RSS fingerprint cleaning and density enhancing method for visible light positioning

Abstract

The invention relates to the technical field of indoor visible light positioning, in particular to an RSS fingerprint cleaning and density enhancing method for visible light positioning, which comprises the following steps: acquiring an initial RSS fingerprint data set through spatial sparse sampling, and cleaning data based on multi-LED spatial correlation and a 3 sigma criterion to eliminate abnormal values; virtual grid points are newly added through Delaunay triangulation interpolation, and an RSS regeneration set of the virtual grid points is generated in combination with an improved geometric constraint extreme learning machine; combining the cleaned data with the virtual data, and outputting a high-resolution RSS fingerprint data set; and finally, establishing a positioning model by using an extreme learning machine to realize accurate positioning. The method effectively improves the RSS data robustness and positioning precision, gives consideration to the real-time performance, can adapt to the indoor complex environment, reduces the system deployment and operation and maintenance cost, and can be widely applied to indoor navigation, asset tracking and other scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of indoor visible light positioning technology, in particular to a RSS fingerprint cleaning and density enhancement method for visible light positioning. BACKGROUND

[0002] In various VLP systems, the RSS fingerprint-based VLP technology has become a research hotspot in this field due to its simple implementation, high cost-effectiveness and excellent positioning accuracy. As a data-driven method, the accuracy of RSS fingerprint positioning directly depends on the density and quality of the location fingerprint: theoretical analysis shows that the greater the spatial density and the higher the quality of the location fingerprint, the better the positioning accuracy. However, the traditional RSS fingerprint collection method has two major limitations: first, a large number of RSS fingerprints need to be obtained through labor-intensive field measurement; second, sampling noise and equipment abnormalities may introduce an incorrect RSS fingerprint set, resulting in a decrease in positioning accuracy.

[0003] Therefore, a RSS fingerprint cleaning and density enhancement method for visible light positioning is proposed to solve the above problems. SUMMARY

[0004] The present application aims to provide a RSS fingerprint cleaning and density enhancement method for visible light positioning to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0006] A RSS fingerprint cleaning and density enhancement method for visible light positioning, comprising the following steps:

[0007] S1, obtaining an initial RSS fingerprint dataset through spatial sparse sampling;

[0008] S2, constructing a RSS fingerprint cleaning method based on multi-LED spatial correlation to clean the initial RSS fingerprint dataset and output a clean dataset;

[0009] S3, constructing a RSS fingerprint density enhancement method based on multi-LED spatial correlation to enhance the density of the clean dataset and generate a RSS regeneration set of virtual grid points;

[0010] S4, combining the cleaned original RSS matrix and the RSS regeneration set of virtual grid points to output a high-robustness high-resolution RSS fingerprint dataset.

[0011] As a preferred solution, step S1 comprises:

[0012] S1-1, discretizing the positioning area into sparse grid points, and calculating the distance between each grid point and The Euclidean distance between each LED is used to construct a dimension of The distance matrix from the initial sparse grid points to each LED. ;

[0013] S1-2, Constructing the Distance Matrix Corresponding received signal strength matrix This forms the initial RSS fingerprint dataset.

[0014] As a preferred embodiment, step S2 includes:

[0015] S2-1, Distance Matrix Each column according to The norm is normalized.

[0016] S2-2. Use leave-one-out cross-validation, starting from [the left-one-out method] in each iteration. A single sample is reserved from the dataset of each sample as the test set, and the rest... One sample is used as the training set;

[0017] S2-3. Using an improved weighted regularized extreme learning machine to learn from the training set, the weight matrix of the output layer is obtained. The following optimization problem can be solved: ,in, W is the weight matrix of the output layer; W is the diagonal weighted matrix with dimension 1. , its first diagonal elements γ is the weighting adjustment coefficient. For the first The prediction error of each training sample; To remove the first Remaining after [number] test samples The RSS matrix of each training sample; The hidden layer output matrix has a dimension of . K is the number of neurons in the hidden layer. , This is the weight matrix from the input layer to the hidden layer. To remove the first After 10 samples, the remaining The normalized distance vector of each training sample. This is the hidden layer bias vector. It is the sigmoid activation function; For regularization parameters; for Norm; This is the Frobenius norm, used to calculate the norm of a matrix;

[0018] S2-4, Based on the weight matrix The RSS estimate of the test sample is obtained. ,in, , For the first The hidden layer output vector of each test sample. For the first The RSS estimate vector of each test sample. For the first Normalized distance vectors of each test sample;

[0019] S2-5. Calculate the RSS estimation error of the current test sample. ,in, For the first RSS estimation error for each test sample; For the first The actual RSS vector of each test sample;

[0020] S2-6, Calculation Mean of RSS estimation error for each test sample with standard deviation ;

[0021] S2-7. Clean the original RSS fingerprint dataset based on the 3σ criterion, removing those that satisfy the 3σ criterion. Abnormal data;

[0022] S2-8. Output the cleaned dataset. ,in, This is the distance matrix from the cleaned grid points to each LED. This represents the number of grid points in the clean dataset after cleaning. This is the clean RSS matrix after cleaning. .

[0023] As a preferred embodiment, step S3 includes:

[0024] S3-1, in the original Based on a sparse grid of points, additional points are added using interpolation methods. For each virtual grid point, calculate its distance to... Calculate the Euclidean distance between each LED and construct a new virtual grid point distance matrix to each LED. ;

[0025] S3-2, Matrix Each column according to The norm is normalized.

[0026] S3-3, Using an improved geometrically constrained limit learning machine on a clean dataset. The learning process yields the weight matrix between the output layer and the hidden layer of the geometrically constrained limit learning machine. The following optimization problem can be solved: ,in, The hidden layer output matrix of the clean dataset. , for Normalized distance vectors of clean grid points; These are the prior weight regularization parameters; This is the prior weight matrix based on the physical propagation model; These are the geometric constraint weighting coefficients; For the first The clean grid point and the first Spatial distance weights of virtual grid points , For the first Normalized distance vectors of virtual grid points; For the first Hidden layer output vector of virtual grid points;

[0027] S3-4, Based on weight matrix ,get RSS regenerated set of virtual grid points ,in, , The RSS regeneration set matrix for virtual grid points; This is the hidden layer output matrix for virtual grid points.

[0028] As a preferred option, the interpolation method in step S3-1 is a linear interpolation method based on Delaunay triangulation, which specifically includes: firstly, performing Delaunay triangulation on the original L sparse grid points, and then uniformly inserting new virtual grid points in each triangular unit.

[0029] As a preferred option, the prior weight matrix This is obtained through the following physical propagation model: ,in, LED emission power; For the gain of the LED transmitting antenna; The gain of the receiving antenna at the receiving end; This represents the transmission distance from the grid point to the LED. The angle of incidence between the LED emitted light and the normal to the receiver. Let be the cosine of the angle of incidence.

[0030] As a preferred option, step S4 specifically involves: verifying the original RSS matrix. The RSS matrix regenerated from N virtual grid points Combined column-wise into a high-resolution RSS matrix , This is the final output matrix of a robust, high-resolution RSS fingerprint dataset.

[0031] As a preferred embodiment, step S5 is also included: processing the high-resolution RSS fingerprint dataset. For visible light positioning systems, an extreme learning machine algorithm is used to establish a signal strength-spatial position mapping model. The input of the signal strength-spatial position mapping model is... The signal strength is used to output the spatial coordinates of the corresponding grid points, enabling precise positioning of the terminal.

[0032] As can be seen from the technical solution provided by the present invention above, the RSS fingerprint cleaning and density enhancement method for visible light positioning provided by the present invention has the following beneficial effects:

[0033] Optimize RSS fingerprint data quality: Remove abnormal signals by using a multi-LED spatial correlation cleaning method to ensure data reliability and lay a stable foundation for subsequent positioning and modeling;

[0034] Improve positioning resolution and accuracy: No need for additional manual intensive sampling. By generating virtual grid points to supplement the fingerprint of sparse areas, the positioning results are more accurate, solving the problem of ambiguous location matching caused by traditional sparse sampling.

[0035] Balancing positioning accuracy and real-time performance: Employing efficient modeling algorithms reduces data processing and model training time while lowering real-time positioning latency, adapting to the needs of dynamic indoor scenarios;

[0036] Enhanced environmental adaptability and stability: It can cope with indoor light interference, equipment aging and other conditions, without the need for frequent full-area resampling and calibration, reducing subsequent maintenance operations and extending the stable operation cycle of the system;

[0037] Lowering the barrier to system deployment: No need to increase LED hardware or large amount of manual sampling costs, simplifying calculation requirements, adaptable to various indoor scenarios, and improving the practicality and promotion of the technology. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the steps of an RSS fingerprint cleaning and density enhancement method for visible light positioning according to the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0040] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0041] like Figure 1 As shown, this embodiment of the invention provides an RSS fingerprint cleaning and density enhancement method for visible light positioning, comprising the following steps:

[0042] S1. Obtain the initial RSS fingerprint dataset through spatial sparse sampling;

[0043] S2. Construct an RSS fingerprint cleaning method based on multi-LED spatial correlation to clean the initial RSS fingerprint dataset and output a clean dataset.

[0044] S3. Construct an RSS fingerprint density enhancement method based on multi-LED spatial correlation to enhance the density of a clean dataset and generate an RSS regenerated set of virtual grid points;

[0045] S4. Combine the cleaned original RSS matrix and the RSS regeneration set of the virtual grid points to output a highly robust, high-resolution RSS fingerprint dataset.

[0046] In this embodiment, step S1 aims to obtain an initial RSS (Received Signal Strength) fingerprint dataset through spatial sparse sampling, providing basic data support for subsequent RSS fingerprint cleaning and density enhancement. This step, by discretizing the positioning area, calculating spatial distance, and constructing a signal strength matrix, associates the physical location of the indoor space with visible light signal features, forming the initial "location-signal" mapping relationship required for positioning. The detailed steps are as follows:

[0047] S1-1, Discretize the positioning area into Given a sparse grid of points, calculate the relationship between each grid point and... The Euclidean distance between each LED is used to construct a dimension of The distance matrix from the initial sparse grid points to each LED. The specific operating procedure is as follows:

[0048] Clearly define the physical boundaries and coordinate system of the positioning area, determine the area range and establish the corresponding coordinate system to ensure the accuracy of subsequent location calculations;

[0049] Discretize the positioning area into A sparse grid of points is used to record the spatial coordinates of each point, ensuring that the grid points are evenly distributed and do not overlap within the region.

[0050] Determine indoor deployment The spatial coordinates of each LED are determined to ensure the accuracy of the LED coordinate information, providing a basis for subsequent distance calculations.

[0051] Calculate each grid point and... The Euclidean distance between LEDs quantifies the spatial relationship between grid points and LEDs;

[0052] Following the rule of "LEDs as rows, grid points as columns," all Euclidean distances are arranged to construct a dimension of [missing information]. The distance matrix from the initial sparse grid points to each LED ;

[0053] Core objective: To transform a continuous indoor space into a discrete "grid point-LED" spatial relationship network, quantify the spatial positions of the two through a distance matrix, provide a physical basis for subsequent distance-based derivation of RSS signal strength, and ensure that the spatial correlation of the initial data is not lost;

[0054] S1-2, Constructing the Distance Matrix Corresponding received signal strength matrix The initial RSS fingerprint dataset is generated through the following steps:

[0055] Based on the physical propagation characteristics of visible light signals, a "distance-RSS" mapping model is established. This model needs to reflect the law of signal strength changing with propagation distance.

[0056] Based on the established mapping model, combined with the distance matrix The Euclidean distance in the grid is used to calculate the RSS value of the corresponding LED signal received at each grid point;

[0057] According to Following the rule of "LEDs as rows, grid points as columns," all RSS values ​​are arranged and constructed in dimension [missing information]. Received signal strength matrix ;

[0058] Distance matrix RSS matrix The data is then correlated and integrated to form an initial RSS fingerprint dataset; each dataset sample corresponds to a grid point, containing information from that grid point to... The distance between LEDs and the receiver The RSS value of each LED contains two parts of information;

[0059] Core objective: To transform the spatial location of grid points (represented by their distance from LEDs) into RSS signal features that can be used for localization, forming an initial "location-signal" fingerprint database. This dataset is the basis for subsequent steps S2 (RSS fingerprint cleaning) to remove abnormal data and S3 (RSS fingerprint density enhancement) to generate a virtual grid point RSS regeneration set. Its completeness and accuracy directly affect the robustness of the final localization result.

[0060] In this embodiment, step S2 is used to construct an RSS fingerprint cleaning method based on the spatial correlation of multiple LEDs. Through improved weighted regularized extreme learning machine modeling and the 3σ criterion, outliers (such as noise interference and deviations caused by measurement errors) in the initial RSS fingerprint dataset are removed, outputting a clean dataset with spatial consistency, providing a reliable data foundation for subsequent density enhancement. The detailed steps are as follows:

[0061] S2-1, Distance Matrix Each column according to The norm is normalized, and the specific steps are as follows:

[0062] The processing object is determined to be the initial sparse grid point distance matrix obtained in step S1. ;

[0063] right Each column (i.e., from a single grid point to all) Perform ℓ2 norm normalization on the distance vectors of each LED to ensure that the distance vector of each grid point satisfies the unit norm constraint.

[0064] Core objective: To unify the data scale of distance vectors, providing standardized input for the stable training of subsequent improved weighted regularized extreme learning machines, and ensuring the model's effective learning of spatial correlations among multiple LEDs;

[0065] S2-2. Use leave-one-out cross-validation, starting from [the left-one-out method] in each iteration. A single sample is reserved from the dataset of each sample as the test set, and the rest... Using 10 samples as the training set, the specific steps are as follows:

[0066] The initial RSS fingerprint dataset obtained in step S1 (including Using a grid point sample as the object, initiate cross-validation iteration;

[0067] In each iteration, from One sample is randomly selected from the total samples to serve as the test set (used to verify the model's generalization ability), and the remaining samples... One sample is used as the training set (for model parameter learning);

[0068] Repeat the above process This ensures that each sample is validated once as a test set, ultimately covering the testing and training of all L samples;

[0069] Core objective: Through full-sample-coverage cross-validation, provide each initial RSS sample with an independent opportunity for model prediction and error assessment, avoiding misjudgment or omission of outliers caused by a single data partition;

[0070] S2-3. Using an improved weighted regularized extreme learning machine to learn from the training set, the weight matrix of the output layer is obtained. The following optimization problem can be solved: ,in, W is the weight matrix of the output layer; W is the diagonal weighted matrix with dimension 1. , its first diagonal elements γ is the weighting adjustment coefficient. For the first The prediction error of each training sample; To remove the first Remaining after [number] test samples The RSS matrix of each training sample; The hidden layer output matrix has a dimension of . K is the number of neurons in the hidden layer. , This is the weight matrix from the input layer to the hidden layer. To remove the first After 10 samples, the remaining The normalized distance vector of each training sample. This is the hidden layer bias vector. It is the sigmoid activation function; For regularization parameters; for Norm; This is the Frobenius norm, used to calculate the norm of a matrix;

[0071] Core objective: To establish a normalized distance vector-RSS mapping model using an improved weighted regularized extreme learning machine, providing reliable weight parameters for RSS estimation of subsequent test samples. ;

[0072] S2-4, Based on the weight matrix The RSS estimate of the test sample is obtained. ,in, , For the first The hidden layer output vector of each test sample. For the first The RSS estimate vector of each test sample. For the first The normalized distance vector of each test sample is obtained through the following steps:

[0073] Extract the first Normalized distance vector of each test sample Substitute into the hidden layer output function Calculate the hidden layer output vector of the test sample. ( For dimension The vector, and (The column vector structure is consistent)

[0074] Using the weight matrix obtained in step S2-3 Through formula Calculate the first RSS estimation vector of each test sample (dimension) The corresponding test sample received (Estimated RSS value of each LED);

[0075] Core objective: To obtain the RSS estimate for each test sample, laying the foundation for subsequent calculation of estimation error and identification of outlier data;

[0076] S2-5. Calculate the RSS estimation error of the current test sample. ,in, For the first RSS estimation error for each test sample; For the first The actual RSS vector of each test sample is obtained through the following steps:

[0077] Extract the first The actual RSS vector of each test sample (From the initial RSS matrix) ) (column) and the estimated RSS vector obtained in step S2-4 .

[0078] pass The norm calculates the deviation between the two, i.e., the estimation error. ( The larger the value, the greater the deviation between the estimated value and the true value, and the more likely the sample is to be abnormal.

[0079] Core objective: To assign a quantified error index to each initial sample, serving as the core basis for subsequent judgments on whether a sample is abnormal;

[0080] S2-6, Calculation Mean of RSS estimation error for each test sample with standard deviation The specific operating steps are as follows:

[0081] Collect all the results obtained in steps S2-5 Estimation error per test sample ;

[0082] Calculate the arithmetic mean of the error This reflects the average estimation bias level across all samples;

[0083] Calculate the standard deviation of the error This reflects the degree of dispersion of the error, that is, the range of fluctuation of the sample error within the statistical mean.

[0084] Core objective: To establish a statistical distribution benchmark for errors and provide a quantitative basis for the "normal error range" of the 3σ criterion;

[0085] S2-7. Clean the original RSS fingerprint dataset based on the 3σ criterion, removing those that satisfy the 3σ criterion. For abnormal data, the specific steps are as follows:

[0086] Set the anomaly detection threshold as follows (Based on the characteristics of normal distribution, 99.7% of normal data will fall within this interval, and data outside the interval are considered abnormal).

[0087] Estimation error for each sample Make a judgment: If the condition is satisfied If the RSS data of the sample deviates from the normal range (possibly due to noise or measurement errors), it is considered abnormal data and removed; if it meets the following conditions... If the result is negative, it is considered normal data and retained in the clean dataset.

[0088] Core objective: To accurately identify and remove outliers in the initial data, ensuring that the remaining data has consistent spatial correlation, and providing high-quality clean data for subsequent density enhancement;

[0089] S2-8. Output the cleaned dataset. ,in, This is the distance matrix from the cleaned grid points to each LED. This represents the number of grid points in the clean dataset after cleaning. This is the clean RSS matrix after cleaning. The specific operating steps are as follows:

[0090] Collect all samples determined to be normal in steps S2-7, and count their number. ( (Due to the removal of some abnormal samples);

[0091] From the initial distance matrix Extract the columns corresponding to normal samples to form a clean distance matrix. (dimension) );

[0092] From the initial RSS matrix Extract the columns corresponding to normal samples to form a clean RSS matrix. (dimension) );

[0093] Integration and Output a clean dataset ;

[0094] Core objective: To output clean RSS fingerprint data with high reliability and strong spatial correlation, providing reliable input for density enhancement in step S3, and ensuring the robustness of the final high-resolution RSS fingerprint dataset from the source.

[0095] In this embodiment, step S3 is used to construct an RSS fingerprint density enhancement method based on multi-LED spatial correlation. Virtual grid points are added to the clean dataset, and the distance-RSS mapping rule is learned through an improved geometrically constrained limit learning machine to generate an RSS regenerated set of virtual grid points. This compensates for the insufficient density of the initial sparse sampling and lays the foundation for subsequent merging of high-resolution datasets. The detailed steps are as follows:

[0096] S3-1, in the original Based on a sparse grid of points, additional points are added using interpolation methods. For each virtual grid point, calculate its distance to... Calculate the Euclidean distance between each LED and construct a new virtual grid point distance matrix to each LED. The interpolation method in step S3-1 is a linear interpolation method based on Delaunay triangulation, specifically including: first, performing Delaunay triangulation on the original L sparse grid points, and then uniformly inserting new virtual grid points in each triangular unit. The specific operation steps are as follows:

[0097] The original obtained in step S1 Based on a sparse grid of points, a linear interpolation method based on Delaunay triangulation is used to add virtual grid points: First, the original grid points are... Delaunay triangulation is performed on sparse grid points to divide the positioning region into several non-overlapping triangular units (satisfying the Delaunay criterion: the circumcircle of any triangular unit does not contain other original grid points, avoiding elongated triangular units to ensure interpolation accuracy); then, new virtual grid points are inserted into each triangular unit according to a uniform interval rule, accumulating the new points. Each virtual grid point records its spatial coordinates.

[0098] Calculate each virtual grid point and The Euclidean distance between LEDs (the calculation logic is the same as the Euclidean distance between "grid point-LED" in S1-1, and the straight-line distance is calculated based on the coordinates of the virtual grid point and the LED).

[0099] Following the rule of "LEDs as rows and virtual grid points as columns," all virtual grid points are arranged according to their Euclidean distances to the LEDs, constructing a dimension of... Distance matrix from virtual grid points to each LED ( For the number of LEDs, (To increase the number of virtual grid points).

[0100] Core objective: To increase grid point density by adding virtual points between the original sparse grid points through Delaunay triangulation and linear interpolation; at the same time, to construct a distance matrix of virtual points to provide a spatial distance basis for the subsequent derivation of the virtual point RSS value, thereby realizing the transformation of the grid structure from sparse to dense.

[0101] S3-2, Matrix Each column according to The norm is normalized, specifically by the following steps:

[0102] The processing object is determined to be the virtual grid point distance matrix obtained in step S3-1. (dimension) );

[0103] right Each column (i.e., a single virtual grid point to all) The distance vectors of each LED are respectively processed. Norm normalization is calculated in the same way as in S2-1: for the distance vector of the IP-th virtual grid point Its normalized vector satisfy ;

[0104] Core objective: To unify the data scale of the distance vectors of virtual grid points, making them consistent with the normalized distance vector format of the clean dataset output by S2, providing standardized input for subsequent model training, and ensuring the consistent transmission of spatial correlation between virtual and clean LED points;

[0105] S3-3, Using an improved geometrically constrained limit learning machine on a clean dataset. The learning process yields the weight matrix between the output layer and the hidden layer of the geometrically constrained limit learning machine. The following optimization problem can be solved: ,in, The hidden layer output matrix of the clean dataset. , for Normalized distance vectors of clean grid points; These are the prior weight regularization parameters; This is the prior weight matrix based on the physical propagation model; These are the geometric constraint weighting coefficients; For the first The clean grid point and the first Spatial distance weights of virtual grid points , For the first Normalized distance vectors of virtual grid points; For the first Hidden layer output vectors for each virtual grid point, prior weight matrix This is obtained through the following physical propagation model:

[0106] ,in, LED emission power; For the gain of the LED transmitting antenna; The gain of the receiving antenna at the receiving end; This represents the transmission distance from the grid point to the LED. The angle of incidence between the LED emitted light and the normal to the receiver. Let be the cosine of the angle of incidence;

[0107] Core objective: To learn a reliable weight matrix by combining the laws of physical propagation with spatial geometric constraints. This provides model parameters for accurate prediction of subsequent virtual grid point RSS values, ensuring that the generated virtual RSS values ​​conform to physical laws and are consistent with the spatial correlation of clean data.

[0108] S3-4, Based on weight matrix ,get RSS regenerated set of virtual grid points ,in, , The RSS regeneration set matrix for virtual grid points; The specific steps for outputting the hidden layer matrix for virtual grid points are as follows:

[0109] Construct the hidden layer output matrix of virtual grid points Based on the virtual grid point normalized distance vector obtained in step S3-2 ( Calculate the hidden layer output vector for each virtual point. ;Will indivual Arranged in columns, forming (dimension) ;

[0110] Based on the weight matrix obtained in step S3-3 The RSS regeneration set matrix of the virtual grid points is calculated using matrix multiplication: ( Dimensions );

[0111] Output the RSS regenerated set of virtual grid points Ensure the matrix has no missing or outlier values ​​(if numerical anomalies exist, backtracking is required). Solve or (Construction process, excluding calculation errors);

[0112] Core objective: To generate virtual grid point RSS values ​​that are spatially correlated with the clean dataset and conform to the laws of physical propagation, thereby enhancing the density of the RSS fingerprint dataset. This provides crucial virtual sample support for merging the high-resolution dataset in step S4, ultimately improving the resolution and accuracy of the positioning system.

[0113] In this embodiment, step S4 specifically involves: verifying the original RSS matrix. The RSS matrix regenerated from N virtual grid points Combined column-wise into a high-resolution RSS matrix , The final output is a highly robust, high-resolution RSS fingerprint dataset matrix;

[0114] Furthermore, the core function of step S4 is to standardize and merge the cleaned, reliable original RSS data with the density-enhanced virtual RSS data, integrating the RSS fingerprint information of the clean original grid points and virtual grid points. The final output is an RSS fingerprint dataset that combines high robustness (due to outlier removal) and high resolution (due to virtual point supplementation), providing core data support for the subsequent accurate positioning of the visible light positioning system. Specific operation steps include:

[0115] S4-1, Input matrix dimension and consistency check:

[0116] Before merging, the key attributes of the two matrices must be verified to avoid merging errors due to data mismatch.

[0117] Dimensional consistency check: Confirm and The number of rows is exactly the same (both are) (Rows); Since the row number corresponds to the number of LEDs, if the row numbers are inconsistent, it means that the two matrices were generated based on different numbers of LEDs, the signal dimensions of the RSS data do not match, and they cannot be merged; it is necessary to backtrack to step S2 or S3 to check for errors in the LED count.

[0118] Data integrity verification: Check whether there are missing values ​​(such as null values ​​or NaN values) in two matrices. Since S2 has completed outlier removal, it is necessary to confirm that there are no missing valid columns. It is necessary to confirm that there are no missing columns due to virtual point generation failure; if there are missing columns, the RSS data of the corresponding grid points must be regenerated (e.g., S2 re-cleaning or S3 recalculating the virtual point RSS).

[0119] Data validity verification: Randomly check the RSS values ​​of two matrices to confirm that the numerical range is consistent with physical laws (no extreme values ​​exceeding -80dBm to -20dBm), avoiding errors in model parameters when S3 generates virtual data (e.g., ...). Outliers may be included due to calculation bias.

[0120] S4-2, Perform column merge operation:

[0121] After confirming that the input matrix is ​​correct, the two matrices are merged using the "column concatenation" method, with the following specific rules:

[0122] Merging logic: The core principle is to "preserve the complete RSS fingerprint of each grid point"—each column of the matrix corresponds to the complete RSS fingerprint of a grid point (including the data received by that grid point). Merging all RSS values ​​of each LED by column ensures that the fingerprint information of each grid point is not split, while directly increasing the total number of grid points (increasing density).

[0123] Merge order: No mandatory order requirement (merge can be done first) of Columns, then spliced ​​together of (or vice versa), but the merging order must be recorded (e.g., before the label). Listed as clean original grid points, then (Listed as virtual grid points), which facilitates the subsequent location of the model and the association of grid point coordinates (clean original grid point coordinates come from S1, and virtual grid point coordinates come from S3-1);

[0124] Calculation of matrix dimensions after merging: High-resolution RSS matrix after merging Dimensions ,in:

[0125] Line count: Keep Row (number of LEDs remains the same);

[0126] Column count: The number of clean, original grid points With the number of virtual grid points The sum (increased total number of grid points, improved resolution);

[0127] S4-3, Output high-resolution RSS fingerprint dataset:

[0128] After merging, a final highly robust, high-resolution RSS fingerprint dataset is generated. The specific output content and requirements are as follows:

[0129] Output matrix definition: The merged matrix is ​​formally defined as follows: And specify its data structure:

[0130] Row index: corresponding LEDs (from 1 to 1) (Label the LED number and coordinates for each row).

[0131] Column index: corresponding Grid points (from 1 to \( + Label each column with the grid point type (clean raw / virtual) and coordinates.

[0132] Element meaning: (No. Line number (Column) represents "the first The grid point receives the first "RSS value of each LED";

[0133] Output format requirements: Store in matrix form (e.g., CSV, MAT format), and include a "grid point information table" recording the information for each column (grid point):

[0134] Grid point ID;

[0135] Type (clean raw / virtual);

[0136] Spatial coordinates (x, y, consistent with the coordinate system of S1 and S3-1);

[0137] Corresponding to the original data source (e.g., clean original grid points correspond to the initial grid point ID of S1, and virtual grid points correspond to the triangulation unit ID of S3-1).

[0138] Data backup and verification: Perform a complete backup, and randomly select 10%-20% of the columns (grid points) to verify the spatial correlation between their RSS values ​​and the corresponding grid point coordinates (e.g., the RSS values ​​of adjacent grid points should be closer) to ensure that the spatial consistency of the merged data is not compromised.

[0139] In this embodiment, the method further includes step S5: processing the high-resolution RSS fingerprint dataset. For visible light positioning systems, an extreme learning machine algorithm is used to establish a signal strength-spatial position mapping model. The input of the signal strength-spatial position mapping model is... The signal strength is used to output the spatial coordinates of the corresponding grid points, enabling precise positioning of the terminal.

[0140] Furthermore, step S5 involves constructing a visible light positioning model based on a high-resolution RSS fingerprint dataset. Through model training, real-time positioning, and accuracy optimization, it achieves accurate estimation of the target location and ultimately outputs a stable and reliable positioning result, providing core technical support for the practical application of indoor visible light positioning systems. The detailed steps are as follows:

[0141] Step S5-1: Divide the training dataset for the localization model:

[0142] The high-resolution RSS fingerprint dataset output in step S4 (dimension) , For the number of LEDs, Based on the total number of grid points, the spatial coordinates of each grid point are associated (clean original grid point coordinates come from S1, and virtual grid point coordinates come from S3-1), forming a set of "RSS fingerprint-coordinate" sample pairs. Each sample has the following structure: For the first RSS vector of grid points (its coordinates);

[0143] A random partitioning strategy is adopted to divide the sample pair set into a training set and a test set: samples are randomly selected according to a preset ratio (e.g., 7:3), with 70% used as the training set (for model parameter learning) and 30% used as the test set (for evaluating the model's generalization ability), ensuring that the spatial distribution of the two sets of samples is consistent (avoiding the training set being concentrated in a certain area, which could lead to model bias).

[0144] Core objective: To provide independent data support for the training and performance evaluation of subsequent localization models, and to achieve the separation and verification of model learning ability and generalization ability;

[0145] Step S5-2: Training the localization model based on Extreme Learning Machine:

[0146] Construct an Extreme Learning Machine (ELM) localization model, with the RSS vectors of grid points as input. (dimension) The output is the coordinates of that grid point. (dimension) );

[0147] Using training set samples as input, the model parameters are solved using the ELM algorithm:

[0148] Randomly initialize the weight matrix from the input layer to the hidden layer. (dimension) , (Number of neurons in the hidden layer) and bias vector (dimension) );

[0149] Calculate the hidden layer output matrix (dimension) ), of which, the List , For activation functions (such as the sigmoid function) );

[0150] The output layer weight matrix is ​​solved by the Moore-Penrose generalized inverse. (dimension) The formula is ,in for The generalized inverse, For the training set coordinate matrix (dimensions) , No. Listed as training samples coordinates );

[0151] Save the parameters of the trained model , , This leads to the formation of an ELM model that can be used for localization prediction;

[0152] Core objective: To establish a nonlinear mapping model of "RSS fingerprint-spatial coordinates" so that the model can infer the spatial coordinates of an unknown location from the RSS signal, thus providing an algorithmic basis for real-time positioning.

[0153] Step S5-3: Real-time positioning and coordinate prediction:

[0154] Real-time acquisition of RSS signals at the target location: synchronous acquisition via receiving equipment (such as a photodetector). The real-time RSS values ​​of each LED are used to form the RSS vector of the target to be located. (dimension) Ensure that the signal acquisition frequency matches the positioning requirements (e.g., 10Hz, to meet real-time requirements);

[0155] Preprocessing real-time RSS vectors: Perform standardization processing consistent with the training data (such as removing DC offset and filtering noise reduction) to eliminate signal fluctuations caused by environmental interference and ensure that the input format is consistent with the model training data;

[0156] Input ELM model predicted coordinates: (Preprocessed coordinates) Input the model trained in step S5-2 and calculate the hidden layer output. Then through Obtain the predicted coordinates of the target location ;

[0157] Core objective: To apply the trained localization model to real-world scenarios, and through real-time signal acquisition and model prediction, output the coordinate estimate of the target location, thus realizing the practical application of the localization function;

[0158] Step S5-4: Positioning accuracy assessment and error analysis

[0159] Calculate the localization error of the test set: Input the RSS vector of the test set samples into the ELM model to obtain the predicted coordinates. , and the actual coordinates In comparison, the positioning error of each sample was calculated using Euclidean distance.

[0160] Statistical error index: Calculates the average localization error of the test set. Error standard deviation And the 90% confidence interval error (the error of 90% of the test samples is less than this value) to comprehensively evaluate the model accuracy;

[0161] Error distribution analysis: Plot histograms and cumulative distribution curves of positioning errors, analyze the areas of concentrated error and the sources of outliers (such as whether the error is too high due to deviations in the generation of virtual points in a certain area), and provide direction for model optimization;

[0162] Core objective: To quantitatively evaluate the performance of the positioning model, identify the model's accuracy level and potential problems, provide data support for subsequent iterative optimization, and ensure that the positioning system meets the accuracy requirements of practical applications;

[0163] Step S5-5: Iterative Model Optimization and Robustness Enhancement

[0164] Adjust model parameters based on error analysis results: If the average error is too high, increase the number of neurons in the hidden layer. Alternatively, change the activation function (e.g., use the ReLU function to enhance nonlinear fitting ability); if the error standard deviation is large, introduce a regularization term during training (e.g., Suppress overfitting;

[0165] Dynamically update the training dataset: Collect new real RSS data regularly (e.g., once a month), clean it through step S2 and add it to the training set, while removing outdated samples (e.g., RSS feature shifted samples caused by environmental changes) to keep the dataset timely.

[0166] Multi-sensor data fusion for optimized positioning: If single RSS positioning has limitations, motion data from inertial measurement units (IMUs) can be fused. Kalman filtering can be used to fuse RSS positioning results with IMU trajectory predictions to reduce positioning jumps caused by instantaneous signal interference.

[0167] Core objective: Through continuous iteration and optimization, address issues such as insufficient accuracy and poor robustness that may arise in positioning models during practical applications, ensuring long-term stable operation of the system and meeting positioning needs in different scenarios;

[0168] Steps S5-6: Final Location Result Output and System Deployment

[0169] Determine the optimal model: Select the ELM model that has the smallest average error on the test set and the highest stability after iterative optimization as the final localization model;

[0170] Encapsulated positioning system: Integrates model parameters, real-time signal acquisition module, and coordinate prediction module into a complete positioning system, providing standardized interfaces (such as TCP / IP protocol) to support interface with upper-layer applications (such as indoor navigation and asset tracking);

[0171] Output positioning results: After system deployment, the coordinates of the target location are output in real time. It also includes location reliability (calculated based on error distribution, such as...) , (This is the current estimated prediction error value), which is used by the application layer for decision-making;

[0172] Core objective: To complete the engineering implementation of the positioning system, output high-precision and highly reliable real-time positioning results, and provide core technical support for the practical application of indoor visible light positioning (such as smart factories and warehouse management).

[0173] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for cleaning and density enhancement of RSS fingerprints for visible light positioning, characterized in that: Includes the following steps: S1. Obtain the initial RSS fingerprint dataset through spatial sparse sampling; S2. Construct an RSS fingerprint cleaning method based on multi-LED spatial correlation to clean the initial RSS fingerprint dataset and output a clean dataset. S3. Construct an RSS fingerprint density enhancement method based on multi-LED spatial correlation to enhance the density of a clean dataset and generate an RSS regenerated set of virtual grid points; S4. Combine the cleaned original RSS matrix and the RSS regeneration set of the virtual grid points to output a highly robust, high-resolution RSS fingerprint dataset.

2. The RSS fingerprint cleaning and density enhancement method for visible light positioning according to claim 1, characterized in that: Step S1 includes: S1-1, Discretize the positioning area into Given a sparse grid of points, calculate the relationship between each grid point and... The Euclidean distance between each LED is used to construct a dimension of The distance matrix from the initial sparse grid points to each LED. ; S1-2, Constructing the Distance Matrix Corresponding received signal strength matrix This forms the initial RSS fingerprint dataset.

3. The RSS fingerprint cleaning and density enhancement method for visible light positioning according to claim 2, characterized in that: Step S2 includes: S2-1, Distance Matrix Each column according to The norm is normalized. S2-2. Use leave-one-out cross-validation, starting from [the left-one-out method] in each iteration. A single sample is reserved from the dataset of each sample as the test set, and the rest... One sample is used as the training set; S2-3. Using an improved weighted regularized extreme learning machine to learn from the training set, the weight matrix of the output layer is obtained. The following optimization problem can be solved: ,in, W is the weight matrix of the output layer; W is the diagonal weighted matrix with dimension 1. , its first diagonal elements γ is the weighting adjustment coefficient. For the first The prediction error of each training sample; To remove the first Remaining after [number] test samples The RSS matrix of each training sample; The hidden layer output matrix has a dimension of . K is the number of neurons in the hidden layer. , This is the weight matrix from the input layer to the hidden layer. To remove the first After 10 samples, the remaining The normalized distance vector of each training sample. This is the hidden layer bias vector. It is the sigmoid activation function; For regularization parameters; for Norm; This is the Frobenius norm, used to calculate the norm of a matrix; S2-4, Based on the weight matrix The RSS estimate of the test sample is obtained. ,in, , For the first The hidden layer output vector of each test sample. For the first The RSS estimate vector of each test sample. For the first Normalized distance vectors of each test sample; S2-5. Calculate the RSS estimation error of the current test sample. ,in, For the first RSS estimation error for each test sample; For the first The actual RSS vector of each test sample; S2-6, Calculation Mean of RSS estimation error for each test sample with standard deviation ; S2-7. Clean the original RSS fingerprint dataset based on the 3σ criterion, removing those that satisfy the 3σ criterion. Abnormal data; S2-8. Output the cleaned dataset. ,in, This is the distance matrix from the cleaned grid points to each LED. This represents the number of grid points in the clean dataset after cleaning. This is the clean RSS matrix after cleaning. .

4. The RSS fingerprint cleaning and density enhancement method for visible light positioning according to claim 3, characterized in that: Step S3 includes: S3-1, in the original Based on a sparse grid of points, additional points are added using interpolation methods. For each virtual grid point, calculate its distance to... Calculate the Euclidean distance between each LED and construct a new virtual grid point distance matrix to each LED. ; S3-2, Matrix Each column according to The norm is normalized. S3-3, Using an improved geometrically constrained limit learning machine on a clean dataset. The learning process yields the weight matrix between the output layer and the hidden layer of the geometrically constrained limit learning machine. The following optimization problem can be solved: ,in, The hidden layer output matrix of the clean dataset. , for Normalized distance vectors of clean grid points; These are the prior weight regularization parameters; This is the prior weight matrix based on the physical propagation model; These are the geometric constraint weighting coefficients; For the first The clean grid point and the first Spatial distance weights of virtual grid points , For the first Normalized distance vectors of virtual grid points; For the first Hidden layer output vector of virtual grid points; S3-4, Based on weight matrix ,get RSS regenerated set of virtual grid points ,in, , The RSS regeneration set matrix for virtual grid points; This is the hidden layer output matrix for virtual grid points.

5. The RSS fingerprint cleaning and density enhancement method for visible light positioning according to claim 4, characterized in that: The interpolation method in step S3-1 is a linear interpolation method based on Delaunay triangulation, which specifically includes: firstly, performing Delaunay triangulation on the original L sparse grid points, and then uniformly inserting new virtual grid points in each triangular unit.

6. The RSS fingerprint cleaning and density enhancement method for visible light positioning according to claim 4, characterized in that: The prior weight matrix This is obtained through the following physical propagation model: ,in, LED emission power; For the gain of the LED transmitting antenna; The gain of the receiving antenna at the receiving end; This represents the transmission distance from the grid point to the LED. The angle of incidence between the LED emitted light and the normal to the receiver. Let be the cosine of the angle of incidence.

7. The RSS fingerprint cleaning and density enhancement method for visible light positioning according to claim 4, characterized in that: Step S4 specifically involves: verifying the original RSS matrix. The RSS matrix regenerated from N virtual grid points Combined column-wise into a high-resolution RSS matrix , This is the final output matrix of a robust, high-resolution RSS fingerprint dataset.

8. The RSS fingerprint cleaning and density enhancement method for visible light positioning according to claim 1, characterized in that: It also includes step S5, which involves processing the high-resolution RSS fingerprint dataset. For visible light positioning systems, an extreme learning machine algorithm is used to establish a signal strength-spatial position mapping model. The input of the signal strength-spatial position mapping model is... The signal strength is used to output the spatial coordinates of the corresponding grid points, enabling precise positioning of the terminal.

Images