Bluetooth indoor positioning method based on Bluetooth AOA and RSSI

Abstract

The invention provides a Bluetooth indoor positioning method based on Bluetooth AOA and RSSI. The method mainly comprises the following steps: acquiring arrival angle information and received signal strength indication information of Bluetooth; preprocessing the arrival angle information and the received signal strength indication to obtain pre-training data of the positioning fingerprint; carrying out feature extraction, feature splicing and classified learning by utilizing an Encoder layer and a Decoder layer of a Transform model, so as to obtain a Bluetooth position positioning model; the method comprises the following steps: measuring arrival angle information and received signal strength indication of Bluetooth, and preprocessing the arrival angle information and the received signal strength indication; and substituting the preprocessed arrival angle information and the received signal strength indication data into a Bluetooth position positioning model to obtain a position estimation result of the target. According to the method, the Bluetooth arrival angle information and the received signal strength indication data are comprehensively utilized, feature extraction and classified learning are performed through the Transform deep learning network, and the method has the advantages of low complexity, high positioning performance, simplicity in implementation and the like.

Description

Technical Field

[0001] The present invention relates to a Bluetooth indoor positioning method, and belongs to the field of positioning navigation and machine learning applications. Background Art

[0002] With the rapid development of the Internet of Things (IoT) and mobile communication technologies, as well as the rapid upgrade of intelligent hardware devices, people are increasingly demanding location-based services, with increasingly higher precision requirements. Consequently, indoor and outdoor positioning technologies have become an integral part of daily life. Indoor positioning technology, in particular, enables precise positioning, navigation, and security management and monitoring in indoor environments, playing a crucial role in improving user experience and boosting productivity. This technology is currently widely used in public places such as hospitals, shopping malls, and airports, as well as in warehouse management, smart home appliances, and logistics sorting. Traditional indoor positioning technologies primarily include ultrasonic positioning and infrared beacon positioning. These solutions suffer from high deployment and maintenance costs, low positioning accuracy, and susceptibility to environmental interference. Consequently, radio frequency (RF)-based indoor positioning technologies, such as radio frequency identification (RFID), wireless networks (WiFi), ultra-wideband (UWB), and Bluetooth Low Energy (BLE), have become a hot topic. BLE-based indoor positioning technology offers advantages such as low cost, low power consumption, wide coverage, and strong adaptability, making it the most reliable RF-based indoor positioning technology.

[0003] Bluetooth low energy positioning measurement parameters include signal strength indicator (RSSI), angle of arrival (AOA), and time of arrival (TOA). RSSI measurements are susceptible to diffraction and reflection from obstacles. AOA positioning, on the other hand, can be described as a nonlinear estimation problem, primarily locating a mobile target by measuring the angle of arrival between the signal's mobile station and base station. With the continuous development of Bluetooth technology, the Bluetooth 5.1 standard specification has added Bluetooth direction-finding technology, further improving positioning accuracy. This has made Bluetooth positioning a popular research area for indoor positioning.

[0004] Compared to traditional indoor positioning technologies, single-use RSSI or AOA positioning suffers from significant environmental impact, limited accuracy, and limited coverage. Therefore, to ensure the stability and accuracy of indoor positioning, multi-feature fusion technology has become a key research direction. Multi-feature fusion technology leverages multiple features to provide more comprehensive spatial information, thereby improving the robustness of the positioning system and reducing the impact of environmental factors. Furthermore, this technology can select appropriate wireless parameters for joint positioning based on application requirements, making it applicable in a wider range of scenarios. Multi-feature fusion technology can also be integrated with machine learning and deep learning methods to train and optimize positioning algorithms, further improving positioning accuracy and robustness. Furthermore, the Transformer model extracts feature values from RSSI and AOA, and its self-attention mechanism correlates these feature values, significantly improving positioning accuracy. The KNN method is used to optimize distance estimation, calculating the closest Euclidean distances for multiple grid points and averaging the coordinates to obtain the location point.

[0005] Therefore, it is necessary to propose a Bluetooth indoor positioning method to solve the above problems. Summary of the Invention

[0006] Technical problem: The purpose of this invention is to provide a Bluetooth positioning method that integrates the received signal strength indication and arrival angle information and combines it with deep learning, which significantly improves the positioning accuracy, shortens the positioning time, and reduces the positioning complexity.

[0007] Technical Solution: This paper proposes a Bluetooth indoor positioning method based on Bluetooth AOA and RSSI. This method combines the advantages of Bluetooth arrival angle information and received signal strength indication, and uses the self-attention mechanism of the Transformer model to fuse the two data to obtain fingerprint data for classification, thereby optimizing positioning accuracy. The specific solution is as follows:

[0008] A Bluetooth indoor positioning method based on Bluetooth AOA and RSSI includes the following steps:

[0009] Step 1: Collect Bluetooth angle of arrival (AOA) and received signal strength indicator (RSSI) data, and pre-process them through linear discriminant analysis and particle filtering to obtain training data for positioning fingerprints.

[0010] Step 2: Use the encoder and decoder layers of the Transformer to extract features and use the self-attention mechanism for feature splicing. Then, use the KNN algorithm for classification learning to obtain the Bluetooth positioning model.

[0011] Step 3: Measure the arrival angle information and received signal strength indicator of the Bluetooth at the target location, perform the operation preprocessing of step 1, and bring the preprocessed data into the Bluetooth location positioning model to obtain the position estimation result of the target location.

[0012] Furthermore, step one specifically includes the following steps:

[0013] Step 11: Use linear discriminant analysis to reduce the dimension of received signal strength knowledge;

[0014] Step 12: Optimize the original arrival angle data through the particle filter algorithm and normalize it to form a positioning fingerprint.

[0015] Furthermore, in step 11, the linear discriminant analysis method includes the following steps:

[0016] (1) Data standardization and category definition;

[0017] 1. Construct N×M dimensional raw data matrix D for RSSI fingerprint library R , where N is the total number of fingerprint samples, M is the RSSI signal dimension; define the data category label set as {y n |y n ∈{1,2,...,C}}, C is the total number of categories divided into the positioning area;

[0018] 2. Calculate the global mean vector μ and the category mean vector μ for each dimension c :

[0019]

[0020]

[0021] Among them, R n is the RSSI signal vector of the nth sample;

[0022] 3. Generate a standardized matrix Its nth sample is:

[0023]

[0024] Where Σ is the covariance matrix;

[0025] (2) Scatter matrix calculation;

[0026] Calculate the intra-class scatter matrix S w :

[0027]

[0028] Calculate the inter-class divergence matrix S b :

[0029]

[0030] (3) Generalized Rayleigh quotient optimization;

[0031] Solve the generalized eigenvalue problem:

[0032] S b w i =λ i S w w i (i=1,2,…,d)

[0033] Select the eigenvectors corresponding to the first d largest eigenvalues to form the projection matrix W = [w1,w2…,w d ], where d≤C-1, the maximum number of dimensions that can be reduced by LDA is determined by the number of categories;

[0034] (4) Data projection and dimensionality reduction

[0035] Project the normalized data into a low-dimensional discriminant space:

[0036]

[0037] (5) Dimensionality reduction selection

[0038] The value of d is selected based on the classification performance verification, or d=C-1 is preset to retain the maximum discriminant information.

[0039] Furthermore, in step 12, the smoothing process of the particle filter includes the following steps:

[0040] (1) Particle initialization and state modeling

[0041] (2) State transition and prediction

[0042] (3) Observation likelihood and weight update

[0043] (4) Resampling and error control

[0044] (5) State estimation and output

[0045] Furthermore, in step (1), particle initialization and state modeling include the following steps:

[0046] State vector definition: The particle state vector includes position, direction and dynamic noise parameters, that is, Where (x, y) is the two-dimensional coordinate, θ is the motion direction angle, v is the velocity, and ω is the angular velocity noise term;

[0047] Initial distribution generation: Generate N particles in the initial positioning area, and the position component follows the Gaussian distribution N(μ p ,Σp ), where μ p is the initial estimated position, Σ p Obtained by mapping the covariance matrix of the RSSI fingerprint library.

[0048] Furthermore, in step (2), state transfer and prediction include the following steps:

[0049] Motion model construction: Pedestrian dead reckoning (PDR) model is used to update particle states:

[0050]

[0051] Where f(·) is the nonlinear motion equation, Q is the process noise covariance matrix, which includes step error and azimuth noise. The step error adopts the RSSI dynamic attenuation model, and the azimuth noise adopts the AoA phase jitter model.

[0052] Dynamic noise compensation: Dynamically adjust noise parameters according to Bluetooth signal strength:

[0053] ω (i) =α*RSSI var +β*AoA std

[0054] Among them, α and β are trainable coefficients, RSSI var and AoA std are RSSI variance and AoA standard deviation respectively.

[0055] Furthermore, in step (3), the observation likelihood and weight update includes the following steps:

[0056] RSSI observation model: The logarithmic path loss model is used to calculate the similarity between particles and the fingerprint library:

[0057]

[0058]

[0059] in, is the RSSI similarity between the i-th particle and the fingerprint library, R m is the RSSI value of the mth signal source actually measured, is the RSSI value of the predicted mth signal source at the position of the i-th particle, P0 is the signal strength at the reference distance, x (i) is the predicted coordinate of the i-th particle, b m is the actual coordinate of the mth signal source;

[0060] AoA observation model: Calculate the angle matching probability based on the antenna array phase difference:

[0061]

[0062] in, is the phase difference received by the antenna array, θ meas is the actual measured arrival angle, σ θ is the AoA measurement noise standard deviation;

[0063] Joint weight calculation:

[0064]

[0065] where σ R and σ θ are the measurement noise standard deviations of RSSI and AoA, respectively.

[0066] Furthermore, in step (4), resampling and error control include the following steps:

[0067] Adaptive resampling strategy: When the number of effective particles When , the system resampling is performed, retaining the high-weight particles and randomly perturbing the positions of the low-weight particles, and the perturbation amplitude is determined by Σ p Dynamic adjustment, where ω (i) is the weight of the i-th particle, N th is the preset threshold for triggering resampling;

[0068] Introduce Random Linear Crossover to enhance diversity: randomly select two particles to generate a new particle x new =γx a +(1-γ)x b , where γ~U(0,1).

[0069] Furthermore, in step (5), state estimation and output include the following steps:

[0070] Weighted fusion estimation: The final position estimate is the weighted mean of the particle set:

[0071]

[0072] in is the weight of the i-th particle at time k, is the state vector of the i-th particle at time k;

[0073] Covariance Update: Calculate the posterior covariance matrix for dynamic noise adjustment:

[0074]

[0075] This covariance is fed back to the process noise model in step (2) to form a closed-loop optimization.

[0076] Furthermore, in step 2, the method for constructing the Bluetooth location positioning model specifically includes the following steps:

[0077] Step 21: Multimodal feature embedding and fusion

[0078] 1. AoA feature high-dimensional mapping: Input the original AoA angle data into the fully connected layer FC AoA , generating a high-dimensional embedding vector where d h is the hidden layer dimension;

[0079] 2. RSSI feature high-dimensional mapping: Input RSSI strength data into an independent fully connected layer PC RSSI , generating a high-dimensional embedding vector

[0080] 3. Cross-modal splicing: E along the feature dimension AoA With E RSSI Perform tensor splicing to obtain joint embedding

[0081] Step 22: Multi-head self-attention feature extraction;

[0082] The multi-head self-attention mechanism (MHSA) is used to joint Conduct context-aware modeling: H attn =LayerNorm(E joint +Dropout(MHSA(E joint )))

[0083] Among them, Dropout performs random regularization on the output of multi-head self-attention, and LayerNorm performs layer normalization on the output of residual connection; MHSA consists of h independent attention heads, each of which performs scaled dot product attention operation:

[0084]

[0085] Among them, Q is the query vector, K is the key vector, and V is the value vector. is the scaling factor, which scales the dot product result of Q and K to avoid training instability;

[0086] Step 23: The output of the self-attention mechanism is passed through a fully connected layer for feature fusion and learned through KNN to obtain a Bluetooth location positioning model.

[0087] Beneficial effects:

[0088] (1) Multimodal data fusion for high positioning accuracy: By combining the azimuth characteristics of Bluetooth Angle of Arrival (AoA) and the distance characteristics of Received Signal Strength (RSSI) (Step 1), the limitations of traditional single signal indicators are overcome. The spatial complementarity of the two data types enables the positioning model to simultaneously perceive the target's direction and distance information, maintaining high positioning accuracy even in complex occlusion environments.

[0089] (2) Robust data processing capabilities: A unified data preprocessing process is used to effectively filter out multipath interference and noise fluctuations in Bluetooth signals. Data quality is improved through operations such as feature normalization and outlier correction, making subsequent model training and prediction more resistant to environmental interference.

[0090] (3) Combining deep feature learning with spatial reasoning: Innovatively applying the Transformer’s self-attention mechanism to the positioning field: The Encoder layer captures the nonlinear relationship between signal features through multi-head attention, and the Decoder layer realizes cross-modal feature fusion, while the KNN

[0091] The classifier performs fast nearest neighbor matching in the reduced-dimensional space. This hybrid architecture combines the representational power of deep learning with the spatial interpretability of traditional algorithms.

[0092] (4) Dynamic Environment Adaptation: Compared to static fingerprint library methods, this solution can automatically learn the spatial pattern changes of signal propagation through the sequence modeling capability of Transformer. When the environment changes dynamically, such as personnel movement and equipment addition and removal, the model can maintain positioning stability through online fine-tuning.

[0093] (5) Low-latency real-time positioning: The end-to-end processing flow achieves a response in seconds: After the pre-processing module completes data cleaning, the model quickly outputs positioning results through parallel attention calculation and KNN fast retrieval, meeting the real-time requirements of indoor navigation, asset tracking, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] Figure 1 Schematic diagram of the flow of the Bluetooth indoor positioning method of the present invention.

[0095] Figure 2 Schematic diagram of the arrival angle of the incident signal in the Bluetooth indoor positioning method of the present invention.

[0096] Figure 3 Schematic diagram of the multi-head self-attention mechanism used in the Bluetooth indoor positioning method of the present invention. DETAILED DESCRIPTION

[0097] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings. Figures 1 to 3 As shown, the present invention discloses a Bluetooth indoor positioning method, which mainly includes the following steps:

[0098] Step 1: Collect Bluetooth angle of arrival (AOA) and received signal strength indicator (RSSI) data, and pre-process them through linear discriminant analysis and particle filtering to obtain training data for positioning fingerprints.

[0099] Step 2: Use the encoder and decoder layers of the Transformer to extract features and use the self-attention mechanism for feature splicing. Then, use the KNN algorithm for classification learning to obtain the Bluetooth positioning model.

[0100] Step 3: Measure the arrival angle information and received signal strength indicator of the Bluetooth at the target location, perform the operation preprocessing of step 1, and bring the preprocessed data into the Bluetooth location positioning model to obtain the position estimation result of the target location.

[0101] In this embodiment, step one specifically includes the following steps:

[0102] Step 11: Use linear discriminant analysis to reduce the dimension of received signal strength knowledge;

[0103] The linear discriminant analysis method includes the following steps:

[0104] (1) Data standardization and category definition;

[0105] 1. Construct N×M dimensional raw data matrix D for RSSI fingerprint library R , where N is the total number of fingerprint samples, M is the RSSI signal dimension; define the data category label set as {y n |y n ∈{1,2,...,C}}, C is the total number of categories divided into the positioning area;

[0106] 2. Calculate the global mean vector μ and the category mean vector μ for each dimension c :

[0107]

[0108]

[0109] Among them, R n is the RSSI signal vector of the nth sample;

[0110] 3. Generate a standardized matrix Its nth sample is:

[0111]

[0112] Where Σ is the covariance matrix;

[0113] (2) Scatter matrix calculation;

[0114] Calculate the intra-class scatter matrix S w :

[0115]

[0116] Calculate the inter-class divergence matrix S b :

[0117]

[0118] (3) Generalized Rayleigh quotient optimization;

[0119] Solve the generalized eigenvalue problem:

[0120] S b w i =λ i S w w i (i=1,2,…,d)

[0121] Select the eigenvectors corresponding to the first d largest eigenvalues to form the projection matrix W = [w1,w2…,w d ], where d≤C-1, the maximum number of dimensions that can be reduced by LDA is determined by the number of categories;

[0122] (4) Data projection and dimensionality reduction

[0123] Project the normalized data into a low-dimensional discriminant space:

[0124]

[0125] (5) Dimensionality reduction selection

[0126] The value of d is selected based on the classification performance verification, or d=C-1 is preset to retain the maximum discriminant information.

[0127] Step 12: Optimize the original arrival angle data through the particle filter algorithm and normalize it to form a positioning fingerprint.

[0128] Among them, the smoothing process of the particle filter includes the following steps:

[0129] (1) Particle initialization and state modeling

[0130] State vector definition: The particle state vector includes position, direction and dynamic noise parameters, that is, Where (x, y) is the two-dimensional coordinate, θ is the motion direction angle, v is the velocity, and ω is the angular velocity noise term;

[0131] Initial distribution generation: Generate N particles in the initial positioning area, and the position component follows the Gaussian distribution N(μ p ,Σ p), where μ p is the initial estimated position, Σ p Obtained by mapping the covariance matrix of the RSSI fingerprint library.

[0132] (2) State transition and prediction

[0133] Motion model construction: Pedestrian dead reckoning (PDR) model is used to update particle states:

[0134]

[0135] Where f(·) is the nonlinear motion equation, Q is the process noise covariance matrix, which includes step error and azimuth noise. The step error adopts the RSSI dynamic attenuation model, and the azimuth noise adopts the AoA phase jitter model.

[0136] Dynamic noise compensation: Dynamically adjust noise parameters according to Bluetooth signal strength:

[0137] ω (i) =α*RSSI var +β*AoA std

[0138] Among them, α and β are trainable coefficients, RSSI var and AoA std are RSSI variance and AoA standard deviation respectively.

[0139] (3) Observation likelihood and weight update

[0140] RSSI observation model: The logarithmic path loss model is used to calculate the similarity between particles and the fingerprint library:

[0141]

[0142]

[0143] in, is the RSSI similarity between the i-th particle and the fingerprint library, R m is the RSSI value of the mth signal source actually measured, is the RSSI value of the predicted mth signal source at the position of the i-th particle, P0 is the signal strength at the reference distance, x (i) is the predicted coordinate of the i-th particle, b m is the actual coordinate of the mth signal source;

[0144] AoA observation model: Calculate the angle matching probability based on the antenna array phase difference:

[0145]

[0146] in, is the phase difference received by the antenna array, θ meas is the actual measured arrival angle, σ θ is the AoA measurement noise standard deviation;

[0147] Joint weight calculation:

[0148]

[0149] where σ R and σ θ are the measurement noise standard deviations of RSSI and AoA, respectively.

[0150] (4) Resampling and error control

[0151] Adaptive resampling strategy: When the number of effective particles When , the system resampling is performed, retaining the high-weight particles and randomly perturbing the positions of the low-weight particles, and the perturbation amplitude is determined by Σ p Dynamic adjustment, where ω (i) is the weight of the i-th particle, N th is the preset threshold for triggering resampling;

[0152] Introduce Random Linear Crossover to enhance diversity: randomly select two particles to generate a new particle x new =γx a +(1-γ)x b , where γ~U(0,1).

[0153] (5) State estimation and output

[0154] Weighted fusion estimation: The final position estimate is the weighted mean of the particle set:

[0155]

[0156] in is the weight of the i-th particle at time k, is the state vector of the i-th particle at time k;

[0157] Covariance Update: Calculate the posterior covariance matrix for dynamic noise adjustment:

[0158]

[0159] This covariance is fed back to the process noise model in step (2) to form a closed-loop optimization.

[0160] In step 2, the method for constructing the Bluetooth location positioning model specifically includes the following steps:

[0161] Step 21: Multimodal feature embedding and fusion

[0162] 1. AoA feature high-dimensional mapping: Input the original AoA angle data into the fully connected layer FC AoA , generating a high-dimensional embedding vector where d h is the hidden layer dimension;

[0163] 2. RSSI feature high-dimensional mapping: Input RSSI intensity data into an independent fully connected layer FC RSSI , generating a high-dimensional embedding vector

[0164] 3. Cross-modal splicing: E along the feature dimension AoA With E RSSI Perform tensor splicing to obtain joint embedding

[0165] Step 22: Multi-head self-attention feature extraction;

[0166] The multi-head self-attention mechanism (MHSA) is used to joint Conduct context-aware modeling: H attn =LayerNorm(E joint +Dropout(MHSA(E joint )))

[0167] Among them, Dropout performs random regularization on the output of multi-head self-attention, and LayerNorm performs layer normalization on the output of residual connection; MHSA consists of h independent attention heads, each of which performs scaled dot product attention operation:

[0168]

[0169] Among them, Q is the query vector, K is the key vector, and V is the value vector. is the scaling factor, which scales the dot product result of Q and K to avoid training instability;

[0170] Step 23: The output of the self-attention mechanism is passed through a fully connected layer for feature fusion and learned through KNN to obtain the Bluetooth location positioning model.

[0171] In step three, after data preprocessing, the RSSI and AOA data are sent to the fully connected layer for mapping and splicing, and then the multi-head self-attention mechanism is used to extract their features.

[0172] The multi-head self-attention mechanism adds more linear layers for keys, queries, and values. These linear layers are trained in parallel and have independent weights. Each value, key, and query provides three outputs instead of just one. These three sets of keys and queries are assigned three different weights. These three weights are then matrix-multiplied with the three values to produce three outputs. These three attentions are concatenated to produce a final attention output.

[0173] Analysis: This technical solution innovatively proposes a multimodal deep learning positioning method that integrates Bluetooth AoA phase features and RSSI intensity features. By constructing a heterogeneous data joint processing framework, the linear discriminant analysis method is used to dynamically reduce the RSSI dimension to eliminate signal redundancy, and the particle filter is combined to achieve anti-interference optimization processing of AoA data. The Transformer architecture based on the multi-head self-attention mechanism is innovatively introduced to extract RSSI path attenuation features and AoA spatial pointing features in parallel feature channels respectively. The cross-attention mechanism is used to achieve deep fusion of cross-modal features, and finally a decision layer with a Softmax temperature adjustment strategy is constructed. In actual measurements in office parks, the scene positioning error is reduced to 0.38 meters, which is higher than the traditional single-mode method. At the same time, relying on TVM compilation optimization, the real-time inference performance of 17.2FPS on the edge is achieved, providing solutions with both algorithm innovation and engineering feasibility for scenarios such as high-precision positioning of Industry 4.0 and tracking of smart medical equipment.

Claims

1. A Bluetooth indoor positioning method based on Bluetooth AOA and RSSI, characterized by: The following steps are involved: Step 1: Collect Bluetooth angle of arrival (AOA) and received signal strength indicator (RSSI) data, and pre-process them through linear discriminant analysis and particle filtering to obtain training data for positioning fingerprints. Step 2: Use the encoder and decoder layers of the Transformer to extract features and use the self-attention mechanism for feature splicing. Then, use the KNN algorithm for classification learning to obtain the Bluetooth positioning model. Step 3: Measure the arrival angle information and received signal strength indicator of the Bluetooth at the target location, perform the operation preprocessing of step 1, and bring the preprocessed data into the Bluetooth location positioning model to obtain the position estimation result of the target location.

2. The Bluetooth indoor positioning method based on Bluetooth AOA and RSSI according to claim 1, characterized in that: Step 1 specifically includes the following steps: Step 11: Use linear discriminant analysis to reduce the dimension of received signal strength knowledge; Step 12: Optimize the original arrival angle data through the particle filter algorithm and normalize it to form a positioning fingerprint.

3. The Bluetooth indoor positioning method based on Bluetooth AOA and RSSI according to claim 2, characterized in that: In step 11, the linear discriminant analysis method includes the following steps: (1) Data standardization and category definition; 1. Construct N×M dimensional raw data matrix D for RSSI fingerprint library R , where N is the total number of fingerprint samples, M is the RSSI signal dimension; define the data category label set as {y n |y n ∈{1,2,...,C}}, C is the total number of categories divided into the positioning area; 2. Calculate the global mean vector μ and the category mean vector μ for each dimension c : Among them, R n is the RSSI signal vector of the nth sample; 3. Generate a standardized matrix Its nth sample is: Where Σ is the covariance matrix; (2) Scatter matrix calculation; Calculate the intra-class scatter matrix S w : Calculate the inter-class scatter matrix S b : (3) Generalized Rayleigh quotient optimization; Solve the generalized eigenvalue problem: S b w i =λ i S w w i (i=1,2,…,d) Select the eigenvectors corresponding to the first d largest eigenvalues to form the projection matrix W = [w1,w2…,w d ], where d≤C-1, the maximum number of dimensions that can be reduced by LDA is determined by the number of categories; (4) Data projection and dimensionality reduction Project the normalized data into a low-dimensional discriminant space: (5) Dimensionality reduction selection The value of d is selected based on the classification performance verification, or d=C-1 is preset to retain the maximum discriminant information.

4. The Bluetooth indoor positioning method based on Bluetooth AOA and RSSI according to claim 2, characterized in that: In step 12, the smoothing process of the particle filter includes the following steps: (1) Particle initialization and state modeling (2) State transition and prediction (3) Observation likelihood and weight update (4) Resampling and error control (5) State estimation and output.

5. The Bluetooth indoor positioning method based on Bluetooth AOA and RSSI according to claim 4, characterized in that: In step (1), particle initialization and state modeling include the following steps: State vector definition: The particle state vector includes position, direction and dynamic noise parameters, that is, Where (x, y) is the two-dimensional coordinate, θ is the motion direction angle, v is the velocity, and ω is the angular velocity noise term; Initial distribution generation: Generate N particles in the initial positioning area, and the position component follows the Gaussian distribution N(μ p ,Σ p ), where μ p is the initial estimated position, Σ p Obtained by mapping the covariance matrix of the RSSI fingerprint library.

6. The Bluetooth indoor positioning method based on Bluetooth AOA and RSSI according to claim 4, characterized in that: In step (2), state transfer and prediction include the following steps: Motion model construction: Pedestrian dead reckoning (PDR) model is used to update particle states: Where f(·) is the nonlinear motion equation, Q is the process noise covariance matrix, which includes step error and azimuth noise. The step error adopts the RSSI dynamic attenuation model, and the azimuth noise adopts the AoA phase jitter model. Dynamic noise compensation: Dynamically adjust noise parameters according to Bluetooth signal strength: oh (i) =α*RSSI var +β*AoA std Among them, α and β are trainable coefficients, RSSI var and AoA std are RSSI variance and AoA standard deviation respectively.

7. The Bluetooth indoor positioning method based on Bluetooth AOA and RSSI according to claim 4, characterized in that: In step (3), the observation likelihood and weight update includes the following steps: RSSI observation model: The logarithmic path loss model is used to calculate the similarity between particles and the fingerprint library: in, is the RSSI similarity between the i-th particle and the fingerprint library, R m is the RSSI value of the mth signal source actually measured, is the RSSI value of the predicted mth signal source at the position of the i-th particle, P0 is the signal strength at the reference distance, x (i) is the predicted coordinate of the i-th particle, b m is the actual coordinate of the mth signal source; AoA observation model: Calculate the angle matching probability based on the antenna array phase difference: in, is the phase difference received by the antenna array, θ meas is the actual measured arrival angle, σ θ is the AoA measurement noise standard deviation; Joint weight calculation: where σ R and σ θ are the measurement noise standard deviations of RSSI and AoA, respectively.

8. The Bluetooth indoor positioning method based on Bluetooth AOA and RSSI according to claim 4, characterized in that: In step (4), resampling and error control include the following steps: Adaptive resampling strategy: When the number of effective particles When , the system resampling is performed, the high-weight particles are retained and the positions of the low-weight particles are randomly disturbed, and the disturbance amplitude is determined by Σ p Dynamic adjustment, where ω (i) is the weight of the i-th particle, N th is the preset threshold for triggering resampling; Introducing random linear crossover to enhance diversity: Randomly select two particles to generate a new particle x new =γx a +(1-γ)x b , where γ~U(0,1).

9. The Bluetooth indoor positioning method based on Bluetooth AOA and RSSI according to claim 4, characterized in that: In step (5), state estimation and output include the following steps: Weighted fusion estimation: The final position estimate is the weighted mean of the particle set: in is the weight of the i-th particle at time k, is the state vector of the i-th particle at time k; Covariance Update: Calculate the posterior covariance matrix for dynamic noise adjustment: This covariance will be fed back to the process noise model in step (2) to form a closed-loop optimization.

10. The Bluetooth indoor positioning method based on Bluetooth AOA and RSSI according to claim 1, characterized in that: In the second step, the method for constructing the Bluetooth location positioning model specifically includes the following steps: Step 21: Multimodal feature embedding and fusion 1. AoA feature high-dimensional mapping: Input the original AoA angle data into the fully connected layer FC AoA , generating a high-dimensional embedding vector where d h is the hidden layer dimension; 2. RSSI feature high-dimensional mapping: Input RSSI intensity data into an independent fully connected layer FC RSSI , generating a high-dimensional embedding vector 3. Cross-modal splicing: E along the feature dimension AoA With E RSSI Perform tensor splicing to obtain joint embedding Step 22: Multi-head self-attention feature extraction; Using multi-head self-attention mechanism to joint Conduct context-aware modeling: H attn =LayerNorm(E joint +Dropout(MHSA(E joint ))) in, Dropout performs random regularization on the output of multi-head self-attention, and LayerNorm performs layer normalization on the output of residual connection; MHSA consists of h independent attention heads, each of which performs scaled dot product attention operation: Among them, Q is the query vector, K is the key vector, and V is the value vector. is the scaling factor, which scales the dot product result of Q and K to avoid training instability; Step 23: The output of the self-attention mechanism is passed through a fully connected layer for feature fusion and learned through KNN to obtain a Bluetooth location positioning model.

Images