interval [T0, T1], the system completes m1 rounds of acquisition, that is, repeats the measurement for all U reference points m1 times, obtaining m1×U rows of sample data. Since the sample size is limited at this time, only a small dataset can be formed, which is suitable as a training set for small-scale neural networks;

[0017] The second stage (data expansion stage): Within the time interval [T0, T2], reference point data continues to be collected, with a total of m2 collections (m2 > m1). At this point, the total number of samples in the RSS fingerprint database increases significantly, reaching m2 × U rows, which is sufficient to support the training needs of large-scale neural networks. Therefore, the number of collections directly determines the size of the dataset; small models can be trained with fewer collections, while large models require more collections and a larger scale of data.

[0018] Step 3: Small model training:

[0019] On the dataset collected in the initial stage, a BP neural network was used as a small model for training.

[0020] The number of input layer nodes is N s ;

[0021] The input is R s =[RSS(u,1),RSS(u,2),...,RSS(u,N)];

[0022] Received signal strength (RSS) under conditions of a small number of samples;

[0023] The number of hidden layer nodes is M s The output layer has 2 nodes, corresponding to the position coordinates (x, y) of the photodetector;

[0024] The connection from the input layer to the hidden layer has a weight matrix w s and bias vector b1 s The connection from the hidden layer to the output layer has a weight matrix v s and bias vector b2 s ;

[0025] Both the hidden layer and the output layer use the non-linear activation function g1. s g2 s : Hidden layer activation function g1 s The Sigmoid function is used, and its form is: Output layer activation function g2 s A linear function is used, in the form of: g2 s (z)=z

[0026] During training, mean squared error (MSE) is used as the loss function.

[0027] When the amount of collected data is small, a small model neural network can be considered to have completed training if it meets one of the following conditions:

[0028] When the loss decreases less than the threshold δ for several consecutive iterations s ,Right now |L (t) -L (t-1) |<δ s , Then the training is considered complete;

[0029] The positioning error meets the requirement: the average positioning error on the validation set. Less than the preset threshold ∈ s When the training is complete, it is considered finished.

[0030] This model is suitable for use when the number of data collections is limited and the amount of data is small. It can converge quickly and output stable prediction results.

[0031] Step 4: Large Model Training

[0032] During the data expansion phase, when the number of data collections reaches m² times and the dataset size reaches a set threshold U... th Then, the Transformer neural network was used as the large model for training;

[0033] The number of input layer nodes is N t ;

[0034] The input is R t = [RSS(u, 1), RSS(u, 2), ..., RSS(u, N)], which corresponds to the received signal strength (RSS) under large-scale sample conditions and is expressed as a sequence;

[0035] The output layer has 2 nodes, corresponding to the position coordinates (x, y) of the photodetector;

[0036] The connection from the input layer to the coding layer has a parameter matrix w t and bias b1 t The connection from the encoding layer to the output layer has a parameter matrix v t and bias b2 t ;

[0037] Both the encoding and output layers use the activation function g1. t g2 t : g1 t Using GELU (Gaussian Error Linear Unit), for any scalar input u: g2 t Here W o h+b o ∈R 2 h is the converged representation of the coding layer output. g2 t (u)=u

[0038] The Transformer neural network enhances the spatial information of the RSS sequence through position encoding, calculates a weighted representation using multi-head attention, and outputs predicted coordinates after mapping through a feedforward network.

[0039] During training, mean squared error (MSE) is used as the loss function.

[0040] As the amount of collected data gradually increases and reaches a large scale, large-scale neural networks must simultaneously meet the following conditions to be considered to have completed training:

[0041] When the loss decreases less than the threshold δ for several consecutive iterations t ,Right now |L (t) -L (t-1) |<δ t , Then it is considered convergent;

[0042] High precision requirement: Average positioning error on the validation set Less than the strict threshold ∈ t When training is complete, ∈ t <∈ s

[0043] Step 5: Runtime Prediction and Output

[0044] During actual positioning, the system receives the RSS data of the point to be located and uses it as input to the currently selected small or large model to obtain the predicted coordinates (x, y). The prediction results are provided by the data output module and can be directly used in indoor navigation, intelligent transportation, or IoT application scenarios.

[0045] Step Six: Sub-model discrimination and selection (based on comparison of multiple predictions)

[0046] To ensure the stability and accuracy of the output results, this invention designs a sub-model data discrimination mechanism. At the same prediction point, small-model neural networks are used respectively. Large model neural networks Multiple predictions are made, resulting in two sets of prediction outcomes. By calculating the mean and variance of each set, the system selects the optimal model based on the following principles:

[0047] Prediction result set:

[0048] Prediction result set: {(x t i, y t，i | i = 1, 2, ..., K}

[0049] Mean:

[0050] variance:

[0051] Mean:

[0052] variance:

[0053] Model selection is based on the comparison results of the mean and variance:

[0054] When μ t =μ s hour

[0055] like This indicates that the larger model exhibits less fluctuation and more stable output at the same prediction location, demonstrating stronger fitting and generalization capabilities. Therefore, the larger model is deemed superior. The results from the larger model are selected.

[0056] like Although the predicted mean was consistent, the larger model exhibited greater fluctuations, indicating that it might have included irrelevant samples or overfitting in the training data, leading to unstable prediction results; therefore, the smaller model was deemed superior. The results from the smaller model were selected.

[0057] When μ t ≠μ s hour:

[0058] like The larger model, while outputting different means, has smaller variance, indicating that it can capture potential data patterns that the smaller model failed to learn, and provides new prediction results with greater stability. Therefore, the larger model is determined to be superior. The results of the larger model are selected.

[0059] like The large and small models predict different mean values, with the large model exhibiting greater fluctuations, indicating that it has not yet converged sufficiently or is undertrained. Therefore, the results from the small model should be used directly.

[0060] Once both the small and large models have completed training, the system compares the prediction results of the two models at fixed time intervals Δt (e.g., 15 minutes) during operation. During the time interval between two adjacent comparisons, the system continues to output the previously selected model until the next comparison is completed and the model selection is updated.

[0061] Step 7; Model Usage Records and Statistics

[0062] During system operation, small model neural networks are recorded. and large model neural networks The usage time, which is represented by the following vector A. T Indicates: A T = [a1, a2, ... a n , ...a N ]; where N represents the number of time periods in the statistics, a n The value is either "0" or "1", where "0" indicates that a small model is currently being used, and "1" indicates that a large model is currently being used. According to vector A... T Records, statistical small model neural network Usage time T s =m1ΔT, utilization rate μ s =T s / (T t +T s ), Statistical Large Model Neural Network Usage time T t =m2ΔT, utilization rate μ t =T t / (T t +T s ), where m1 is A T The number of "0"s in m2 is A T The number of "1"s in the text.

[0063] Step 8: Calculation and Dynamic Adjustment of Computational Load:

[0064] The total computational cost of one forward propagation is C S =O(N)s ·N s +2·M s )≈0(M s ·N s )

[0065] The total computational cost of one forward propagation is

[0066] The total computational load of the entire visible light positioning deep learning fusion system within the statistical period was: Total = C s ×μ s +C t ×μ t

[0067] The system statistically analyzes the usage and computational load of small and large models over different time periods, forming a historical database to reflect the computational overhead characteristics under different data scales and switching strategies. Based on historical trends, the parameters T1 and T2 in claim 2 are dynamically adjusted. 1) When the large model usage rate μ t1 Exceeding rated utilization μ t0 That is, μ t1 >μ t0 Using ΔT1 as the time step, the setting time of parameter T1 is extended by T1+iΔT1, where T1+iΔT1<T2, i=1,2,3... This method improves the training effect of the small model and increases the utilization rate of the small model in actual use. 2) When At that time, with ΔT2 as the time step, the setting time of parameter T2 is extended by T2+jΔT2, j=1,2,3..., in order to improve the training effect of large models;

[0068] Step 8: Final Output

[0069] The final output selected by the sub-model data discrimination system is used as the localization output of the deep learning fusion system and transmitted to the application through the data output system. Attached Figure Description

[0070] Figure 1 Schematic diagram of the overall structure of a deep learning fusion system for visible light positioning

[0071] Figure 2 Data Acquisition and Fingerprint Database Construction System

[0072] Figure 3 Sub-model data discrimination system Detailed Implementation

[0073] To better understand the technical solution of this invention, the following flowchart is attached. Figure 1 The present invention will be described in detail with reference to specific embodiments.

[0074] Step 1: The experimental area is set up as an indoor room of 6m×6m×3m; five LED lights are placed at the four corners and center of the ceiling as beacons, each with a transmission power of 1W and an operating wavelength of 450-650nm; the target terminal is equipped with a photodetector (PD) with a detection area of ​​1cm². 2 The field of view is 60° and the installation height is 0.8m. 100 reference points are collected in the experimental area using a 0.5m×0.5m grid. The RSS signal of each reference point is collected 50 times and the average value is taken as the RSS fingerprint of that point. The RSS unit is dBm.

[0075] Step 2: In the initial sampling phase T1 = 10 samplings (approximately 5000 samples), with 100 reference points. Each point is sampled 50 times and the average is taken. Therefore, the sample size for each sampling is approximately 500 samples for training, and each sampling time is 10 minutes.

[0076] The collected RSS values ​​are first normalized, and outliers are removed; Input layer: 5 nodes, corresponding to the RSS of 5 LEDs respectively; Hidden layer: 1 layer, with a total of 20 neurons; Output layer: two nodes;

[0077] Parameter settings: Weight initialization method: Xavier initialization; Hidden layer activation function: Sigmoid; Output layer activation function: Linear function; Learning rate: 0.01; Optimizer: Adam;

[0078] Training details: Batch size: 32; Maximum number of iterations: 1000; Validation set: divided into 20% of the sampled data;

[0079] Training completion criteria: The loss decreases by less than 10% over 20 consecutive iterations. -4 Training is considered complete when either condition is met (e.g., convergence is achieved, or the average positioning error of the validation set is less than 0.5m).

[0080] Step 3: When the amount of collected data increases to T2 = 100 collections (approximately 50,000 samples), the large-scale model training phase begins. At this point, each collection session still takes 10 minutes. The large-scale model uses a Transformer neural network.

[0081] The collected RSS values ​​are first normalized, and outliers are removed; the RSS sequence is represented as a vector of length 5 (corresponding to 5 LEDs) as the input sequence;

[0082] Input layer: sequence length is 5, feature dimension d = 64;

[0083] Input layer overlay position encoding, using a sine / cosine function form:

[0084] Encoding layers: There are 2 layers in total, each layer contains:

[0085] Multi-head self-attention mechanism, number of heads h=4, attention calculation formula is: Where Q, K, and V are the input query, key, and value vectors, respectively;

[0086] The feedforward fully connected network consists of two linear transformation layers with GELU activation in between.

[0087] Residual connectivity and layer normalization (LayerNorm);

[0088] Output layer: Fully connected layer, compressing the feature dimension from 64 to 2; Optimizer: Adam, learning rate set to 0.001; Activation function: Identifier mapping; Batch size: 64; Maximum number of iterations: 5000; Training set to validation set ratio: 8:2;

[0089] Criteria for determining whether large model training is complete: When the loss decreases by less than 10% over 20 consecutive iterations. -5 Training is considered complete when the network converges and the average localization error on the validation set is less than 0.2m.

[0090] During the positioning and operation phase, in this embodiment of the invention, the small model and the large model are called 10 times each at the same prediction point to obtain two sets of prediction results.

[0091] If the difference in the means of the two sets of predictions is less than 5m, the means are considered consistent; if the difference in variance is greater than 0.01m... 2 If the variance is significantly different, then it is determined that there is a significant difference.

[0092] With consistent means and smaller variance in larger models, output the results of larger models;

[0093] The means are the same, but the larger model has a larger variance; output the results of the smaller model.

[0094] The means are inconsistent, but the larger model has a smaller variance, so the output is the result of the larger model.

[0095] The means are inconsistent and the larger model has a larger variance, so the output is the result of the smaller model.

[0096] Step 4: During the experiment, the system performs model discrimination every 15 minutes and records the result as a binary vector A.T = [a1, a2, ... a n , ...a N ];

[0097] When the discrimination result is a small model, the system automatically records "0" in the database;

[0098] When the discrimination result is a large model, record it as "1";

[0099] Within a 24-hour statistical period, the system performed 96 discriminations (Δt = 15 min).

[0100] The small model was selected 62 times (a i =0); the large model was selected 34 times (a i =1);

[0101] Small model usage time: T s =62×15min=930min≈15.5h;

[0102] Duration of large model usage: T t =34×15min=510min≈8.5h;

[0103] Small model utilization rate: r1 = 62 / 96 ≈ 64.6%; Large model utilization rate: r2 = 34 / 96 ≈ 35.4%.

[0104] Step 5: The computational cost of a single forward propagation for the small model is approximately 10. 3 Multiple multiplications and additions; the computational cost of a single forward propagation for a large model is approximately 10. 5 Multiplication and addition;

[0105] The total computational load of the entire visible light positioning deep learning fusion system within the statistical period was: Total = C s ×μ s +C t ×μ t Here is the rated utilization rate of the large model μ t0 Set to 35%.

[0106] When the system detects that, based on historical trends, parameters T1 and T2 in claim 2 are dynamically adjusted: 1) When the large model usage rate μ t1 Exceeding rated utilization μ t0 That is, μ t1 >μ t0Using ΔT1 as the time step, the setting time of parameter T1 is extended by T1+iΔT1, where T1+iΔT1<T2, i=1,2,3... This method improves the training effect of the small model and increases the utilization rate of the small model in actual use. 2) When When ΔT2 is used as the time step, the setting time of parameter T2 is extended by T2+jΔT2, j=1,2,3..., in this way to improve the training effect of large models; here m is the index, which can be 3-5.

[0107] Based on the statistical results of the historical database, the system performs rolling parameter optimization in three dimensions: daily, weekly, and monthly, to ensure a balance between positioning accuracy and computational load.

[0108] Step 5: The prediction results determined by the sub-model discrimination system are used as the system's positioning output and transmitted to the application end through the data output module.

Claims

1. A deep learning fusion system for visible light positioning, characterized in that, The deep learning fusion system includes: A deep learning model data acquisition system (11) comprises the following components: N (N≥3) light-emitting diodes (LEDs) (101) are arranged on an indoor plane as beacons; a target terminal is equipped with a single photodetector (PD) (102); visible light signals from each LED are received sequentially at a preset reference point u; the RSS measurement value at that location is obtained; and the actual position coordinates (x, y, y) of the reference point are recorded simultaneously. u y u This is used to construct an RSS fingerprint database, which serves as both input data and a label fingerprint database for the neural network. After all reference points are collected in the k-th iteration, a U-shaped data set is formed. Where RSS(u, n) represents the RSS value of the nth LED measured at reference point u; Small model neural network This is a sub-model of deep learning, with M1 model parameters. The computational complexity (C) required for one forward propagation is C. s ; Large model neural networks This belongs to another deep learning sub-model, with M2 parameters, where M2 > M1. The computational complexity (C) required for one forward propagation of this model is C. t C t >C s ; The sub-model data discrimination system (14) is responsible for selecting the small model neural network. and large model neural networks The output result; The deep learning fusion system data output system (15) outputs the position coordinates predicted according to the model.

2. As described in claim 1: Data collection employs a phased, cumulative approach, involving multiple data collections across different time intervals during system operation. This results in datasets of varying sizes at each stage. Training methods include the following: 1) Collect data from the reference point m1 times within the time interval [T0, t1] to construct a small model neural network. The dataset required for training; 2) A total of m2 data acquisitions are completed within the time interval [T0, T2], forming a large-scale neural network model. The training dataset satisfies T2>T1 and m2>m1; 3) and All are trained and predicted based on the data in this fingerprint database. The input format may vary depending on the network structure, but all use the same RSS data as the source. 4) When the amount of data collected is in the initial stage, for Training will be conducted; as the amount of collected data gradually increases and reaches the preset threshold U... th At that time, Conduct training.

3. The system as described in claim 1, characterized in that, The system operation method includes the following steps: 1) The output of the system is based on the output of the currently selected sub-model; 2) The sub-model data discrimination system (14) selects the sub-model output according to the following method: At the same prediction point, for and Multiple predictions are made separately, and the prediction result of each network at this point is given as: Small model prediction result set: {(x s,i ,y s，i )|i=1,2,...,K} Large model prediction result set: {(x t,i ,y t，i )|i=1,2,...,K} Its predicted mean and variance are defined as follows: Small model mean: Small model variance: Large model mean: Large model variance: Model selection is based on the comparison results of the mean and variance: 1) If This indicates that the larger model exhibits less fluctuation and more stable output at the same prediction position, demonstrating stronger fitting and generalization capabilities. Therefore, the larger model is deemed superior; in this case, the larger model neural network is selected. The output result; 2) If Although the predicted mean is consistent, the larger model exhibits greater fluctuations, indicating that it may have included irrelevant samples or overfitting in the training data, leading to unstable prediction results. Therefore, the smaller model is deemed superior; thus, the smaller neural network model is selected. The output result; 3) When μ t ≠μ s Time: If The larger model, while outputting different means, has a smaller variance, indicating that it can capture potential data patterns that the smaller model failed to learn, and provides new predictions with greater stability. Therefore, the larger model is deemed superior; in this case, a large-model neural network is chosen. The output result; 4) When μ t ≠μ s At that time, if If the large and small models predict different mean values, and the large model exhibits greater fluctuations, it indicates that the large model has not yet converged sufficiently or is undertrained. In this case, the small model should be considered directly, and a small model neural network should be selected. The output result.

4. The method according to claim 1 or 2, A backpropagation (BP) neural network is selected, consisting of an input layer, hidden layers, and an output layer: 1) The number of nodes in the input layer is N s The input is R s = [RSS(u, 1), RSS(u, 2), ..., RSS(u, N)], corresponding to the received signal strength (RSS) under a small number of sample conditions; 2) The number of hidden layer nodes is M s The output layer has 2 nodes, corresponding to the position coordinates (x, y) of the photodetector; 3) The connection from the input layer to the hidden layer has a weight matrix w. s and bias vector b1 s The connection from the hidden layer to the output layer has a weight matrix v s and bias vector b2 s ; 4) Both the hidden layer and the output layer use the non-linear activation function g1. s g2 s : Hidden layer activation function g1 s The Sigmoid function is used, and its form is: Output layer activation function g2 s A linear function is used, in the form of: g2 s (z)=z 5) During training, mean squared error (MSE) is used as the loss function. The criteria for determining whether training is complete are: When the amount of collected data is small, a small model neural network can be considered to have completed training if it meets one of the following conditions: When the loss decreases less than the threshold δ for several consecutive iterations s ,Right now |L (t) -L (t-1) |<d s , Then the training is considered complete; The positioning error meets the requirement: the average positioning error on the validation set. Less than the preset threshold ∈ s When the training is complete, it is considered finished.

5. The method according to claim 1 or 2, The Transformer neural network is selected, consisting of an input layer, an encoding layer, and an output layer: 1) The number of nodes in the input layer is N t The input is R t = [RSS(u,1), RSS(u,2), ..., RSS(u,N)], which corresponds to the received signal strength (RSS) under large-scale sample conditions and is expressed as a sequence; 2) The coding layer consists of a multi-head self-attention mechanism and a feedforward network. The multi-head self-attention mechanism is used to capture the dependencies between each RSS signal, and the feedforward network is used for nonlinear feature transformation. 3) The output layer has 2 nodes, corresponding to the position coordinates (x, y) of the photodetector; 4) The connection from the input layer to the coding layer has a parameter matrix w t and bias b1 t The connection from the encoding layer to the output layer has a parameter matrix v t and bias b2 t ; 5) Both the encoding and output layers use the activation function g1. t g2 t : g1 t Using GELU (Gaussian Error Linear Unit), for any scalar input u: g2 t Here W o h+b o ∈R 2 h is the converged representation of the coding layer output. g2 t (u)=u; 6) During training, mean squared error (MSE) is used as the loss function. 7) The Transformer neural network enhances the spatial information of the RSS sequence through position encoding, calculates the weighted representation using multi-head attention, and outputs the predicted coordinates after mapping through a feedforward network; Sufficient data: The number of training samples U is not less than the set threshold U th ,Right now U≥U th ; Criteria for determining training completion; As the amount of collected data gradually increases and reaches a large scale, large-scale neural networks must simultaneously meet the following conditions to be considered to have completed training: When the loss decreases less than the threshold δ for several consecutive iterations t ,Right now |L (t) -L (t-1) |<d t , Then it is considered convergent; High precision requirement: Average positioning error on the validation set Less than the strict threshold ∈ t When training is complete, ∈ t <∈ s .

6. The method according to claim 1 or 2, characterized in that: when (12) and (13) After training is completed and model switching is performed, the system compares the prediction results of the small and large models at a preset time interval Δt, so as to dynamically determine whether to use the small model neural network or the large model neural network during the operation of the system. In the time interval Δt between two adjacent comparisons, the system continues to use the small model neural network or the large model neural network selected in the previous determination for positioning output until the next comparison is completed and the new model selection result is determined.

7. The system as described in claim 3, characterized in that: When the system is running, it records the small model neural network N. s (12) and large model neural networks (13) The usage time, which uses the following vector A T Indicates: A T = [a1, a2, ... a n , ...a N ]; where N represents the number of time periods in the statistics, a n The value is "0" or "1", where "0" indicates that a small model is currently being used and "1" indicates that a large model is currently being used.

8. The system as described in claim 7, characterized in that: According to vector A T The record, statistical small model neural network N s (12) Usage time T s =m1ΔT, utilization rate μ s =T s / (T t +T s ), Statistical Large Model Neural Network Usage time T1 = m2ΔT, utilization rate μ t =T t / (T t +T s ), where m1 is A T The number of "0"s in m2 is A T The number of "1"s in the text, ΔT is the value of a statistical time period as described in claim 7.

9. The system as described in claim 8: characterized in that: The total computational load of the entire visible light positioning deep learning fusion system within the statistical period is: Total=C s ×μ s +C t ×μ t in, The total computational cost of one forward propagation is: C S =O(N s ·N s +2·M s )≈O(M s ·N s ), The total computational cost of one forward propagation is:

10. The system according to claim 8, characterized in that: Based on historical trends, the parameters T1 and T2 in claim 2 will be dynamically adjusted: 1) When the large model usage rate μ t1 Exceeding rated utilization μ t0 That is, μ t1 >μ t0 Using ΔT1 as the time step, the setting time of parameter T1 is extended by T1+iΔT1, where T1+iΔT1<T2, i=1,2,3... This method improves the training effect of the small model and increases the utilization rate of the small model in actual use. 2) When When m is a specific value; with ΔT2 as the time step, the setting time of parameter T2 is extended by T2+jΔT2, j=1,2,3..., in this way to improve the training effect of large models; The system statistically analyzes the usage and computational load of small and large models over different time periods to form a historical database, which reflects the computational overhead characteristics under different data scales and switching strategies.

Images