A bluetooth AOA base station self-calibration positioning method and system based on deep learning MUSIC optimization

Abstract

The application belongs to the technical field of indoor wireless positioning, Bluetooth AOA angle of arrival positioning and array signal processing, and discloses a Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization, wherein a signal covariance matrix is adaptively constructed, multipath false spectrum peaks are automatically filtered, and accurate AOA angles of each base station are output; based on the signal confidence output by deep learning, the positioning error index of each base station is updated in real time, and dynamic positioning weights are allocated; cross circle positioning is carried out in combination with the angles of arrival of multiple base stations, the intersection point with the largest number of intersections is selected as a label coordinate, base station coordinates are synchronously and iteratively updated, and online self-calibration of the base station is completed; the application solves the problem that the traditional Bluetooth AOA positioning is susceptible to indoor multipath interference, array hardware phase drift and coordinate offset caused by base station layout, thereby continuously attenuating positioning accuracy, and manual repeated movement for calibration labels is not needed, so that all-weather adaptive base station calibration can be realized, and indoor positioning stability and positioning accuracy are improved.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the fields of indoor wireless positioning, Bluetooth AOA angle of arrival positioning, and array signal processing technology, and particularly relates to a Bluetooth AOA base station self-calibration positioning method and system based on deep learning MUSIC optimization. Background Technology

[0002] In the field of IoT smart spatial positioning technology, AOA (Angle-of-Arrival) is the mainstream high-precision indoor positioning solution. Bluetooth 5.1 and above, equipped with a CTE (Constant Tone Extended Channel), can collect carrier phase information through an array antenna to calculate the signal incident angle. It then uses multi-base station triangulation to solve the tag's three-dimensional coordinates. With its advantages of low power consumption, low cost, and easy deployment, it is widely used in indoor scenarios such as smart campuses, indoor warehouses, and building security. The MUSIC (Multi-Signal Classification) algorithm is a classic array angle measurement algorithm, often used for Bluetooth AOA phase angle of arrival calculation due to its ultra-high angle resolution.

[0003] Existing traditional Bluetooth AOA+MUSIC positioning solutions have several inherent drawbacks: First, indoor environments are affected by walls, furniture, and foot traffic, resulting in numerous reflected multipath signals. Coherent multipath components can cause spatial spectral peak splitting and generate a large number of pseudo-peaks in the MUSIC algorithm, easily leading to errors in angle of arrival calculation. Second, Bluetooth array antenna hardware is affected by ambient temperature and crystal oscillator drift, resulting in continuous timing phase deviations. Fixed-parameter MUSIC algorithms cannot dynamically compensate for hardware drift errors, and positioning accuracy will continuously decrease over time. Third, after base station deployment, slight spatial coordinate shifts can occur due to loose installation or environmental disturbances. Traditional solutions rely on manual periodic on-site relocation of calibration tags to calibrate the coordinates of each base station, resulting in a large calibration workload and an inability to achieve 24 / 7 online self-calibration. Fourth, traditional MUSIC requires searching for spectral peaks across the entire angle domain, incurring high computational overhead and making it difficult to deploy in a lightweight manner on edge embedded gateways.

[0004] Current improvements to Bluetooth AOA positioning mostly focus on optimizing Bluetooth module hardware power consumption, modifying base station structure and shape, and simply adjusting antenna array arrangement. They lack optimization at the solution level that integrates deep learning with multipath suppression, phase drift correction, and online self-calibration of the MUSIC algorithm. They cannot simultaneously solve the three core problems of multipath interference, hardware phase timing drift, and long-term base station coordinate offset. Ultimately, this results in poor 3D positioning accuracy and insufficient system stability for indoor target terminals, making it difficult to meet the long-term continuous high-precision positioning requirements in smart campus scenarios. Therefore, there is an urgent need to design a new positioning solution method.

[0005] Based on the above analysis, the urgent technical problems that need to be solved in the existing technology are:

[0006] (1) Most of the existing improvement schemes for Bluetooth AOA positioning focus on optimizing the hardware power consumption of Bluetooth modules, modifying the structure and shape of base stations, and simply adjusting the antenna array arrangement. They lack optimization at the solution level that combines deep learning to integrate multipath suppression, phase drift correction, and online self-calibration of the MUSIC algorithm.

[0007] (2) It is impossible to solve the three core problems of multipath interference, hardware phase timing drift and long-term offset of base station coordinates at the same time. Ultimately, the accuracy of three-dimensional positioning of indoor target terminals is poor and the system stability is insufficient, making it difficult to meet the long-term continuous high-precision positioning requirements in smart campus scenarios. Therefore, it is urgent to design a new positioning solution method. Summary of the Invention

[0008] To address the problems of existing technologies, this invention provides a Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization. By preprocessing phase time-series data using a dual-channel CNN-LSTM deep learning network to filter out multipath clutter and dynamically compensate for hardware phase drift, and employing a deep-enhanced DA-MUSIC network to adaptively solve for the accurate AOA angle of arrival, combined with dynamic weighting of signal confidence, online iterative self-calibration of the base station is achieved. This overcomes the shortcomings of traditional Bluetooth AOA positioning, such as severe multipath interference, inability to compensate for hardware phase drift, the need for periodic manual calibration of the base station, and continuous decay of positioning accuracy, significantly improving indoor 3D positioning accuracy and long-term system stability.

[0009] This invention is implemented as follows: A Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization includes:

[0010] S1, Data Acquisition:

[0011] At least three Bluetooth 5.1 AOA array base stations and several known coordinate calibration tags are deployed indoors. The array antennas continuously collect the raw phase timing data of CTE carrier IQ and synchronously collect the RSSI signal strength of each channel to construct the raw phase timing dataset.

[0012] S2, Multipath Interference Deep Learning Preprocessing:

[0013] A dual-channel CNN-LSTM hybrid neural network is used to process the original phase dataset; the spatial convolution channel extracts the phase spatial features of the array antenna to distinguish between direct Bluetooth signals and wall-reflected multipath clutter; the temporal LSTM channel captures the phase temporal drift features and outputs the hardware phase deviation dynamic correction coefficient to obtain the pure phase difference sequence after filtering multipath.

[0014] S3, Depth-enhanced DA-MUSIC Angle of Arrival Calculation:

[0015] The corrected phase sequence is input into the DA-MUSIC deep learning network. The array hardware error is pre-compensated through the recurrent network, the signal covariance matrix is ​​adaptively calculated, multipath pseudo-spectral peaks are automatically filtered, and the accurate AOA angle of arrival for each base station is output.

[0016] S4, Dynamically Weighted Base Station Self-Calibration and Coordinate Calculation:

[0017] The positioning error index of each base station is updated in real time based on the signal confidence output of deep learning, and dynamic positioning weights are assigned. Cross-circle positioning is performed by combining the angle of arrival of multiple base stations, and the intersection point with the most intersections is selected as the label coordinates. The base station coordinates are updated synchronously and iteratively to complete the online self-calibration of the base station.

[0018] Furthermore, the data acquisition in step S1 involves deploying at least three Bluetooth 5.1 AOA array base stations and several known coordinate calibration tags indoors. The array antennas continuously acquire the raw phase timing data of the CTE carrier IQ and synchronously acquire the RSSI signal strength of each channel to construct the raw phase timing dataset, including the following sub-steps:

[0019] Step S101 Spatial Layout Planning and Hardware Deployment:

[0020] Within the indoor positioning area, at least three Bluetooth 5.1 AOA array receiving base stations are deployed in a triangular network configuration, with the distance between any two base stations controlled between 3m and 12m to ensure full spatial coverage and enable multi-angle cross-angle measurement. Several passive Bluetooth calibration tags are deployed on the ground, walls, and desktops in the space. The three-dimensional spatial coordinates of each calibration tag are accurately measured in advance using a total station and entered into the system. These coordinates are used as the true value tag for the dataset. All AOA base stations are equipped with uniform linear array receiving antennas, and the Bluetooth CTE (Constant Tone Extended) phase angle measurement function is enabled in hardware.

[0021] Step S102 Signal periodic broadcast and synchronization timing configuration:

[0022] The calibration tag is set to continuously broadcast Bluetooth CTE modulated carrier signals with a fixed transmission period of 20ms. All receiving base stations are connected to the same clock synchronization module to complete the timing alignment of multiple base stations, ensuring that all array antennas synchronously sample and receive the wireless signals transmitted by the tag, and avoiding phase timing deviations caused by asynchronous sampling.

[0023] Step S103 Parallel acquisition of multi-channel raw sensor data:

[0024] Each base station has multiple receiving antennas in its array, which synchronously collect the original phase timing data of the incident Bluetooth carrier in IQ quadrature sampling frame by frame. At the same time, each channel of the antenna independently collects the RSSI received signal strength value of the corresponding link. The IQ phase sequence of all channels of the array and the RSSI value of each channel at the same sampling time are packaged and cached to generate a single frame sampling sample.

[0025] Step S104 Multi-scene traversal sampling and dataset annotation construction:

[0026] The study sequentially switched between three typical indoor propagation scenarios: ① an open, unobstructed scene; ② a scene occluded by static obstacles such as tables, chairs, and walls; and ③ a scene occluded by dynamic obstacles due to frequent movement of people. Sufficient temporal samples were continuously collected for each scenario, and each sampled frame was bound and labeled with the known true 3D coordinates of the corresponding calibration label at the current moment. All labeled samples from all scenarios were then aggregated and uniformly divided into training, validation, and test sets. Finally, a raw phase temporal dataset with ground truth labels was constructed for subsequent training and inference of the dual-channel CNN-LSTM network.

[0027] Furthermore, the formula for calculating the basic phase difference arrival angle involved in the spatial layout and timing synchronization steps in step S1 includes the following sub-steps:

[0028] In step S105, three Bluetooth 5.1 AOA array receiving base stations (base station 1, base station 2, and base station 3) are deployed indoors in a triangular configuration. All base stations are connected to the same clock synchronization module to achieve timing alignment. The deployed passive Bluetooth calibration tags continuously broadcast Bluetooth CTE modulated carrier signals at a fixed period of 20ms. Each of the three base stations is equipped with a uniform linear array receiving antenna array to synchronously collect the same source carrier signals.

[0029] For the linear array antenna inside base station 1, two groups of antenna elements are selected for angle calculation: Let the first group of antennas inside base station 1 be antenna... ,antenna The center-to-center distance between the two antennas is The second group of antennas is an antenna. ,antenna The center-to-center distance between the two antennas is Bluetooth signals from the same source arriving at different antennas will experience phase shifts due to inconsistent transmission path lengths. The complementary angle of signal incidence can be calculated using the phase difference.

[0030] The formula for calculating the direction angle is:

[0031] in: : The wavelength of the CTE Bluetooth carrier signal emitted by the calibration tag;

[0032] These correspond to the pairing of two sets of antennas within base station 1;

[0033] :antenna With antenna The phase difference value of received signals from the same source;

[0034] :antenna With antenna The phase difference value of received signals from the same source;

[0035] :antenna , The center distance between them;

[0036] :antenna , The center distance between them;

[0037] :based on , The calculated signal incidence co-angle is 1.

[0038] :based on , The calculated signal incidence co-angle is 2;

[0039] Similarly, following the exact same formula and calculation logic, the complementary angle of incidence corresponding to base station 2 is calculated respectively. The incident complementary angle corresponding to base station 3 ;

[0040] Step S106: Retrieve the coordinates of base station 1 pre-calibrated by the total station. Base station 2 coordinates 3-coordinate base station and the three-dimensional coordinates of the calibration tag true value A global three-dimensional virtual model was built by combining the length, width and height dimensions of the indoor space, and the positions of the three base stations and all the antenna units inside the model were marked.

[0041] The incident complementary angle obtained using base station 1 Determine the spatial incident direction of the calibration tag relative to base station 1; in Within its plane, with antenna Draw a circular trajectory with the center of the circle and the incident direction as the radial axis, denoted as circle. ;

[0042] The incident complementary angle obtained using base station 1 Determine the calibration tag relative to the antenna The spatial incident direction; in Within its plane, with antenna Draw a circular trajectory with the center of the circle and the incident direction as the radial axis, denoted as circle. Simultaneously repeat the above circle-drawing operation on base stations 2 and 3 to generate circles sequentially. and circle Gradually and synchronously increase the radius values ​​of all circular trajectories to obtain the circle. With circle The common straight line obtained by the intersection ; Antenna Incident ray and straight line towards the calibration label The intersections yield the initial intersection point from the perspective of base station 1. ;

[0043] The intersection points obtained from solving the three base stations individually To conduct spatial intersection, select the spatial point where the three rays intersect with the highest degree of overlap. This point is the real-time calculated coordinate of the calibration tag; the calculated coordinates with true coordinates The comparison yields the basic angle measurement error, which is used for subsequent deep learning dataset annotation.

[0044] Furthermore, the S2 multipath interference deep learning preprocessing step includes the following sub-steps:

[0045] Step S201: Construct the network input matrix and temporal sample tensor with dimension annotations.

[0046] Read the raw IQ phase data and RSSI data of all array antenna channels of base station 1, base station 2, and base station 3 respectively. Define: base station number as... Each base station has an internal antenna channel numbered as follows: ;set up For the first Taiwan base station, the first The root antenna at time The original acquired phase value, This refers to the RSSI signal strength value synchronously acquired for the corresponding channel;

[0047] The phase and RSSI data of all base stations and all antennas at a single moment are concatenated in spatial order to construct a single-moment spatial input matrix. Then stack them continuously. A time-series sampling frame is used to construct the input tensor of the four-dimensional network. ,in This refers to the number of sample batches. For timing step size, The total number of base stations. The number of antenna channels per base station is used as the sole input to a dual-channel CNN-LSTM hybrid neural network.

[0048] Step S202 CNN Spatial Convolution Channels: Spatial Phase Feature Extraction and Multipath Detection

[0049] The training labels for the multipath interference confidence coefficients are generated by the direct geometric phase corresponding to the calibration label ground truth coordinates, the measured phase residual, the RSSI drop, the phase consistency of adjacent array elements, and the multi-base station intersection residual. The training loss includes at least the multipath classification loss, the phase consistency loss, and the timing / channel smoothing loss.

[0050] For the input matrix Two-dimensional convolution operations are performed to extract the spatial phase distribution features between base stations and between antenna channels, and the spatial feature map is output. Define the spatial characteristic threshold of the direct signal. Multipath clutter characteristic threshold ;

[0051] For arbitrary channel phase characteristics :like If the channel is determined to be a direct and valid signal; It is determined to be a weak multipath interference signal; if The signal was determined to be invalid due to wall reflection or blockage of clutter.

[0052] Simultaneously output the confidence coefficient of multipath interference for each channel. , , used to characterize the Taiwan base station The purity of the channel signal is used to complete the screening of multipath noise and feature purification in the spatial dimension;

[0053] Step S203 LSTM Timing Channel: Phase Timing Drift Fitting and Hardware Bias Coefficient Solving

[0054] The dynamic hardware phase deviation correction coefficient is updated iteratively using the geometric phase residual derived from the total station's true coordinates as the supervisory true value, or using the geometric consistency between the compensated phase, angle of arrival, and multi-base station intersection coordinates as a self-supervised constraint.

[0055] Continuous time phase sequence The input LSTM timing network is fitted with the timing phase drift caused by antenna hardware temperature drift, crystal oscillator error and environmental disturbance through a triple gating structure of input gate, forget gate and output gate.

[0056] Definition of the first base station The dynamic hardware phase deviation correction factor for the root antenna is The network iteratively updates all data based on the temporal feature loss. It outputs the independent phase deviation compensation value for each antenna channel in real time, realizing the dynamic quantification of hardware system errors;

[0057] Step S204: Multipath filtering + phase compensation to generate a globally clean phase difference sequence.

[0058] Using timing deviation coefficient The original phase value is corrected channel by channel to obtain the compensated accurate phase: ;

[0059] Combining multipath confidence coefficients Weighted filtering is applied to anomalous phase points to eliminate phase abrupt noise caused by wall reflections and multipath occlusion; the corrected pure phase is then filtered. Perform channel differential calculations to solve for the pairwise phase differences between the internal antennas of base station 1, base station 2, and base station 3. , ;

[0060] The final output completely filters out multipath clutter, eliminates timing drift, and compensates for hardware errors. The global clean phase difference sequence matrix serves as the precise input data for subsequent DA-MUSIC network angle of arrival calculation.

[0061] Furthermore, step S3, the depth-enhanced DA-MUSIC angle of arrival calculation step, includes the following sub-steps:

[0062] Step S301 DA-MUSIC network input loading and array hardware error pre-compensation:

[0063] The global pure phase difference sequence matrix output in step S2 As dedicated input data for the DA-MUSIC deep learning network, the base station number is defined. The phase difference between paired antenna channels in a single base station is Hardware phase deviation correction coefficients for each antenna channel obtained based on S2 Construct a global hardware error compensation matrix for the base station array. The pure phase sequence is pre-corrected a second time using an error compensation matrix to eliminate fixed system errors caused by factory inconsistency deviations and installation tilt deviations of the array antenna, thus obtaining the final corrected phase dataset. ;

[0064] Step S302: Construction of the adaptive signal covariance matrix:

[0065] Based on the correction phase dataset Construct the received signal vector of a single base station array ,in For the first Taiwan base station The array signal column vector at time step; abandoning the traditional MUSIC algorithm's fixed covariance matrix construction method, the real-time signal covariance matrix is ​​obtained through adaptive iterative computation using the DA-MUSIC network: In the formula, For the first The array signal covariance matrix of the base station, For mathematical expectation calculation, For matrix conjugate transpose operation, it accurately adapts to signal characteristic changes in indoor dynamic multipath interference scenarios;

[0066] Step S303: Feature Space Decomposition and Multipath Pseudo-Peak Suppression

[0067] The covariance matrix obtained by adaptive solution Perform eigenvalue decomposition to obtain the eigenvalue set. Based on the corresponding eigenvector matrix, the signal subspace and noise subspace are distinguished according to the magnitude of the eigenvalues; an effective signal eigenvalue threshold is defined. , will be greater than The eigenvalues ​​corresponding to the vectors are classified into the signal subspace. less than or equal to The eigenvalues ​​corresponding to the vectors are classified into the noise subspace. To address the issues of spectral peak splitting and spurious peak interference caused by indoor coherent multipath signals, the noise subspace matrix is ​​adaptively corrected through network weights to suppress multipath spurious peaks derived from wall reflections and occlusions, while preserving the spectral peak characteristics of the true direct signal.

[0068] Step S304: Deep learning regression to accurately solve the AOA (Angle of Arrival) for each base station.

[0069] The DA-MUSIC network relies on an optimized signal subspace. Noise subspace At the same time, the array antenna spacing parameter is introduced. and radio frequency signal wavelength This type of array structure uses prior information; it completely abandons the traditional MUSIC algorithm's method of solving angles by searching the spatial spectrum peaks through a global angle traversal. Instead, it leverages the powerful nonlinear mapping and fitting capabilities of deep neural networks to complete regression calculations, enabling it to directly predict the two sets of high-precision signal incidence complementary angles corresponding to the m-th receiving base station. The angle of arrival symbols for each base station antenna array are defined as follows: This represents the angle of arrival of the signal calculated by each of the two antenna arrays at base station 1. The angle of arrival of the signal is calculated by each of the two antenna arrays at base station No. 2. The angle of arrival of the signal is calculated by each of the two antenna arrays at base station No. 3.

[0070] While outputting the angle estimates, the network simultaneously generates the confidence coefficient for each set of angle solutions, expressed as follows: The coefficient ranges from 0 to 1, and its value can quantitatively characterize the reliability and estimation accuracy of the AOA angle solution results of a single base station. The confidence value will serve as a key weight parameter to support the subsequent dynamic weighted adaptive self-calibration iteration process of base station coordinates, effectively weaken the error interference caused by low-precision angle observations, and continuously improve the overall spatial positioning stability.

[0071] The DA-MUSIC network includes a pseudo-peak suppression head, an angle regression head, and a confidence output head. The pseudo-peak suppression head generates a pseudo-peak mask based on the spatial spectrum peak value, eigenvalue interval, multipath confidence, and phase deviation. The angle regression head outputs the angle of arrival, and the confidence output head outputs the angle confidence that is negatively correlated with the angle error.

[0072] Furthermore, step S4, the dynamic weighted base station self-calibration and coordinate calculation step, includes the following sub-steps:

[0073] Step S401: Solving for the dynamic error exponent and real-time weighting coefficients:

[0074] The confidence coefficient is calculated based on the angles output by each base station in step S3 above. Construct two sets of dynamically adjustable weights that can be updated in real time. ;in The weighting coefficient representing the positional deviation Characterizing the angular error weighting coefficient; All can be based on the confidence coefficient of each base station Real-time assignment and adjustment are completed, and the value range of both is uniformly limited to the closed interval [0,1].

[0075] Set variables For the first Cumulative historical location deviation for each base station For the first The instantaneous error of the angle calculation at the current moment of each base station is obtained by substituting the above variables into the dynamic error exponent calculation formula. Dynamic error index of base station:

[0076]

[0077] Dynamic error index obtained for all base stations After performing normalization, the specific positioning weighting coefficients for each base station are derived, and the calculation formula is as follows:

[0078]

[0079] The normalized base station positioning weights satisfy the normalization constraint:

[0080]

[0081] The weighting follows these rules: the smaller the dynamic error index of a base station, the higher the confidence level of its angle calculation, and the larger the corresponding spatial positioning weighting coefficient. Conversely, base stations with large error indices and low angle calculation accuracy will be assigned smaller positioning weights to mitigate the negative impact of low-precision observation data on the overall positioning results.

[0082] Step S402: Solve for the initial coordinates of the tag using multi-base station AOA angle ray cross-positioning.

[0083] Retrieve the calibration spatial coordinates of three base stations , , Combined with the precise angle of arrival of each base station A system of two-dimensional angle incident ray equations is constructed for each base station; the predicted intersection points of the tags corresponding to base station 1, base station 2, and base station 3 are solved using a spatial geometric intersection algorithm. , , The spatial overlap and ray intersection count of the three sets of intersection points were statistically analyzed, and the intersection point with the highest overlap and smallest adaptation angle error was selected as the initial estimated coordinates of the calibration label. ;

[0084] Step S403: Calculation of positioning residuals and solution of base station coordinate deviation:

[0085] Retrieve the true coordinates of the calibration labels pre-calibrated by the total station. The initial estimated coordinates are compared with the true coordinates to solve for the 3D spatial positioning residual vector.

[0086] ,

[0087] Combining the dynamic weights of each base station The offset error of the current spatial coordinates of each base station is derived in reverse. The error allocation rule is as follows: the smaller the weight and the larger the error exponent of the base station, the larger the coordinate offset correction amount is allocated, so as to accurately match the real-time positioning error status of each base station.

[0088] Step S404 Iteratively update the base station coordinates to achieve online adaptive self-calibration:

[0089] Based on the base station coordinate offset error obtained from the solution Combined with preset iterative correction step size The coordinates of the three base stations are updated iteratively in real time, and the update formula is as follows: ,in For the first The updated calibration coordinates of the base station are continuously executed in a loop, encompassing data acquisition, phase preprocessing, angle calculation, and residual correction. The base station coordinate parameters are iteratively optimized until the spatial positioning residual is corrected. To preset the positioning accuracy threshold, base station convergence calibration is completed. This step does not require manual re-deployment or movement of calibration tags. It relies on time-series deep learning for continuous iterative correction to achieve 24 / 7 online adaptive calibration of the base station. Finally, the coordinates of the calibrated base station are used to achieve high-precision and stable positioning of the Bluetooth tag under test.

[0090] The base station coordinate offset error is determined by the tag positioning residual, the Jacobian matrix of the angle observation function on the base station coordinates, and the angle confidence weight. The residual is then distributed as the base station coordinate correction amount through weighted least squares or gradient descent.

[0091] Another objective of this invention is to provide a Bluetooth AOA base station self-calibration positioning system based on deep learning MUSIC optimization, comprising:

[0092] The data acquisition module is used for data acquisition.

[0093] The preprocessing module is used for deep learning preprocessing of multipath interference;

[0094] Angle of arrival calculation module, used for depth-enhanced DA-MUSIC angle of arrival calculation;

[0095] The coordinate calculation module is used for dynamic weighted base station self-calibration and coordinate calculation.

[0096] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, the computer program being executed by the processor causing the processor to perform the steps of the Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization.

[0097] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the Bluetooth AOA base station self-calibration positioning method optimized based on deep learning MUSIC.

[0098] Another objective of this invention is to provide an information data processing terminal for implementing the Bluetooth AOA base station self-calibration positioning system optimized based on deep learning MUSIC.

[0099] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0100] This invention employs a dual-channel CNN-LSTM hybrid neural network to perform parallel processing on the raw Bluetooth phase timing data. The spatial convolution channel distinguishes between the direct Bluetooth signal and multipath clutter caused by walls and human movement based on the spatial distribution characteristics of the array antenna phase. The temporal LSTM channel continuously captures the phase timing drift caused by hardware crystal oscillator temperature drift and environmental disturbances, and solves the hardware phase deviation correction coefficient in real time. Based on the correction coefficient, the raw IQ phase data is dynamically compensated frame by frame and channel by channel. The advantage of this is that it can eliminate indoor multipath interference noise layer by layer from the signal acquisition source and compensate for the system phase error caused by long-term operation of the array antenna. It completely avoids the shortcomings of traditional fixed parameter algorithms that cannot adapt to the dynamic and changing indoor environment and the continuous accumulation of hardware drift error, effectively improving the purity of the raw phase data and laying a reliable data foundation for subsequent accurate angle of arrival calculation.

[0101] This invention also replaces the traditional classic MUSIC algorithm's fixed covariance matrix construction and global angle traversal peak finding operation logic with a deep enhanced DA-MUSIC deep learning network. The network can adaptively construct the time-varying signal covariance matrix based on real-time phase data, autonomously complete feature space decomposition and correct noise subspace weights, and specifically suppress the problems of spectral peak splitting and pseudo-spectral peak interference induced by coherent multipath signals. Finally, it directly outputs the accurate AOA of each base station and the angle calculation confidence by relying on the nonlinear regression of the neural network. The advantage of this is that it eliminates the redundant computational overhead brought by the large-scale angle traversal of traditional MUSIC, and at the same time eliminates the angle measurement deviation caused by multipath clutter, which greatly improves the accuracy of Bluetooth signal incident angle of arrival measurement and reduces the probability of angle calculation error.

[0102] This invention also dynamically assigns weight coefficients to the angle confidence output by deep learning, calculates the positioning error index of each base station in real time and allocates differentiated positioning weights, integrates the arrival angles of multiple base stations to perform three-dimensional spatial angle intersection to solve the tag coordinates, compares the calculated coordinates with the true value coordinates of the tags calibrated by the total station to obtain the positioning residual, and iteratively corrects the three-dimensional deployment coordinates of all receiving base stations. The entire process does not require manual repetition of moving calibration tags to perform calibration work. Relying on time series data for continuous iteration, it achieves 24 / 7 all-weather online adaptive self-calibration of base stations. The advantage of this is that it can offset the coordinate offset error caused by loose installation, external force disturbance, and environmental deformation of base stations in real time, dynamically adapt to the long-term operating status of the system, eliminate the maintenance cost of manual on-site calibration, and significantly improve the fault tolerance and long-term stability of the entire positioning system in actual deployment and operation.

[0103] In addition, this invention performs lightweight pruning and optimization of CNN-LSTM and DA-MUSIC networks, compressing the number of model parameters and computational load. It can be directly deployed in edge computing embedded gateways in smart campus scenarios to complete local real-time inference operations without relying on cloud-based high-performance servers for data processing, thus reducing data transmission latency and deployment costs. The positioning accuracy decays very little under long-term uninterrupted operation of the entire system, maintaining centimeter-level positioning effects in both static indoor scenes and dynamic pedestrian scenarios. This comprehensively improves the positioning accuracy, environmental adaptability, engineering practicality, and robustness of the Bluetooth AOA indoor 3D positioning system.

[0104] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0105] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0106] This invention targets high-frequency indoor positioning scenarios such as smart campuses, indoor warehouses, hospital buildings, factory parks, asset tracking, and personnel safety management. Based on existing Bluetooth 5.1 AOA array base station hardware, it improves positioning accuracy and continuous operational stability through CNN-LSTM phase preprocessing, DA-MUSIC angle-of-arrival calculation, and a dynamic weighted base station self-calibration process. Compared to traditional solutions relying on manual periodic calibration, this invention reduces the costs of repeated tag movement, manual on-site measurement, downtime debugging, and subsequent maintenance, lowering the frequency of system deployment and maintenance. Simultaneously, the lightweight deep learning model can be deployed locally for inference on the edge gateway, reducing cloud transmission latency and server dependence, facilitating the creation of scalable indoor positioning products. Because this invention can maintain centimeter-level positioning capabilities in complex real-world environments such as static occlusion, personnel movement, and minor base station offsets, its commercial value lies not only in the sale of positioning terminals but also in ongoing service revenue from positioning platform services, campus security management, warehouse route optimization, asset inventory, personnel trajectory analysis, and building digital twins.

[0107] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0108] Existing Bluetooth AOA positioning improvement schemes mostly focus on optimizing Bluetooth module power consumption, modifying antenna array structure, adjusting single MUSIC algorithm parameters, or optimizing offline calibration processes. These typically only address one problem at a time: multipath interference, phase drift, or base station installation errors. They lack a technical solution that unifies multipath identification, hardware phase drift compensation, MUSIC spurious peak suppression, angle confidence output, and base station online self-calibration into a single closed-loop solution link. This invention fills this gap by using total station truth labels and raw IQ / RSSI time-series data as a foundation. A CNN outputs multipath confidence, an LSTM outputs dynamic phase deviation coefficients, and a DA-MUSIC outputs the AOA angle and angle confidence after spurious peak suppression. The label positioning residual is then back-allocated to each base station coordinate correction amount through weights, Jacobian, or gradient relationships. This technology chain transforms the Bluetooth AOA system from the traditional model of "fixed base station coordinates + fixed algorithm parameters + manual periodic calibration" to a model of "data-driven phase purification + physical constraint angle calculation + online closed-loop self-calibration", filling the systemic technical gap in the industry for long-term adaptive positioning.

[0109] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully:

[0110] Indoor Bluetooth AOA positioning has long faced three interdependent and difficult-to-solve technical challenges: First, coherent multipath propagation caused by walls, tables, chairs, and human movement can lead to MUSIC spatial spectrum peak splitting or the generation of spurious peaks, causing the estimated angle of arrival to deviate from the true direct path. Second, crystal oscillator drift, temperature changes, and array channel differences can cause timing phase deviations, leading to accuracy degradation of fixed-parameter algorithms over time. Third, loosening, collisions, or changes in the spatial environment after base station installation can cause slight shifts in base station coordinates, which traditional systems typically require manual remeasurement and recalibration. These problems often occur simultaneously in actual engineering deployments, and simply improving hardware accuracy, performing single offline calibrations, or using traditional MUSIC peak finding methods cannot provide a sustainable solution. This invention suppresses multipath anomalous phase through CNN confidence, dynamically compensates for phase drift through LSTM, outputs angle results with confidence through DA-MUSIC, and drives iterative correction of base station coordinates through label truth residuals. This enables the system to continuously converge without downtime or repeated manual repositioning of calibration labels, thus providing a complete technical path for the positioning challenge of "high accuracy, low maintenance, and long-term stability" in complex indoor environments.

[0111] (4) The technical solution of the present invention overcomes technical bias:

[0112] Traditional technical approaches typically assume that MUSIC-type subspace algorithms rely on fixed covariance matrix construction and global angle traversal for peak finding. Deep learning models are often considered difficult to directly intervene in array signal processing, as this could compromise the algorithm's physical interpretability. Furthermore, in engineering, base station coordinates are usually treated as fixed external parameters after installation, and accuracy degradation is addressed through manual retesting, calibration, or hardware replacement, rather than real-time correction during positioning calculations. This invention overcomes these technical biases: it does not completely replace array signal processing with a black-box network. Instead, it introduces interpretable confidence levels, phase deviation coefficients, pseudo-peak masks, angle confidence levels, and Jacobian residual allocation mechanisms at the phase input layer, covariance / pseudospectral processing layer, angle output layer, and coordinate self-calibration layer, respectively. This ensures that the deep learning output is constrained by geometric truth, array priors, and positioning residuals. Thus, this invention proves that deep learning and MUSIC physical models are not mutually exclusive, but can form a hybrid solution framework that is traceable, reproducible, and iteratively available online. It also breaks through the conventional idea that base station coordinates can only be manually calibrated offline, and realizes dynamic self-calibration of base station coordinates during system operation. Attached Figure Description

[0113] Figure 1 This is a flowchart of a Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization provided in an embodiment of the present invention.

[0114] Figure 2 This is a block diagram of a Bluetooth AOA base station self-calibration positioning system based on deep learning MUSIC optimization provided in an embodiment of the present invention.

[0115] Figure 3 This is a flowchart of the overall method provided in the embodiments of the present invention.

[0116] Figure 4 This is a schematic diagram of the CNN-LSTM dual-channel network structure in step S2 of the embodiment of the present invention.

[0117] Figure 5 This is a diagram of the campus classroom deployment architecture for the experimental site provided in an embodiment of the present invention.

[0118] Figure 6 This is a flowchart of another Bluetooth AOA base station self-calibration positioning method provided in an embodiment of the present invention.

[0119] Figure 7 This is another Bluetooth AOA base station self-calibration positioning system module diagram provided in an embodiment of the present invention.

[0120] Figure 8 This is a comparison chart of static and dynamic scene positioning errors provided in an embodiment of the present invention.

[0121] Figure 9 This is a cumulative distribution curve of positioning error provided in an embodiment of the present invention.

[0122] Figure 10 This is a diagram illustrating the phase drift and self-calibration suppression effect during continuous operation, as provided in an embodiment of the present invention.

[0123] Figure 11 This is a graph showing the contribution of each technical module provided in the embodiments of the present invention to the improvement of positioning accuracy.

[0124] Figure 12 This is a graph showing the suppression effect of DA-MUSIC on multipath pseudo-spectral peaks provided in this embodiment of the invention.

[0125] Figure 13 This is a graph showing the correspondence between angle confidence and positioning error provided in an embodiment of the present invention. Detailed Implementation

[0126] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0127] like Figure 1 As shown, the Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization provided by this embodiment of the invention includes the following steps:

[0128] S1, Data Acquisition:

[0129] At least three Bluetooth 5.1 AOA array base stations and several known coordinate calibration tags are deployed indoors. The array antennas continuously collect the raw phase timing data of CTE carrier IQ and synchronously collect the RSSI signal strength of each channel to construct the raw phase timing dataset.

[0130] S2, Multipath Interference Deep Learning Preprocessing:

[0131] A dual-channel CNN-LSTM hybrid neural network is used to process the original phase dataset; the spatial convolution channel extracts the phase spatial features of the array antenna to distinguish between direct Bluetooth signals and wall-reflected multipath clutter; the temporal LSTM channel captures the phase temporal drift features and outputs the hardware phase deviation dynamic correction coefficient to obtain the pure phase difference sequence after filtering multipath.

[0132] S3, Depth-enhanced DA-MUSIC Angle of Arrival Calculation:

[0133] The corrected phase sequence is input into the DA-MUSIC deep learning network. The array hardware error is pre-compensated through the recurrent network, the signal covariance matrix is ​​adaptively calculated, multipath pseudo-spectral peaks are automatically filtered, and the accurate AOA angle of arrival for each base station is output.

[0134] S4, Dynamically Weighted Base Station Self-Calibration and Coordinate Calculation:

[0135] The positioning error index of each base station is updated in real time based on the signal confidence output of deep learning, and dynamic positioning weights are assigned. Cross-circle positioning is performed by combining the angle of arrival of multiple base stations, and the intersection point with the most intersections is selected as the label coordinates. The base station coordinates are updated synchronously and iteratively to complete the online self-calibration of the base station.

[0136] The data acquisition in step S1 of this embodiment of the invention involves: deploying at least three Bluetooth 5.1 AOA array base stations and several known coordinate calibration tags indoors; continuously acquiring CTE carrier IQ raw phase timing data and synchronously acquiring RSSI signal strength of each channel to construct the raw phase timing dataset, including the following sub-steps:

[0137] Step S101 Spatial Layout Planning and Hardware Deployment:

[0138] Within the indoor positioning area, at least three Bluetooth 5.1 AOA array receiving base stations are deployed in a triangular network configuration, with the distance between any two base stations controlled between 3m and 12m to ensure full spatial coverage and enable multi-angle cross-angle measurement. Several passive Bluetooth calibration tags are deployed on the ground, walls, and desktops in the space. The three-dimensional spatial coordinates of each calibration tag are accurately measured in advance using a total station and entered into the system. These coordinates are used as the true value tag for the dataset. All AOA base stations are equipped with uniform linear array receiving antennas, and the Bluetooth CTE (Constant Tone Extended) phase angle measurement function is enabled in hardware.

[0139] Step S102 Signal periodic broadcast and synchronization timing configuration:

[0140] The calibration tag is set to continuously broadcast Bluetooth CTE modulated carrier signals with a fixed transmission period of 20ms. All receiving base stations are connected to the same clock synchronization module to complete the timing alignment of multiple base stations, ensuring that all array antennas synchronously sample and receive the wireless signals transmitted by the tag, and avoiding phase timing deviations caused by asynchronous sampling.

[0141] Step S103 Parallel acquisition of multi-channel raw sensor data:

[0142] Each base station has multiple receiving antennas in its array, which synchronously collect the original phase timing data of the incident Bluetooth carrier in IQ quadrature sampling frame by frame. At the same time, each channel of the antenna independently collects the RSSI received signal strength value of the corresponding link. The IQ phase sequence of all channels of the array and the RSSI value of each channel at the same sampling time are packaged and cached to generate a single frame sampling sample.

[0143] Step S104 Multi-scene traversal sampling and dataset annotation construction:

[0144] The study sequentially switched between three typical indoor propagation scenarios: ① an open, unobstructed scene; ② a scene occluded by static obstacles such as tables, chairs, and walls; and ③ a scene occluded by dynamic obstacles due to frequent movement of people. Sufficient temporal samples were continuously collected for each scenario, and each sampled frame was bound and labeled with the known true 3D coordinates of the corresponding calibration label at the current moment. All labeled samples from all scenarios were then aggregated and uniformly divided into training, validation, and test sets. Finally, a raw phase temporal dataset with ground truth labels was constructed for subsequent training and inference of the dual-channel CNN-LSTM network.

[0145] The definition of the formula for calculating the basic phase difference arrival angle involved in the spatial layout and timing synchronization steps in step S1 provided in this embodiment of the invention includes the following sub-steps:

[0146] In step S105, three Bluetooth 5.1 AOA array receiving base stations (base station 1, base station 2, and base station 3) are deployed indoors in a triangular configuration. All base stations are connected to the same clock synchronization module to achieve timing alignment. The deployed passive Bluetooth calibration tags continuously broadcast Bluetooth CTE modulated carrier signals at a fixed period of 20ms. Each of the three base stations is equipped with a uniform linear array receiving antenna array to synchronously collect the same source carrier signals.

[0147] For the linear array antenna inside base station 1, two groups of antenna elements are selected for angle calculation: Let the first group of antennas inside base station 1 be antenna... ,antenna The center-to-center distance between the two antennas is The second group of antennas is an antenna. ,antenna The center-to-center distance between the two antennas is Bluetooth signals from the same source arriving at different antennas will experience phase shifts due to inconsistent transmission path lengths. The complementary angle of signal incidence can be calculated using the phase difference.

[0148] The formula for calculating the direction angle is:

[0149] in: : The wavelength of the CTE Bluetooth carrier signal emitted by the calibration tag;

[0150] These correspond to the pairing of two sets of antennas within base station 1;

[0151] :antenna With antenna The phase difference value of received signals from the same source;

[0152] :antenna With antenna The phase difference value of received signals from the same source;

[0153] :antenna , The center distance between them;

[0154] :antenna , The center distance between them;

[0155] :based on , The calculated signal incidence co-angle is 1.

[0156] :based on , The calculated signal incidence co-angle is 2;

[0157] Similarly, following the exact same formula and calculation logic, the complementary angle of incidence corresponding to base station 2 is calculated respectively. The incident complementary angle corresponding to base station 3 ;

[0158] Step S106: Retrieve the coordinates of base station 1 pre-calibrated by the total station. Base station 2 coordinates 3-coordinate base station and the three-dimensional coordinates of the calibration tag true value A global three-dimensional virtual model was built by combining the length, width and height dimensions of the indoor space, and the positions of the three base stations and all the antenna units inside the model were marked.

[0159] The incident complementary angle obtained using base station 1 Determine the spatial incident direction of the calibration tag relative to base station 1; in Within its plane, with antenna Draw a circular trajectory with the center of the circle and the incident direction as the radial axis, denoted as circle. ;

[0160] The incident complementary angle obtained using base station 1 Determine the calibration tag relative to the antenna The spatial incident direction; in Within its plane, with antenna Draw a circular trajectory with the center of the circle and the incident direction as the radial axis, denoted as circle. Simultaneously repeat the above circle-drawing operation on base stations 2 and 3 to generate circles sequentially. and circle Gradually and synchronously increase the radius values ​​of all circular trajectories to obtain the circle. With circle The common straight line obtained by the intersection ; Antenna Incident ray and straight line towards the calibration label The intersections yield the initial intersection point from the perspective of base station 1. ;

[0161] The intersection points obtained from solving the three base stations individually To conduct spatial intersection, select the spatial point where the three rays intersect with the highest degree of overlap. This point is the real-time calculated coordinate of the calibration tag; the calculated coordinates with true coordinates The comparison yields the basic angle measurement error, which is used for subsequent deep learning dataset annotation.

[0162] The S2 multipath interference deep learning preprocessing steps provided in this embodiment of the invention include the following sub-steps:

[0163] Step S201: Construct the network input matrix and temporal sample tensor with dimension annotations.

[0164] Read the raw IQ phase data and RSSI data of all array antenna channels of base station 1, base station 2, and base station 3 respectively. Define: base station number as... Each base station has an internal antenna channel numbered as follows: ;set up For the first Taiwan base station, the first The root antenna at time The original acquired phase value, This refers to the RSSI signal strength value synchronously acquired for the corresponding channel;

[0165] The phase and RSSI data of all base stations and all antennas at a single moment are concatenated in spatial order to construct a single-moment spatial input matrix. Then stack them continuously. A time-series sampling frame is used to construct the input tensor of the four-dimensional network. ,in This refers to the number of sample batches. For timing step size, The total number of base stations. The number of antenna channels per base station is used as the sole input to a dual-channel CNN-LSTM hybrid neural network.

[0166] Step S202 CNN Spatial Convolution Channels: Spatial Phase Feature Extraction and Multipath Detection

[0167] For the input matrix Two-dimensional convolution operations are performed to extract the spatial phase distribution features between base stations and between antenna channels, and the spatial feature map is output. Define the spatial characteristic threshold of the direct signal. Multipath clutter characteristic threshold ;

[0168] For arbitrary channel phase characteristics :like If the channel is determined to be a direct and valid signal; It is determined to be a weak multipath interference signal; if The signal was determined to be invalid due to wall reflection or blockage of clutter.

[0169] Simultaneously output the confidence coefficient of multipath interference for each channel. , , used to characterize the base station The purity of the channel signal is used to complete the screening of multipath noise and feature purification in the spatial dimension;

[0170] Step S203 LSTM Timing Channel: Phase Timing Drift Fitting and Hardware Bias Coefficient Solving

[0171] Continuous time phase sequence The input LSTM timing network is fitted with the timing phase drift caused by antenna hardware temperature drift, crystal oscillator error and environmental disturbance through a triple gating structure of input gate, forget gate and output gate.

[0172] Definition of the first base station The dynamic hardware phase deviation correction factor for the root antenna is The network iteratively updates all data based on the temporal feature loss. It outputs the independent phase deviation compensation value for each antenna channel in real time, realizing the dynamic quantification of hardware system errors;

[0173] Step S204: Multipath filtering + phase compensation to generate a globally clean phase difference sequence.

[0174] Using timing deviation coefficient The original phase value is corrected channel by channel to obtain the compensated accurate phase: ;

[0175] Combining multipath confidence coefficients Weighted filtering is applied to anomalous phase points to eliminate phase abrupt noise caused by wall reflections and multipath occlusion; the corrected pure phase is then filtered. Perform channel differential calculations to solve for the pairwise phase differences between the internal antennas of base station 1, base station 2, and base station 3. , ;

[0176] The final output completely filters out multipath clutter, eliminates timing drift, and compensates for hardware errors. The global clean phase difference sequence matrix serves as the precise input data for subsequent DA-MUSIC network angle of arrival calculation.

[0177] The S3 depth-enhanced DA-MUSIC angle-of-arrival calculation step provided in this embodiment of the invention includes the following sub-steps:

[0178] Step S301 DA-MUSIC network input loading and array hardware error pre-compensation:

[0179] The global pure phase difference sequence matrix output in step S2 As dedicated input data for the DA-MUSIC deep learning network, the base station number is defined. The phase difference between paired antenna channels in a single base station is Hardware phase deviation correction coefficients for each antenna channel obtained based on S2 Construct a global hardware error compensation matrix for the base station array. The pure phase sequence is pre-corrected a second time using an error compensation matrix to eliminate fixed system errors caused by factory inconsistency deviations and installation tilt deviations of the array antenna, thus obtaining the final corrected phase dataset. ;

[0180] Step S302: Construction of the adaptive signal covariance matrix:

[0181] Based on the correction phase dataset Construct the received signal vector of a single base station array ,in For the first Taiwan base station The array signal column vector at time step; abandoning the traditional MUSIC algorithm's fixed covariance matrix construction method, the real-time signal covariance matrix is ​​obtained through adaptive iterative computation using the DA-MUSIC network: In the formula, For the first The array signal covariance matrix of the base station, For mathematical expectation calculation, For matrix conjugate transpose operation, it accurately adapts to signal characteristic changes in indoor dynamic multipath interference scenarios;

[0182] Step S303: Feature Space Decomposition and Multipath Pseudo-Peak Suppression

[0183] The covariance matrix obtained by adaptive solution Perform eigenvalue decomposition to obtain the eigenvalue set. Based on the corresponding eigenvector matrix, the signal subspace and noise subspace are distinguished according to the magnitude of the eigenvalues; an effective signal eigenvalue threshold is defined. , will be greater than The eigenvalues ​​corresponding to the vectors are classified into the signal subspace. less than or equal to The eigenvalues ​​corresponding to the vectors are classified into the noise subspace. To address the issues of spectral peak splitting and spurious peak interference caused by indoor coherent multipath signals, the noise subspace matrix is ​​adaptively corrected through network weights to suppress multipath spurious peaks derived from wall reflections and occlusions, while preserving the spectral peak characteristics of the true direct signal.

[0184] Step S304: Deep learning regression to accurately solve the AOA (Angle of Arrival) for each base station.

[0185] The DA-MUSIC network relies on an optimized signal subspace. Noise subspace At the same time, the array antenna spacing parameter is introduced. and radio frequency signal wavelength This type of array structure uses prior information; it completely abandons the traditional MUSIC algorithm's method of solving angles by searching the spatial spectrum peaks through a global angle traversal. Instead, it leverages the powerful nonlinear mapping and fitting capabilities of deep neural networks to complete regression calculations, enabling it to directly predict the two sets of high-precision signal incidence complementary angles corresponding to the m-th receiving base station. The angle of arrival symbols for each base station antenna array are defined as follows: This represents the angle of arrival of the signal calculated by each of the two antenna arrays at base station 1. The angle of arrival of the signal is calculated by each of the two antenna arrays at base station No. 2. The angle of arrival of the signal is calculated by each of the two antenna arrays at base station No. 3.

[0186] While outputting the angle estimates, the network simultaneously generates the confidence coefficient for each set of angle solutions, expressed as follows: The coefficient ranges from 0 to 1, and its value can quantitatively characterize the reliability and estimation accuracy of the AOA angle solution results of a single base station. The confidence level will serve as a key weight parameter to support the subsequent dynamic weighted adaptive self-calibration iteration process of base station coordinates, effectively weaken the error interference caused by low-precision angle observations, and continuously improve the overall spatial positioning stability.

[0187] The step S4 dynamic weighted base station self-calibration and coordinate calculation step provided in this embodiment of the invention includes the following sub-steps:

[0188] Step S401: Solving for the dynamic error exponent and real-time weighting coefficients:

[0189] The confidence coefficient is calculated based on the angles output by each base station in step S3 above. Construct two sets of dynamically adjustable weights that can be updated in real time. ;in The weighting coefficient representing the positional deviation Characterizing the angular error weighting coefficient; All can be based on the confidence coefficient of each base station Real-time assignment and adjustment are completed, and the value range of both is uniformly limited to the closed interval [0,1].

[0190] Set variables For the first Cumulative historical location deviation for each base station For the first The instantaneous error of the angle calculation at the current moment of each base station is obtained by substituting the above variables into the dynamic error exponent calculation formula. Dynamic error index of base station:

[0191]

[0192] Dynamic error index obtained for all base stations After performing normalization, the specific positioning weighting coefficients for each base station are derived, and the calculation formula is as follows:

[0193]

[0194] The normalized base station positioning weights satisfy the normalization constraint:

[0195]

[0196] The weighting follows these rules: the smaller the dynamic error index of a base station, the higher the confidence level of its angle calculation, and the larger the corresponding spatial positioning weighting coefficient. Conversely, base stations with large error indices and low angle calculation accuracy will be assigned smaller positioning weights to mitigate the negative impact of low-precision observation data on the overall positioning results.

[0197] Step S402: Solve for the initial coordinates of the tag using multi-base station AOA angle ray cross-positioning.

[0198] Retrieve the calibration spatial coordinates of three base stations , , Combined with the precise angle of arrival of each base station A system of two-dimensional angle incident ray equations is constructed for each base station; the predicted intersection points of the tags corresponding to base station 1, base station 2, and base station 3 are solved using a spatial geometric intersection algorithm. , , The spatial overlap and ray intersection count of the three sets of intersection points were statistically analyzed, and the intersection point with the highest overlap and smallest adaptation angle error was selected as the initial estimated coordinates of the calibration label. ;

[0199] Step S403: Calculation of positioning residuals and solution of base station coordinate deviation:

[0200] Retrieve the true coordinates of the calibration labels pre-calibrated by the total station. The initial estimated coordinates are compared with the true coordinates to solve for the 3D spatial positioning residual vector.

[0201] ,

[0202] Combining the dynamic weights of each base station The offset error of the current spatial coordinates of each base station is derived in reverse. The error allocation rule is as follows: the smaller the weight and the larger the error exponent of the base station, the larger the coordinate offset correction amount is allocated, so as to accurately match the real-time positioning error status of each base station.

[0203] Step S404 Iteratively update the base station coordinates to achieve online adaptive self-calibration:

[0204] Based on the base station coordinate offset error obtained from the solution Combined with preset iterative correction step size The coordinates of the three base stations are updated iteratively in real time, and the update formula is as follows: ,in For the first The updated calibration coordinates of the base station are continuously executed in a loop, encompassing data acquisition, phase preprocessing, angle calculation, and residual correction. The base station coordinate parameters are iteratively optimized until the spatial positioning residual is corrected. To preset the positioning accuracy threshold, base station convergence calibration is completed. This step does not require manual re-deployment or movement of calibration tags. It relies on time-series deep learning for continuous iterative correction, achieving 24 / 7 online adaptive calibration of the base station. Finally, the coordinates of the calibrated base station are used to achieve high-precision and stable positioning of the Bluetooth tag under test.

[0205] like Figure 2 As shown, an embodiment of the present invention provides a Bluetooth AOA base station self-calibration positioning system based on deep learning MUSIC optimization, comprising:

[0206] The data acquisition module is used for data acquisition.

[0207] The preprocessing module is used for deep learning preprocessing of multipath interference;

[0208] Angle of arrival calculation module, used for depth-enhanced DA-MUSIC angle of arrival calculation;

[0209] The coordinate calculation module is used for dynamic weighted base station self-calibration and coordinate calculation.

[0210] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, the computer program being executed by the processor causing the processor to perform the steps of the Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization.

[0211] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the Bluetooth AOA base station self-calibration positioning method optimized based on deep learning MUSIC.

[0212] Another objective of this invention is to provide an information data processing terminal for implementing the Bluetooth AOA base station self-calibration positioning system optimized based on deep learning MUSIC.

[0213] Specific implementation of the present invention:

[0214] This embodiment uses a standard smart campus classroom as the indoor test space. The overall test space parameters are fixed: length 9m, width 7.2m, and ceiling height 3.6m. The space includes static obstructions such as desks and chairs, as well as dynamic human walking obstruction scenarios, adapting to the Bluetooth indoor positioning requirements of a typical classroom teaching environment. The Bluetooth hardware used in this embodiment is a Bluetooth 5.1 compliant AOA array base station with a carrier operating wavelength of... Antenna array element spacing In this embodiment, all threshold parameters, distance parameters, and error grading parameters have been calibrated through multiple sets of comparative experiments. The specific values ​​are as follows:

[0215] First horizontal height = 0.8m, first standard height = 0.3m, second standard height = 0.1m; first directional distance = 0.4m, second directional distance = 0.8m; first adjustment height = 0.12m, second adjustment height = 0.25m; first lateral distance = 0.15m, second lateral distance = 0.3m; iterative convergence accuracy threshold. .

[0216] like Figure 3This invention discloses a Bluetooth AOA base station self-calibration and positioning method based on deep learning MUSIC optimization. The method consists of four main execution steps: S1 data acquisition, S2 multipath interference deep learning preprocessing, S3 deep-enhanced DA-MUSIC angle-of-arrival calculation, and S4 dynamic weighted base station self-calibration and coordinate calculation. The detailed implementation process step-by-step, based on a classroom testing scenario, is as follows:

[0217] Step S1 Data Acquisition Step

[0218] Step S101 Spatial Deployment Planning and Hardware Deployment: Inside a classroom test space with a ceiling height of 3.6m, a length of 9m, and a width of 7.2m, three Bluetooth 5.1 AOA array receiving base stations (Base Station 1, Base Station 2, and Base Station 3) are deployed in an equilateral triangular network configuration. The center-to-center distance between any two base stations is uniformly set to 6m. All three base stations are deployed at a height of 2.8m above the ground, which can completely cover the entire classroom space and realize multi-angle cross-angle measurement. A total of 8 passive Bluetooth calibration tags are deployed on the classroom floor, desk surfaces, and wall facades. The three-dimensional spatial coordinates of each calibration tag are calibrated one by one using a high-precision total station, and the calibrated coordinates are entered into the system as the true coordinates of the dataset. Each base station is equipped with a 3-element uniform linear receiving antenna array, and the Bluetooth CTE constant tone extended phase angle measurement function is enabled on all base station hardware.

[0219] Step S102 Signal Period Broadcast and Synchronization Timing Configuration: All passive calibration tags are uniformly set to a broadcast period of 20ms, continuously transmitting Bluetooth CTE modulated carrier signals; Base Station 1, Base Station 2, and Base Station 3 are connected to the same clock synchronization module to complete nanosecond-level timing alignment, ensuring that all array antennas of the three base stations synchronously sample and receive Bluetooth carrier signals, completely avoiding the phase timing deviation problem caused by asynchronous sampling.

[0220] Step S103 Parallel Acquisition of Raw Time-Series Data and Dataset Construction: Define Base Station Numbers Single base station antenna channel number ; Representing the base station The root antenna at time The raw IQ phase values ​​were acquired. To measure the RSSI signal strength of the corresponding channel, sampling data was continuously collected for three scenarios: open space without table or chair obstruction, static table and chair obstruction, and dynamic obstruction due to multiple people moving around. 12,000 time-series samples were collected for each scenario. The phase data and RSSI data at the same sampling time were bound one by one with the ground truth coordinates of the calibration label, and the training set:validation set:test set = 7:2:1 was divided to construct the original phase time-series input dataset.

[0221] During the sampling process, antennas 1 and 3 inside a single base station receive Bluetooth signals from the same source, resulting in a phase difference. Using the direction angle formula Calculate the complementary angle of incidence of the signal; solve for two sets of complementary angles of incidence for base stations 1, 2, and 3 respectively. This provides basic angle parameters for subsequent spatial rendezvous and positioning.

[0222] Step S104 3D Virtual Space Modeling: Retrieve the spatial parameters of the classroom (length, width, and height), and deploy the three-dimensional coordinates of the three base stations. A 1:1 scale 3D virtual model of the real classroom was built, and the spatial positions of all base stations, array antennas and calibration tags were marked in the virtual model. Based on the incident complementary angles of each group, angular radiation circles were drawn, and the intersection points of the intersecting lines of the circles and the incident rays were obtained to obtain the preliminary positions of the tags and complete the basic geometric positioning deduction.

[0223] like Figure 4 Step S2: Multipath interference deep learning preprocessing step (processed by dual-channel CNN-LSTM network)

[0224] Step S201: Construction of the four-dimensional input tensor of the network: The phase and RSSI time series data acquired in S1 are concatenated to construct a single-time spatial matrix. By stacking 20 consecutive temporal samples, a four-dimensional input tensor is generated. ,batch With a timing step size of 20, a total number of base stations of 3, and a single base station antenna channel of 3, this tensor is input into a dual-channel CNN-LSTM hybrid neural network.

[0225] Step S202: CNN spatial convolution channel extracts spatial features and distinguishes multipath clutter: The two-dimensional convolutional layer extracts features from the phase space matrix and outputs a spatial feature map. Set the direct signal threshold. Multipath clutter threshold When channel features The signal was determined to be a direct and valid signal. It was determined to be weak multipath interference; The interference was identified as wall or human body reflection noise; multipath interference confidence coefficients were output channel by channel. This completes the spatial dimension noise filtering.

[0226] In this embodiment, the multipath interference confidence coefficients output by the CNN spatial convolution channel are used to generate training labels based on the calibration label ground truth coordinates, array antenna coordinates, and the original IQ / RSSI sampling data. Specifically, for the first... base station The original phase of each antenna channel at time t and The true coordinates of the calibration label calibrated by the total station Calculate the geometric phase of the direct path using base station / array element coordinates. And calculate the phase residual. Then, by combining the RSSI drop, the phase consistency of adjacent array elements, and the ray intersection residual of the three base stations, multipath soft tags are generated. When the phase residual is below the first threshold and the RSSI does not drop abruptly, Setting it to 1 indicates that the direct projection is reliable; when the phase residual is in the middle threshold range, A value of 0.5 indicates weak multipath propagation; when the phase residual changes abruptly, the RSSI drops sharply, and the ray intersection residual increases simultaneously, A value of 0 indicates strong reflection or blocking of clutter.

[0227] The CNN training loss function is defined as follows: The BCE term constrains the network to distinguish between direct, weak, and strong multipath signals; the phase consistency term constrains the phase after confidence filtering to be consistent with the geometrically direct phase; and the TV smoothing term constrains the continuity of confidence between adjacent antenna channels and adjacent time frames, preventing single-frame noise from causing the erroneous deletion of effective channels. This demonstrates that the multipath confidence output by the CNN is supported by explicit training labels and a reproducible loss function.

[0228] The first term in the formula, the BCE binary cross-entropy term, is used to constrain the network based on multipath soft labels. Distinguishing between direct, weak multipath, and strong multipath signals; the second phase consistency constraint term, after confidence screening using weighted L2 constraints, corrects the phase. Direct phase with geometric theory To maintain consistency; the third term, TV total variation smoothing, is used to constrain the multipath confidence of adjacent antenna channels and adjacent time frames. To maintain continuous smoothness and avoid the accidental removal of effective line-of-sight channels due to instantaneous single-frame noise. The entire loss function is combined with geometrically derived multipath soft labels.

[0229] Joint training demonstrates that the multipath confidence of the CNN output in this invention has clear and reproducible ground truth labels and a standardized loss function theoretically supported by these.

[0230] Step S203: Fit phase drift to LSTM timing channels and solve for hardware bias coefficients: This involves fitting the continuous timing phase sequence... The system inputs data into an LSTM memory cell to capture the phase timing variations caused by ambient temperature and crystal oscillator drift, and iteratively solves for the dynamic hardware phase deviation correction coefficient for each antenna. .

[0231] The true values ​​of the LSTM dynamic hardware phase bias correction coefficients can be constructed from the geometric phase residuals. During the training phase, the true coordinates of the stationary calibration labels are used. Calculate the geometric distance of each element relative to the label. The theoretical geometric phase is obtained. After expanding the measured phase, the dynamic hardware phase deviation teacher value can be constructed as follows: ,in The LSTM output is used to estimate the multipath residuals based on the multipath confidence weighting of the CNN. and with Used as supervised labels for training.

[0232] During long-term online operation, when frame-by-frame manual calibration of hardware deviation truth values ​​is unavailable, a self-supervised approach is used to construct deviation coefficients: constrained by the calibration tag truth coordinates or the previous round of high-confidence positioning coordinates, the compensated phase, DA-MUSIC output angle, and multi-base station intersection coordinates simultaneously satisfy geometric consistency. The self-supervised loss can be written as... The first constraint is geometric consistency, the second constraint is the temporal continuity of temperature drift and crystal oscillator drift, and the third constraint is to avoid unbounded drift of the deviation coefficient.

[0233] Step S204 Phase Compensation and Pure Phase Difference Sequence Output: According to the formula Compensate for hardware system errors channel by channel; combined with confidence coefficients Eliminate phase abrupt changes caused by multipath propagation; perform differential operations on the corrected phase to solve for the pure phase difference between each pair of antennas at the base station. , The final output is a global pure phase difference matrix. It is then sent to the DA-MUSIC network.

[0234] Step S3: Depth Enhancement DA-MUSIC Angle of Arrival Calculation Steps:

[0235] Step S301 Secondary pre-compensation for array hardware errors: This involves pre-compensating the pure phase difference matrix. Enter the DA-MUSIC network and retrieve Construct a global hardware error compensation matrix The phase data is then recalibrated to eliminate fixed array errors caused by antenna installation tilt angle and factory inconsistencies, resulting in the final calibrated phase data set. .

[0236] Step S302 Adaptively construct the signal covariance matrix: Construct the signal vector of a single base station array The time-varying covariance matrix is ​​solved by adaptive iteration through a network. Unlike the traditional MUSIC fixed covariance matrix construction method, it is adapted to indoor dynamic and multipath environments.

[0237] Step S303 Feature space decomposition and multipath pseudo-peak suppression: for Perform eigenvalue decomposition and set an effective signal eigenvalue threshold. ; eigenvalue greater than The corresponding vector is divided into signal subspaces. The rest are designated as noise subspace. The network adaptively corrects the noise subspace weights to suppress spectral peak splitting and false spectral peaks caused by coherent multipath signals, retaining only the true spectral peak characteristics of direct Bluetooth signals.

[0238] Step S304 Deep learning regression outputs accurate angle of arrival and confidence score: Abandoning the traditional global angle traversal peak-finding mode, it relies on neural network nonlinear regression to directly output two sets of incident complementary angles for each of base station 1, base station 2, and base station 3. Synchronously output the confidence level of angle calculation for each base station The value ranges from 0 to 1, and the higher the confidence level, the higher the accuracy of the angle calculation.

[0239] The specific interface definition of the DA-MUSIC network is as follows: . in, This is the global pure phase difference matrix output by S2. The multipath confidence matrix output by the CNN. This is the hardware phase offset matrix output by the LSTM. This includes the initial values ​​of carrier wavelength λ, element spacing, element coordinates, and base station coordinates; M is the number of base stations, K is the number of elements per base station, and G is the number of angle grids.

[0240] DA-MUSIC internally includes a pseudo-spectral peak suppression head, PeakSuppressHead, whose input is the traditional MUSIC pseudo-spectrum. Local peak prominence, eigenvalue interval Multipath confidence and phase deviation Pseudo-peak mask on the output angle grid The suppressed spatial spectrum is The AngleHead, a regression head, outputs the angle of arrival based on the suppressed spatial spectrum, signal subspace, noise subspace, and array prior. The confidence header outputs confidence based on the maximum spectral peak probability, peak spacing, angular regression variance, and input multipath confidence. During training, angle supervision loss, spurious peak masking loss, and confidence calibration loss are used together to constrain the network, so that the spurious peak suppression results and angle confidence output form a clear network interface.

[0241] Step S4: Dynamically Weighted Base Station Self-Calibration and Coordinate Calculation

[0242] Step S401 Real-time calculation of dynamic error index and positioning weight: Dynamic error index calculation formula: ;in in accordance with Real-time dynamic assignment The historical location of the base station has accumulated deviations. The real-time angle error is used to obtain the positioning weights of each base station through normalization. The total weight is always equal to 1, and base stations with smaller errors and higher confidence are assigned greater positioning weights.

[0243] Step S402 Solving the initial coordinates of the tag by multi-base station angle intersection: retrieve the spatial coordinates of the three base stations, construct a system of spatial angle ray equations by combining the arrival angles of each AOA, and solve for the three sets of intersection points respectively. The number of intersecting rays and their spatial overlap are statistically analyzed, and the point with the most intersections is selected as the initial estimated coordinates for the label. .

[0244] Step S403: Determine the positioning residual classification and perform initial position correction: Retrieve the true coordinates of the calibration label. Calculate the height direction error height Planar distance error ;

[0245] 1. Error height ≥ 0.3m (first standard height): judged as first-level horizontal error;

[0246] 2. 0.1m ≤ error height < 0.3m: judged as level two horizontal error;

[0247] 3. Error height < 0.1m: judged as no horizontal error;

[0248] Planar orientation error classification:

[0249] 1. (Distance in the first direction): No directional error;

[0250] 2. (Second direction distance): Determined as Level II / No directional error based on the level of horizontal error;

[0251] 3. Determine the first / second level of directional error based on the level of horizontal error;

[0252] Based on the error level, the estimated coordinate points are corrected by vertical translation and horizontal offset. Level 1 horizontal error: vertical offset of 0.12m; Level 2 horizontal error: offset of 0.25m after remeasurement; Level 1 directional error: horizontal movement of 0.15m; Level 2 directional error: horizontal movement of 0.3m after remeasurement.

[0253] Step S404: Iterative update of base station coordinates and online adaptive convergence calibration: Calculate the 3D positioning residual vector. Set the iteration step size Base station coordinate update formula: ;Continuously iterate and correct the base station coordinates throughout the entire process from S1 to S4 until the spatial residual modulus is reached. The calibration is considered complete.

[0254] When allocating the tag positioning residual to the coordinate correction of each base station, the Jacobian matrix of the angle observation model is used for derivation. Let the... The coordinates of the base station are The true coordinates of the calibration tag are The azimuth observation output by DA-MUSIC is Base station The theoretical azimuth function to the label is Angular residual is Its Jacobian with respect to base station coordinates is , ,in If pitch angle is used, the pitch angle function and its counterpart can be constructed simultaneously. The partial derivatives are used to achieve three-dimensional joint correction.

[0255] The angle residuals of the three base stations are stacked as follows: Stack the Jacobians into H and set the DA-MUSIC angle confidence. Convert to weights The base station coordinate correction can be obtained by weighted least squares. ,in , This is the regularization coefficient. During online lightweight iteration, gradient updates can be applied to each base station. Therefore, base stations with low confidence and large error exponents will receive larger coordinate corrections, while base stations with high confidence will receive smaller corrections, thus aligning with the original dynamic weight allocation rules.

[0256] Supports the S4 dynamic weighting strategy: high-confidence angle observations receive higher positioning weights, while low-confidence observations are weakened.

[0257] like Figure 5This embodiment eliminates the need for manual repositioning of calibration tags. Relying on temporal deep learning, it continuously iterates and optimizes the base station deployment coordinates around the clock, enabling 24 / 7 online adaptive self-calibration. After calibration convergence, the average indoor static positioning error can be controlled within 6cm, and the positioning error in dynamic walking scenarios is less than 12cm. Compared with the traditional fixed MUSIC Bluetooth AOA positioning solution, the positioning accuracy is improved by more than 45%, effectively overcoming the positioning accuracy attenuation problems caused by indoor multipath interference, hardware phase drift, and long-term coordinate offset of base stations. It is highly suitable for use in indoor personnel positioning scenarios in smart campus classrooms.

[0258] The aforementioned static 6cm, dynamic 12cm, and improvement of over 45% technical effects are supported by a unified statistical standard: for each test sample Calculate the three-dimensional Euclidean positioning error ;

[0259] In the formula For the first Group sample localization coordinate estimation To calibrate the true coordinates of the sample for the total station.

[0260] The static scenarios include two types of conditions: open and unobstructed, and static occlusion by tables and chairs, totaling 2400 test samples. The dynamic scenarios involve continuous movement of people, resulting in human occlusion, totaling 1200 test samples. Four error evaluation indicators are uniformly output: mean error (MAE), root mean square error (RMSE), 50th percentile positioning error (P50), and 90th percentile positioning error (P90).

[0261] The benchmark algorithm used in comparison adopted the traditional fixed-parameter MUSIC Bluetooth AOA positioning scheme. The experimental hardware, base station deployment, calibration tag ground truth, test dataset, and spatial intersection solution module were completely consistent. The benchmark algorithm did not incorporate CNN multipath confidence filtering, LSTM dynamic phase drift compensation, DA-MUSIC pseudo-peak suppression and angle confidence weighting, nor did it perform online iterative self-calibration of base station coordinates based on the Jacobian matrix.

[0262] The accuracy improvement ratio is calculated based on the global average positioning error, using the following formula:

[0263] Improvement rate =

[0264] The average 3D positioning error of the benchmark algorithm The average three-dimensional positioning error of the complete solution of this invention.

[0265] Table 1. Comparison of positioning error and accuracy improvement between fixed MUSIC and the method of the present invention in static and dynamic scenarios.

[0266]

[0267] Figure 6 This is a flowchart of another Bluetooth AOA base station self-calibration positioning method. The method involves synchronously collecting CTE signals, IQ phase data, and RSSI data from each array base station. Spatial phase features are extracted using CNN to generate multipath confidence scores. LSTM is used to fit the phase drift coefficients during continuous operation. A clean phase difference sequence is formed through multipath filtering. Error compensation and angle calculation are performed using DA-MUSIC. Positioning weights and positioning residuals are then calculated based on the angle confidence scores, and the coordinates of each base station are iteratively updated.

[0268] Figure 7 This is a block diagram of another Bluetooth AOA base station self-calibration positioning system. The system includes a perception layer, a data layer, a phase preprocessing module, a positioning processing layer, and a closed-loop calibration layer. The array base station and calibration tag acquire CTE signals, and the synchronous data acquisition module outputs IQ phase and RSSI. The CNN spatial feature unit and LSTM temporal correction unit generate multipath confidence and dynamic bias coefficients, respectively. The DA-MUSIC angle calculation module outputs the AOA estimate, and online calibration is completed by the dynamic weight, positioning residual, and base station coordinate update modules.

[0269] Figure 8 This is a comparison chart of localization errors in static and dynamic scenes. The chart compares the MAE and RMSE indices of the fixed MUSIC control method and the present invention in static and dynamic test scenarios. The present invention reduces the static MAE from 10.9 cm to 5.8 cm and the static RMSE from 13.6 cm to 7.2 cm; the dynamic MAE decreases from 21.6 cm to 11.7 cm and the dynamic RMSE decreases from 27.1 cm to 14.9 cm, indicating that the present invention has a significant suppression effect on both static errors and dynamic disturbances.

[0270] Figure 9 The figure shows the cumulative distribution curve of positioning errors. It illustrates the cumulative probability distribution of 3D positioning errors under four conditions: static fixed MUSIC, static present invention, dynamic fixed MUSIC, and dynamic present invention. The curve corresponding to the present invention shifts towards a lower error region overall. Specifically, in the static scenario, the main positioning error can be controlled to around 6 cm, and in the dynamic scenario, the main positioning error can be controlled to around 12 cm. This demonstrates that the present invention can improve the concentration, stability, and upper limit control capability of positioning results.

[0271] Figure 10The figure shows the effect of phase drift suppression and self-calibration during continuous operation. The continuous operation time is plotted on the x-axis, and the average positioning error on the y-axis. The figure compares the cumulative phase drift results of the fixed MUSIC method with the positioning results after LSTM compensation and online self-calibration in this invention. As the operation time increases, the error of the fixed MUSIC method continuously increases, while the error of this invention remains within a low range, indicating that temporal drift modeling and online coordinate updates of the base station can effectively suppress accuracy degradation during long-term operation.

[0272] Figure 11 This figure shows the contribution of each technical module to the improvement of positioning accuracy. The graph compares the changes in the overall MAE according to the order of addition of CNN multipath confidence, LSTM phase compensation, DA-MUSIC pseudo-peak suppression, and base station online self-calibration. The overall error decreased sequentially from 14.5 cm with the fixed MUSIC to 11.8 cm, 9.8 cm, 8.6 cm, and 7.8 cm, with an overall reduction of 46.2%. This indicates that each technical module creates synergistic gains in multipath identification, drift compensation, angle calculation, and coordinate calibration.

[0273] Figure 12 This image shows the suppression effect of DA-MUSIC on multipath pseudo-spectral peaks. The figure compares the spatial spectrum outputs of traditional MUSIC and DA-MUSIC under the same received signal conditions. Traditional MUSIC is easily affected by reflections from indoor walls, floors, and obstacles, leading to the formation of multiple pseudo-spectral peaks with high amplitudes around the true direct peak. This invention adaptively corrects the covariance matrix and peak response by combining the multipath confidence of the CNN output, retaining the main peak corresponding to the direct path while weakening the side peaks formed by low-confidence reflection paths, thus improving the uniqueness and accuracy of AOA angle estimation.

[0274] Figure 13 This is a graph showing the relationship between angle confidence and positioning error. The graph illustrates the change between the angle confidence of the CNN output and the corresponding positioning residual, explaining the basis for dynamic weight calculation. As angle confidence increases, the degree to which the observed angle is affected by multipath reflection and noise interference decreases, resulting in an overall reduction in positioning error. During the S4 dynamic weighting process, high-confidence angle observations are assigned higher positioning weights, while low-confidence observations are assigned lower weights, thus preventing abnormal AOA results from dominating multi-base station fusion positioning.

[0275] In the description of this invention, unless otherwise stated, "a plurality of" means two or more; the terms "upper," "lower," "left," "right," "inner," "outer," "front end," "rear end," "head," "tail," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0276] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0277] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization, characterized in that, Includes the following steps: S1, Data Acquisition: Deploy at least three Bluetooth 5.1 AOA array base stations and several calibration tags with known coordinates indoors, continuously collect the raw phase timing data of the CTE carrier through the array antenna, and simultaneously collect the RSSI signal strength of each channel to construct the raw phase timing dataset. S2, Multipath Interference Deep Learning Preprocessing: The original phase time series dataset is processed using a dual-channel CNN-LSTM hybrid neural network. The spatial convolution channel is used to extract the phase spatial features of the array antenna and distinguish between direct Bluetooth signals and multipath clutter formed by wall reflections. The temporal LSTM channel is used to extract phase time drift features and output the dynamic correction coefficient of hardware phase deviation to obtain a clean phase difference sequence after filtering out multipath interference. S3, Deep Enhanced DA-MUSIC Angle of Arrival Calculation: The pure phase difference sequence is input into the DA-MUSIC deep learning network. The array hardware error is pre-compensated through the recurrent network, the signal covariance matrix is ​​adaptively calculated, multipath pseudo-spectral peaks are filtered out, and the AOA arrival angle corresponding to each base station is output. S4, Dynamically Weighted Base Station Self-Calibration and Coordinate Calculation: Based on the signal confidence level output by the deep learning network, the positioning error index of each base station is updated in real time, and dynamic positioning weights are assigned to each base station. Cross-circle positioning is performed by combining the angles of arrival of multiple base stations, and the intersection point with the most intersections is selected as the label coordinates. The coordinates of each base station are updated iteratively in real time to complete the online self-calibration of the base stations.

2. The Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization as described in claim 1, characterized in that, Step S1 includes the following sub-steps: S101, Spatial Layout Planning and Hardware Deployment: Within the indoor space to be located, at least three Bluetooth 5.1 AOA array receiving base stations are deployed in a triangular network configuration, with a distance of 3m to 12m between any two base stations to ensure full coverage of the space to be located and to achieve multi-angle cross-angle measurement. Several Bluetooth calibration tags are deployed on the indoor floor, walls, and desktop. The three-dimensional spatial coordinates of each calibration tag are pre-determined using a total station and entered into the system. The three-dimensional spatial coordinates are used as the ground truth labels of the dataset. Each AOA array receiving base station is equipped with a uniform linear array receiving antenna and the Bluetooth CTE phase angle measurement function is enabled. S102, Signal periodic broadcast and synchronization timing configuration: The calibration tag is configured to continuously broadcast a Bluetooth CTE modulated carrier signal at a fixed transmission period of 20ms; all array receiving base stations are connected to the same clock synchronization module to complete the timing alignment between multiple base stations, so that each array antenna synchronously samples the wireless signal transmitted by the calibration tag to eliminate the phase timing deviation caused by asynchronous sampling; S103, Parallel acquisition of multi-channel raw sensor data: The original phase timing data of the incident Bluetooth carrier is synchronously collected frame by frame by multiple receiving antennas in each array receiving base station, and the RSSI received signal strength of the corresponding link is collected through each antenna channel. The IQ phase sequence and RSSI value of each channel of the array corresponding to the same sampling time are packaged and cached to generate a single frame sampling sample. S104, Multi-scene traversal sampling and dataset annotation construction: Sampling was performed sequentially in open, unobstructed scenes, static obstacle occlusion scenes formed by tables, chairs, and walls, and dynamic occlusion scenes formed by frequent movement of people. Temporal samples were continuously collected in each scene, and each frame of sampled data was bound and labeled with the known real 3D coordinates of the calibration label corresponding to the current time. Annotated samples from various scenarios are collected and divided into training, validation, and test sets to construct an original phase-time series dataset with ground truth labels, which is used for training and inference of a dual-channel CNN-LSTM hybrid neural network.

3. The Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization as described in claim 2, characterized in that, The formula for calculating the basic phase difference angle of arrival involved in the spatial deployment and timing synchronization steps in step S1 includes the following sub-steps: In step S105, three Bluetooth 5.1 AOA array receiving base stations (base station 1, base station 2, and base station 3) are deployed indoors in a triangular configuration. All base stations are connected to the same clock synchronization module to achieve timing alignment. The deployed passive Bluetooth calibration tags continuously broadcast Bluetooth CTE modulated carrier signals at a fixed period of 20ms. Each of the three base stations is equipped with a uniform linear array receiving antenna array to synchronously collect the same source carrier signals. For the linear array antenna inside base station 1, two groups of antenna elements are selected for angle calculation: Let the first group of antennas inside base station 1 be antenna... ,antenna The center-to-center distance between the two antennas is The second group of antennas is an antenna. ,antenna The center-to-center distance between the two antennas is Bluetooth signals from the same source arriving at different antennas will experience phase shifts due to inconsistent transmission path lengths. The complementary angle of signal incidence can be calculated using the phase difference. The formula for calculating the direction angle is: ; in: : The wavelength of the CTE Bluetooth carrier signal emitted by the calibration tag; These correspond to the pairing of two sets of antennas within base station 1; :antenna With antenna The phase difference value of received signals from the same source; :antenna With antenna The phase difference value of received signals from the same source; :antenna , The center distance between them; :antenna , The center distance between them; :based on , The calculated signal incidence co-angle is 1. :based on , The calculated signal incidence co-angle is 2; Similarly, following the exact same formula and calculation logic, the complementary angle of incidence corresponding to base station 2 is calculated respectively. The incident complementary angle corresponding to base station 3 ; Step S106: Retrieve the coordinates of base station 1 pre-calibrated by the total station. Base station 2 coordinates 3-coordinate base station and the three-dimensional coordinates of the calibration tag true value A global three-dimensional virtual model was built by combining the length, width and height dimensions of the indoor space, and the positions of the three base stations and all the antenna units inside the model were marked. The incident complementary angle obtained using base station 1 Determine the spatial incident direction of the calibration tag relative to base station 1; in Within its plane, with antenna Draw a circular trajectory with the center of the circle and the incident direction as the radial axis, denoted as circle. ; The incident complementary angle obtained using base station 1 Determine the calibration tag relative to the antenna The spatial incident direction; in Within its plane, with antenna Draw a circular trajectory with the center of the circle and the incident direction as the radial axis, denoted as circle. Simultaneously repeat the above circle-drawing operation on base stations 2 and 3 to generate circles sequentially. and circle Gradually and synchronously increase the radius values ​​of all circular trajectories to obtain the circle. With circle The common straight line obtained by the intersection ; Antenna Incident ray and straight line towards the calibration label The intersections yield the initial intersection point from the perspective of base station 1. ; The intersection points obtained from solving the three base stations individually To conduct spatial intersection, select the spatial point where the three rays intersect with the highest degree of overlap. This point is the real-time calculated coordinate of the calibration tag; the calculated coordinates with true coordinates The comparison yields the basic angle measurement error, which is used for subsequent deep learning dataset annotation.

4. The Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization as described in claim 1, characterized in that, The S2 multipath interference deep learning preprocessing step includes the following sub-steps: Step S201: Construct the network input matrix and temporal sample tensor with dimension annotations. Read the raw IQ phase data and RSSI data of all array antenna channels of base station 1, base station 2, and base station 3 respectively. Define: base station number as... Each base station has an internal antenna channel numbered as follows: ;set up For the first Taiwan base station, the first The root antenna at time The original acquired phase value, This refers to the RSSI signal strength value synchronously acquired for the corresponding channel; The phase and RSSI data of all base stations and all antennas at a single moment are concatenated in spatial order to construct a single-moment spatial input matrix. Then stack them continuously. A time-series sampling frame is used to construct the input tensor of the four-dimensional network. ,in This refers to the number of sample batches. For timing step size, The total number of base stations. The number of antenna channels per base station is used as the sole input to a dual-channel CNN-LSTM hybrid neural network. Step S202 CNN Spatial Convolution Channels: Spatial Phase Feature Extraction and Multipath Detection The training labels for the multipath interference confidence coefficients are generated by the direct geometric phase corresponding to the calibration label ground truth coordinates, the measured phase residual, the RSSI drop, the phase consistency of adjacent array elements, and the multi-base station intersection residual. The training loss includes at least the multipath classification loss, the phase consistency loss, and the timing / channel smoothing loss. For the input matrix Two-dimensional convolution operations are performed to extract the spatial phase distribution features between base stations and between antenna channels, and the spatial feature map is output. Define the spatial characteristic threshold of the direct signal. Multipath clutter characteristic threshold ; For arbitrary channel phase characteristics :like If the channel is determined to be a direct and valid signal; It is determined to be a weak multipath interference signal; if The signal was determined to be invalid due to wall reflection or blockage of clutter. Simultaneously output the confidence coefficient of multipath interference for each channel. , , used to characterize the base station The purity of the channel signal is used to complete the screening of multipath noise and feature purification in the spatial dimension; Step S203 LSTM Timing Channel: Phase Timing Drift Fitting and Hardware Bias Coefficient Solving The dynamic hardware phase deviation correction coefficient is updated iteratively using the geometric phase residual derived from the total station's true coordinates as the supervisory true value, or using the geometric consistency between the compensated phase, angle of arrival, and multi-base station intersection coordinates as a self-supervised constraint. Continuous time phase sequence The input LSTM timing network is fitted with the timing phase drift caused by antenna hardware temperature drift, crystal oscillator error and environmental disturbance through a triple gating structure of input gate, forget gate and output gate. Definition of the first base station The dynamic hardware phase deviation correction factor for the root antenna is The network iteratively updates all data based on the temporal feature loss. It outputs the independent phase deviation compensation value for each antenna channel in real time, realizing the dynamic quantification of hardware system errors; Step S204: Multipath filtering + phase compensation to generate a globally clean phase difference sequence. Using timing deviation coefficient The original phase value is corrected channel by channel to obtain the compensated accurate phase: ; Combining multipath confidence coefficients Weighted filtering is applied to anomalous phase points to eliminate phase abrupt noise caused by wall reflections and multipath occlusion; the corrected pure phase is then filtered. Perform channel differential calculations to solve for the pairwise phase differences between the internal antennas of base station 1, base station 2, and base station 3. , ; The final output completely filters out multipath clutter, eliminates timing drift, and compensates for hardware errors. The global clean phase difference sequence matrix serves as the precise input data for subsequent DA-MUSIC network angle of arrival calculation.

5. The Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization as described in claim 1, characterized in that, The S3 depth-enhanced DA-MUSIC angle of arrival calculation step includes the following sub-steps: Step S301 DA-MUSIC network input loading and array hardware error pre-compensation: The global pure phase difference sequence matrix output in step S2 As dedicated input data for the DA-MUSIC deep learning network, the base station number is defined. The phase difference between paired antenna channels in a single base station is ; Hardware phase deviation correction coefficients for each antenna channel obtained based on S2 Construct a global hardware error compensation matrix for the base station array. The pure phase sequence is pre-corrected a second time using an error compensation matrix to eliminate fixed system errors caused by factory inconsistency deviations and installation tilt deviations of the array antenna, thus obtaining the final corrected phase dataset. ; Step S302: Construction of the adaptive signal covariance matrix: Based on the correction phase dataset Construct the received signal vector of a single base station array ,in For the first Taiwan base station The array signal column vector at time step; abandoning the traditional MUSIC algorithm's fixed covariance matrix construction method, the real-time signal covariance matrix is ​​obtained through adaptive iterative computation using the DA-MUSIC network: In the formula, For the first The array signal covariance matrix of the base station, For mathematical expectation calculation, For matrix conjugate transpose operation, it accurately adapts to signal characteristic changes in indoor dynamic multipath interference scenarios; Step S303: Feature Space Decomposition and Multipath Pseudo-Peak Suppression The covariance matrix obtained by adaptive solution Perform eigenvalue decomposition to obtain the eigenvalue set. Based on the corresponding eigenvector matrix, the signal subspace and noise subspace are distinguished according to the magnitude of the eigenvalues; an effective signal eigenvalue threshold is defined. , will be greater than The eigenvalues ​​corresponding to the vectors are classified into the signal subspace. less than or equal to The eigenvalues ​​corresponding to the vectors are classified into the noise subspace. To address the issues of spectral peak splitting and spurious peak interference caused by indoor coherent multipath signals, the noise subspace matrix is ​​adaptively corrected through network weights to suppress multipath spurious peaks derived from wall reflections and occlusions, while preserving the spectral peak characteristics of the true direct signal. Step S304: Deep learning regression to accurately solve the AOA (Angle of Arrival) for each base station. The DA-MUSIC network relies on an optimized signal subspace. Noise subspace At the same time, the array antenna spacing parameter is introduced. and radio frequency signal wavelength This type of array structure uses prior information; it completely abandons the traditional MUSIC algorithm's method of solving angles by searching the spatial spectrum peaks through a global angle traversal. Instead, it leverages the powerful nonlinear mapping and fitting capabilities of deep neural networks to complete regression calculations, enabling it to directly predict the two sets of high-precision signal incidence complementary angles corresponding to the m-th receiving base station. The angle of arrival symbols for each base station antenna array are defined as follows: This represents the angle of arrival of the signal calculated by each of the two antenna arrays at base station 1. The angle of arrival of the signal is calculated by each of the two antenna arrays at base station No.

2. The angle of arrival of the signal is calculated by each of the two antenna arrays at base station No.

3. While outputting the angle estimates, the network simultaneously generates the confidence coefficient for each set of angle solutions, expressed as follows: The value of this coefficient ranges from 0 to 1, and its magnitude can quantitatively characterize the reliability and estimation accuracy of the AOA angle solution results for a single base station. The confidence score will serve as a key weighting parameter to support the subsequent dynamic weighted adaptive self-calibration iteration process of base station coordinates, effectively weakening the error interference caused by low-precision angle observations and continuously improving the overall spatial positioning stability. The DA-MUSIC network includes a pseudo-peak suppression head, an angle regression head, and a confidence output head. The pseudo-peak suppression head generates a pseudo-peak mask based on the spatial spectrum peak value, eigenvalue interval, multipath confidence, and phase deviation. The angle regression head outputs the angle of arrival, and the confidence output head outputs the angle confidence that is negatively correlated with the angle error.

6. The Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization as described in claim 1, characterized in that, Step S4, dynamic weighted base station self-calibration and coordinate calculation, includes the following sub-steps: Step S401: Solving for the dynamic error exponent and real-time weighting coefficients: The confidence coefficient is calculated based on the angles output by each base station in step S3 above. Construct two sets of dynamically adjustable weights that can be updated in real time. ;in The weighting coefficient representing the positional deviation Characterizing the angular error weighting coefficient; All can be based on the confidence coefficient of each base station Real-time assignment and adjustment are completed, and the value range of both is uniformly limited to the closed interval [0,1]. Set variables For the first Cumulative historical location deviation for each base station For the first The instantaneous error of the angle calculation at the current moment of each base station is obtained by substituting the above variables into the dynamic error exponent calculation formula. Dynamic error index of base station: ； Dynamic error index obtained for all base stations After performing normalization, the specific positioning weighting coefficients for each base station are derived, and the calculation formula is as follows: ； The normalized base station positioning weights satisfy the normalization constraint: ； The weight allocation follows the following rules: the smaller the dynamic error index of the base station, the higher the confidence level of the angle calculation of the base station, and the larger the corresponding spatial positioning weighting coefficient. Conversely, base stations with larger error exponents and lower angle calculation accuracy will be assigned smaller positioning weights to mitigate the negative impact of low-precision observation data on the overall positioning results. Step S402: Solve for the initial coordinates of the tag using multi-base station AOA angle ray cross-positioning. Retrieve the calibration spatial coordinates of three base stations , , Combined with the precise angle of arrival of each base station A system of two-dimensional angle incident ray equations is constructed for each base station; the predicted intersection points of the tags corresponding to base station 1, base station 2, and base station 3 are solved using a spatial geometric intersection algorithm. , , ; The spatial overlap and ray intersection counts of the three sets of intersection points were statistically analyzed, and the intersection point with the highest overlap and smallest adaptation angle error was selected as the initial estimated coordinates of the calibration label. ; Step S403: Calculation of positioning residuals and solution of base station coordinate deviation: Retrieve the true coordinates of the calibration labels pre-calibrated by the total station. The initial estimated coordinates are compared with the true coordinates to solve for the 3D spatial positioning residual vector. ， Combining the dynamic weights of each base station The offset error of the current spatial coordinates of each base station is derived in reverse. The error allocation rule is as follows: the smaller the weight and the larger the error exponent of the base station, the larger the coordinate offset correction amount is allocated, so as to accurately match the real-time positioning error status of each base station. Step S404 Iteratively update the base station coordinates to achieve online adaptive self-calibration: Based on the base station coordinate offset error obtained from the solution Combined with preset iterative correction step size The coordinates of the three base stations are updated iteratively in real time, and the update formula is as follows: ,in For the first The updated calibration coordinates of the base station are continuously executed in a loop, encompassing data acquisition, phase preprocessing, angle calculation, and residual correction. The base station coordinate parameters are iteratively optimized until the spatial positioning residual is corrected. To preset the positioning accuracy threshold, base station convergence calibration is completed. This step does not require manual re-deployment or movement of calibration tags. It relies on time-series deep learning for continuous iterative correction to achieve 24 / 7 online adaptive calibration of the base station. Finally, the coordinates of the calibrated base station are used to achieve high-precision and stable positioning of the Bluetooth tag under test. The base station coordinate offset error is determined by the tag positioning residual, the Jacobian matrix of the angle observation function on the base station coordinates, and the angle confidence weight. The residual is then distributed as the base station coordinate correction amount through weighted least squares or gradient descent.

7. A Bluetooth AOA base station self-calibration positioning system based on deep learning MUSIC optimization, implementing the Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization as described in any one of claims 1-6, characterized in that, The Bluetooth AOA base station self-calibration positioning system based on deep learning MUSIC optimization includes: The data acquisition module is used for data acquisition. The preprocessing module is used for deep learning preprocessing of multipath interference; Angle of arrival calculation module, used for depth-enhanced DA-MUSIC angle of arrival calculation; The coordinate calculation module is used for dynamic weighted base station self-calibration and coordinate calculation.

8. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the Bluetooth AOA base station self-calibration positioning method based on deep learning MUSIC optimization as described in any one of claims 1-6.

10. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the Bluetooth AOA base station self-calibration positioning system based on deep learning MUSIC optimization as described in claim 7.

Images