Bluetooth AOA signal outlier elimination method and system based on deep learning

Abstract

The invention relates to a Bluetooth AOA signal outlier elimination method and system based on deep learning, and the method comprises the steps: initializing system configuration, and cooperatively collecting Bluetooth signals through employing a plurality of Bluetooth base stations; acquiring IQ data in a plurality of Bluetooth base stations, and performing data preprocessing; according to the IQ data, performing calculation to obtain a preliminary positioning result of the Bluetooth terminal; constructing an input data set of the deep learning model and designing a network structure of the deep learning model; training a deep learning model according to the input data set; the training precision of the deep learning model is judged, if the training precision meets the requirement, IQ data collected in real time is input into the trained deep learning model, and a real-time positioning result of the Bluetooth terminal is obtained; otherwise, returning to adjust the network structure; and optimizing the real-time positioning result by using a Kalman filtering algorithm to obtain a final positioning result. Compared with the prior art, the method has the advantages of high positioning accuracy and strong robustness.

Description

Technical Field

[0001] This invention relates to the field of Bluetooth indoor positioning, and in particular to a method and system for removing outliers from Bluetooth AOA signals based on deep learning. Background Technology

[0002] With the rapid development of IoT technology, location-based services are playing an increasingly important role in smart logistics, smart homes, and autonomous driving. Bluetooth technology, due to its low power consumption, ease of deployment, and simple integration, has been widely used in indoor positioning. Among these, Bluetooth Angle of Arrival (AOA) technology, a commonly used positioning method, is widely adopted due to its high accuracy and strong anti-interference capabilities. However, existing Bluetooth AOA technology still faces many challenges in practical applications, such as the effects of multipath propagation, noise, and phase shift, leading to insufficient positioning accuracy and a lack of robustness. Furthermore, traditional positioning technologies struggle to meet the high-precision, real-time positioning requirements in complex and variable indoor environments, such as hospitals, shopping malls, and large office buildings.

[0003] To address these issues, researchers have begun exploring new technological solutions. Among them, deep learning-based methods have emerged as a promising indoor positioning technology due to their ability to learn features directly from raw data, reducing human intervention. However, existing deep learning methods still have limitations when dealing with multi-channel speech separation problems, such as difficulty in fully utilizing spatial information and a lack of adaptability to environmental changes. Therefore, how to apply deep learning technology to Bluetooth AOA signal processing to improve positioning accuracy and robustness has become a current research hotspot and challenge.

[0004] Against this backdrop, a Bluetooth AOA signal outlier removal method based on embedded deep learning has emerged. This method aims to remove outlier interference by performing deep processing on the Bluetooth AOA signal, thereby improving positioning accuracy and providing a new solution for indoor positioning technology.

[0005] Several invention patents have been issued to address the issue of Bluetooth indoor positioning accuracy. For example: CN115840190A discloses a high-precision positioning method based on the fusion of Bluetooth AOA and deep learning. This method acquires the raw sampled I and Q phase data of the Bluetooth terminal by using the AOA master node Bluetooth base station and multiple passive node Bluetooth base stations, and sends the obtained IQ phase data to a PC processing terminal. The PC processing terminal inputs the obtained IQ phase data into a trained neural network model to obtain the real-time positioning result of the Bluetooth terminal. Based on the real-time positioning result, it determines whether the Bluetooth terminal's motion state is dynamic or static; if dynamic, the positioning result is optimized based on an extended Kalman filter algorithm; if static, the positioning result is optimized based on a multi-point averaging filter algorithm. However, in practical applications, this method may face problems such as signal reflection interference, antenna switching time delay, and multipath effects caused by various obstructions and reflective objects in indoor environments.

[0006] CN112040394A proposes a Bluetooth positioning method based on AI deep learning. The method works as follows: an AOA positioning base station acquires the phase data of the first signal transmitted by an AOA signal source. The AOA positioning base station then sends this acquired phase data to an AI server. The AI ​​server, based on a trained neural network model and the acquired phase data, determines the location of the AOA signal source transmitting the first signal. This method introduces AI technology into the AOA positioning field. By sampling the raw phase values ​​received from the positioning signal source by multiple antennas, it further abstracts the phase difference, converts it into angles, and trains the relevant phase data onto a known grid-based spatial coordinate system through combinations of different antenna angles. However, in practical applications, this method still faces the challenge of further optimizing the neural network model to improve positioning accuracy and speed.

[0007] The existing technology has the following drawbacks: 1. Traditional Bluetooth AOA positioning technology is easily affected by signal reflection interference, antenna switching time delay, and multipath effects such as various obstructions and reflective objects in indoor environments in practical applications, resulting in low positioning accuracy and lack of robustness.

[0008] 2. Existing deep learning-based indoor positioning methods have limitations when dealing with multi-channel signal separation problems, such as difficulty in fully utilizing spatial information and lack of adaptability to environmental changes.

[0009] 3. Current deep learning models struggle to meet the high-precision, real-time positioning requirements when dealing with complex and ever-changing indoor environments, and are unable to effectively handle multi-source heterogeneous data.

[0010] 4. Existing Bluetooth positioning methods face the challenge of further optimizing neural network models to improve positioning accuracy and speed in practical applications, lacking more advanced deep learning technologies and richer feature information.

[0011] 5. Traditional positioning technologies lack effective integration and utilization mechanisms when processing multi-source heterogeneous data, making it difficult to fully explore the potential information in the data, thus affecting the accuracy and efficiency of positioning. Summary of the Invention

[0012] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method and system for removing outliers from Bluetooth AOA signals based on deep learning.

[0013] The objective of this invention can be achieved through the following technical solutions: This invention provides a deep learning-based method for removing outliers in Bluetooth AOA signals, comprising: Step 1: Initialize system configuration, then use multiple Bluetooth base stations to collaboratively collect Bluetooth AOA signals; acquire IQ data from the multiple Bluetooth base stations and perform data preprocessing; calculate the preliminary positioning result of the Bluetooth terminal based on the IQ data; construct the input dataset of the deep learning model and design the network structure of the deep learning model based on the preprocessed data and the preliminary positioning result; train the deep learning model based on the input dataset. Step 2: Determine the training accuracy of the deep learning model. If the training accuracy meets the requirements, input the real-time collected IQ data into the trained deep learning model to obtain the real-time positioning result of the Bluetooth terminal; otherwise, return to adjust the network structure. Step 3: Optimize the real-time positioning results using the Kalman filter algorithm to obtain the final positioning result.

[0014] Furthermore, the initialization of system configuration specifically includes: Configure the basic parameters of the multiple Bluetooth base stations, including center frequency, bandwidth, and modulation scheme; initialize the state vector of the Kalman filter; set the training parameters of the deep learning model, including learning rate, batch size, and maximum number of iterations.

[0015] Furthermore, the Bluetooth AOA signal acquisition process specifically includes: The multiple Bluetooth base stations collaboratively collect the AOA signals sent by the Bluetooth terminal, and a master node Bluetooth base station and a passive node Bluetooth base station are set up; the master node Bluetooth base station and the passive node Bluetooth base station adopt different antenna array structures.

[0016] Furthermore, step one also includes the processing of IQ data, specifically including: The IQ data acquired from multiple Bluetooth base stations is synchronized using either a master-slave base station time synchronization method or a GPS synchronization method.

[0017] Furthermore, data preprocessing specifically includes: The IQ data is denoised using a threshold removal method based on median and standard deviation or a threshold removal method based on PCA to remove outliers. Calculate the phase difference of the IQ data, and use cross-correlation or subspace methods to eliminate multipath effects in the IQ data.

[0018] Furthermore, the design of the network structure for deep learning models specifically includes: Configure the input layer, hidden layer, and output layer of the network structure; design the number of nodes in the input layer, hidden layer, and output layer of the deep learning model's network structure, respectively.

[0019] Furthermore, the process of training a deep learning model also includes: When constructing the input dataset, the input dataset is divided into training samples and test samples; and the number of training samples and test samples is determined respectively; the parameters of the deep learning model are optimized using the Adam optimization algorithm or the SGD optimization algorithm. After the real-time collected IQ data is input into the trained deep learning model, the deep learning model extracts features from the IQ data, generates feature vectors, and sends them to a multilayer perceptron. The multilayer perceptron analyzes the feature vectors, outputs a score representing whether the IQ data is an outlier, and sends it to the Softmax function for processing. The Softmax function ultimately gives the probability distribution of whether the IQ data is an outlier.

[0020] Furthermore, determining the training accuracy of the deep learning model specifically includes: Set a training accuracy threshold; if the training accuracy of the deep learning model is less than or equal to the training accuracy threshold, then return the parameters for adjusting the network structure; if the accuracy is greater than the training accuracy threshold, then input the real-time collected IQ data into the trained deep learning model.

[0021] Furthermore, step two also includes: After the collected IQ data is input into the trained deep learning model, the deep learning model outputs the real-time positioning coordinates of the Bluetooth terminal, including x, y and z axis coordinates. When optimizing the real-time positioning results using the Kalman filter algorithm, a specific filtering time constant is set.

[0022] This invention provides a Bluetooth AOA signal outlier removal system based on deep learning, including a memory and a processor. The memory stores a computer program, and the processor calls the computer program to execute the steps of any of the methods described above.

[0023] Compared with the prior art, the present invention has the following advantages: (1) This invention introduces embedded deep learning technology, constructs a deep learning model, and uses multiple Bluetooth base stations to collaboratively collect Bluetooth AOA signals; acquires IQ data from multiple Bluetooth base stations, and trains the deep learning model using a precisely set training set; processes the real-time Bluetooth AOA signal using the trained deep learning model to obtain the real-time positioning angle, and optimizes the real-time positioning angle using the Kalman filter algorithm. It achieves deep processing and analysis of Bluetooth AOA signals, effectively eliminating outlier interference, improving positioning accuracy, overcoming the shortcomings of traditional Bluetooth AOA technology which is susceptible to multipath effects in complex indoor environments, and significantly improving the robustness and reliability of indoor positioning; and based on the feature learning capability of deep learning, it automatically extracts the effective features of Bluetooth AOA signals, avoiding manual intervention and improving the automation and efficiency of positioning.

[0024] (2) This invention uses a trained deep learning model to collaboratively process IQ data from multiple Bluetooth base stations; it uses a deep learning model to directly process the multi-channel signal separation problem, making full use of spatial information, improving adaptability to environmental changes, and overcoming the limitations of existing technologies in making full use of spatial information; through deep integration and processing of multi-source heterogeneous data, it fully explores the potential information of the data, improves the accuracy and efficiency of positioning, and overcomes the shortcomings of insufficient processing of multi-source heterogeneous data in traditional technologies.

[0025] (3) This invention achieves high-precision, real-time positioning of complex and ever-changing indoor environments by optimizing the deep learning model and combining it with the efficient computing power of embedded hardware, thus meeting the positioning requirements of high precision and real-time performance. Attached Figure Description

[0026] Figure 1 This is a flowchart of a Bluetooth AOA signal outlier removal method based on deep learning provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the outlier removal process of a deep learning-based Bluetooth AOA signal outlier removal method provided in this embodiment of the invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0028] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0029] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0030] Example 1 This invention provides a deep learning-based method for removing outliers from Bluetooth AOA signals, the method comprising the following steps: S1: Initialize system configuration, then use multiple Bluetooth base stations to collaboratively collect Bluetooth AOA signals; acquire IQ data from multiple Bluetooth base stations and perform data preprocessing; calculate the preliminary positioning result of the Bluetooth terminal based on the IQ data; construct the input dataset of the deep learning model and design the network structure of the deep learning model based on the preprocessed data and the preliminary positioning result; train the deep learning model based on the input dataset. S101: Configure the basic parameters of the Bluetooth base station, including center frequency, bandwidth, modulation method, etc. Initialize the state vector of the Kalman filter; Configure the training parameters of the deep learning model, including learning rate, batch size, maximum number of iterations, etc.

[0031] S102: The AOA signal sent by the Bluetooth terminal is collected through the coordinated efforts of multiple base stations; Acquire IQ data collected by each base station; Perform time synchronization processing on IQ data.

[0032] S103: Denoising the IQ data and removing outliers; Calculate phase differences and eliminate multipath effects; Angle conversion is performed to obtain preliminary positioning results for the Bluetooth terminal.

[0033] S2: Determine the training accuracy of the deep learning model. If the training accuracy meets the requirements, input the real-time collected IQ data into the trained deep learning model to obtain the real-time positioning result of the Bluetooth terminal; otherwise, return to adjust the network structure. S201: Construct the input dataset for the deep learning model; Design the network structure of a deep learning model, including the input layer, hidden layer, and output layer; Train the model and optimize its parameters; Evaluate the model's performance. If the performance does not meet the requirements, return to step 402 to continue training. Specifically, After real-time collected IQ data is input into a trained deep learning model, the deep learning model extracts features from the IQ data, generates feature vectors, and sends them to a multilayer perceptron. The multilayer perceptron analyzes the feature vectors, outputs a score indicating whether the IQ data is an outlier, and sends it to the Softmax function for processing. The Softmax function ultimately provides the probability distribution of whether the IQ data is an outlier.

[0034] S3: Optimize the real-time positioning results using the Kalman filter algorithm to obtain the final positioning result.

[0035] S301: The collected IQ data is input into the trained deep learning model; The model outputs the real-time location coordinates of the Bluetooth terminal; Kalman filtering is applied to the real-time positioning results; Output the final location result.

[0036] Example 2 This invention provides a Bluetooth AOA signal outlier removal system based on deep learning, including a memory and a processor. The memory stores a computer program, and the processor calls the computer program to execute the steps of any of the methods in Embodiment 1. The specific implementation steps are as follows: Step 1: Initialize system configuration.

[0037] Step 101: Configure the basic parameters of the Bluetooth base station, including a center frequency of 2.4GHz, a bandwidth of 2MHz, and a modulation method of GFSK; initialize the state vector of the Kalman filter as [x=0, y=0, vx=0, vy=0]; set the training parameters of the deep learning model, including a learning rate of 0.01, a batch size of 32, and a maximum number of iterations of 1000.

[0038] Step 102: Collect AOA signals sent by Bluetooth terminals through multiple base stations in a coordinated manner. The master node Bluetooth base station adopts a four-antenna ring array structure, and the passive node adopts two antennas. Obtain the IQ data collected by each base station, using an 8-bit complex sample format and a sampling rate of 1kHz.

[0039] Step 103: Perform time synchronization processing on the IQ data, using a master-slave base station time synchronization method, with synchronization accuracy controlled within ±0.1μs.

[0040] Step 2: Perform signal preprocessing.

[0041] Step 301: Denoise the IQ data by using a threshold removal method based on median and standard deviation to remove outliers; calculate the phase difference and use a cross-correlation method to eliminate multipath effects; perform angle conversion to obtain the preliminary positioning results of the Bluetooth terminal, with positioning accuracy controlled within ±1m.

[0042] Step 302: Construct the input dataset for the deep learning model, which includes 50,000 training samples and 10,000 test samples.

[0043] Step 303: Design the network structure of the deep learning model, including the input layer (64 nodes), the hidden layer (128 nodes), and the output layer (3 nodes).

[0044] Step 3: Train the deep learning model.

[0045] Step 401: Train the model and optimize the model parameters using the Adam optimization algorithm; evaluate the model's performance. If the model accuracy is not higher than 95%, return to adjust the network structure parameters. If the accuracy reaches 95% or higher, proceed to step 4.

[0046] Step 4: Achieve real-time positioning.

[0047] Step 501: Input the collected IQ data into the trained deep learning model; the model outputs the real-time positioning coordinates of the Bluetooth terminal, including x, y and z axis coordinates.

[0048] Step 502: Perform Kalman filtering on the real-time positioning results, with a filtering time constant of 0.1s.

[0049] Step 503: Output the final positioning result, with positioning accuracy controlled within ±0.3m.

[0050] Example 3 This invention provides a Bluetooth AOA signal outlier removal system based on deep learning, including a memory and a processor. The memory stores a computer program, and the processor calls the computer program to execute the steps of any of the methods in Embodiment 1. The specific implementation steps are as follows: Step 1: Initialize system configuration.

[0051] Step 101: Configure the basic parameters of the Bluetooth base station, including a center frequency of 2.45GHz, a bandwidth of 1.5MHz, and a modulation method of π / 4-DQPSK; initialize the state vector of the Kalman filter as [x=0, y=0, vx=0, vy=0]; set the training parameters of the deep learning model, including a learning rate of 0.005, a batch size of 64, and a maximum number of iterations of 2000.

[0052] Step 102: Collect AOA signals sent by Bluetooth terminals through multiple base stations in a coordinated manner. The master node Bluetooth base station adopts a six-antenna ring array structure, and the passive node adopts a four-antenna structure. Obtain the IQ data collected by each base station, using a 16-bit complex sample format and a sampling rate of 2kHz.

[0053] Step 103: Perform time synchronization processing on the IQ data, using GPS synchronization, with synchronization accuracy controlled within ±0.05μs.

[0054] Step 2: Perform signal preprocessing.

[0055] Step 301: Denoise the IQ data by using a PCA-based threshold removal method to remove outliers; calculate the phase difference and use a subspace method to eliminate multipath effects; perform angle conversion to obtain the preliminary positioning results of the Bluetooth terminal, with positioning accuracy controlled within ±0.5m.

[0056] Step 302: Construct the input dataset for the deep learning model, which includes 80,000 training samples and 20,000 test samples.

[0057] Step 303: Design the network structure of the deep learning model, including the input layer (128 nodes), the hidden layer (256 nodes), and the output layer (3 nodes).

[0058] Step 3: Train the deep learning model.

[0059] Step 401: Train the model and optimize the model parameters using the SGD optimization algorithm; evaluate the model performance. If the model accuracy is not higher than 98%, return to adjust the network structure parameters. If the accuracy reaches 98% or higher, proceed to step 4.

[0060] Step 4: Achieve real-time positioning.

[0061] Step 501: Input the collected IQ data into the trained deep learning model; the model outputs the real-time positioning coordinates of the Bluetooth terminal, including x, y and z axis coordinates.

[0062] Step 502: Perform Kalman filtering on the real-time positioning results, with a filtering time constant of 0.05s.

[0063] Step 503: Output the final positioning result, with positioning accuracy controlled within ±0.1m.

[0064] Glossary: Bluetooth AOA positioning technology: a high-precision indoor positioning solution based on the Bluetooth 5.1 protocol specification.

[0065] The core principle is as follows: when a Bluetooth device emits a special directional signal, a positioning base station equipped with an antenna array receives the signal and accurately measures the phase difference of the signal reaching different antenna elements, thereby calculating the angle of arrival of the signal. By measuring the angle of the same target from multiple base stations, the target's position coordinates can be calculated using triangulation. This technology improves the accuracy of Bluetooth positioning from the traditional "meter level" to "sub-meter level," achieving centimeter-level to decimeter-level positioning in complex indoor environments.

[0066] Multipath effect: a common phenomenon in wireless communication, refers to the phenomenon that when a radio signal travels from the transmitter to the receiver, it passes through multiple paths such as direct transmission, reflection, and diffraction, resulting in multiple signal copies with different delays, phases, and amplitudes, which are then superimposed at the receiver.

[0067] This superposition leads to fluctuating signal amplitude and phase distortion, severely interfering with communication quality and positioning accuracy. In positioning systems, multipath effects can confuse signals along direct paths, causing ranging or direction-finding errors, which is one of the main technical challenges of high-precision positioning.

[0068] Kalman filter algorithm: It is an efficient optimal recursive digital signal processing algorithm used to estimate the internal state of a dynamic system from a series of noisy observation data.

[0069] Its core idea is to combine the system's predictive model with actual observation data, and then iteratively perform a "prediction-update" process. The algorithm first predicts the current state based on the system's state and motion model from the previous moment, and then uses the current observations to weight and correct the prediction, thus obtaining the optimal estimate of the system state at the current moment. It is particularly suitable for real-time, continuous state estimation of dynamic systems that change over time, and is widely used in navigation, tracking, and control.

[0070] Outliers: also known as anomalies or outliers, are individual observations in a dataset that deviate significantly from their true values ​​or the overall trend of the majority of the data.

[0071] Outliers are typically caused by transient, non-systematic factors such as measurement errors, signal interference, sensor malfunctions, or sudden environmental changes. The presence of outliers can significantly distort the accuracy of data analysis and model building. Therefore, in fields such as signal processing, statistics, and machine learning, they must be identified and processed using specialized filtering or elimination algorithms to ensure the reliability and robustness of the results.

[0072] Threshold-based removal methods using median and standard deviation are robust outlier detection techniques.

[0073] This method leverages the median's insensitivity to outliers to represent the data center and estimates data dispersion based on the absolute median difference. A threshold range is constructed by calculating "median ± k × standard deviation," and all data points falling outside this range are considered outliers and removed. This method effectively resists the interference of outliers themselves on threshold calculation, is more stable than traditional methods based on the mean and standard deviation, and is particularly suitable for data cleaning scenarios with outliers.

[0074] Cross-correlation method: a digital signal processing technique used to measure the similarity of two signals at different time offsets.

[0075] The core of this method is to obtain a sequence of cross-correlation functions by calculating the inner product of one signal and another signal at different time delays. The time delay corresponding to the peak of this function represents the time difference between the two signals. In positioning applications, this method can identify the main peak corresponding to the direct path from multipath aliasing signals, thereby accurately estimating the time difference of arrival and providing an effective means to combat multipath interference.

[0076] Adam optimization algorithm: It is a deep learning adaptive learning rate optimization algorithm that combines the momentum method and RMSProp.

[0077] Its core principle involves calculating the first and second moments of the gradient, which are the gradient mean and uncentered variance, respectively, and then correcting for biases to dynamically adjust the learning rate for each parameter. This method combines the advantages of maintaining momentum along historical gradient directions with the adaptive scaling of the learning rate by RMSProp, enabling efficient and stable updates of network parameters. It is widely used in the training process of deep neural networks.

[0078] IQ data: where I represents the in-phase component and Q represents the quadrature component, together forming a complex number I+jQ, which fully describes the instantaneous state of the signal. By down-converting and ADC sampling, the high-frequency carrier is converted into a baseband IQ sequence, providing the original signal basis for phase difference measurement, spectrum analysis, etc., and forming the foundation for high-precision positioning and communication demodulation.

[0079] GPS synchronization method: refers to the technology of using the high-precision timing signal provided by the Global Positioning System to achieve a unified time base among multiple distributed devices.

[0080] Its core principle is to receive navigation messages containing precise UTC time information broadcast by GPS satellites and extract synchronization pulse signals through calculation. This method can provide microsecond-level or even nanosecond-level time synchronization for systems such as communication base stations and sensor networks, effectively solving the coordination deviation problem caused by clock drift, and is a key foundation for ensuring the consistency of global system behavior.

[0081] The PCA-based threshold removal method is a multidimensional data anomaly detection technique based on principal component analysis.

[0082] This method projects high-dimensional data onto a feature subspace composed of principal components and calculates the reconstruction error of each sample in the residual subspace. When the reconstruction error of a sample exceeds a statistical threshold determined by the principal component eigenvalues, it is identified as an outlier. This method can effectively capture hidden anomalies that disrupt the overall correlation of the data and is particularly suitable for quality control and fault diagnosis of multi-dimensional sensor data.

[0083] Subspace methods are a class of high-resolution signal processing techniques based on matrix eigenvalue decomposition.

[0084] The core idea is to decompose the covariance matrix of the received signal into mutually orthogonal signal and noise subspaces. Utilizing the orthogonality between the signal steering vector and the noise subspace, accurate estimation of signal parameters is achieved through spectral peak search. This type of method overcomes the Rayleigh limit of traditional beamforming, enabling super-resolution spectral estimation and finding wide application in fields such as direction-of-arrival estimation, frequency analysis, and modern spectrum sensing.

[0085] The Softmax function is a mathematical function that maps a set of arbitrary real-valued scores (logits) to a probability distribution vector. Its core principle is exponential normalization: first, each input value is exponentially calculated to amplify the differences between scores and ensure a positive result; then, each exponential value is divided by the sum of all exponential values, resulting in a sum of 1 for all output values. This allows the function's output to be interpreted as a probability distribution over mutually exclusive classes. It is widely used in the last layer of multi-class neural networks to transform the network's raw scores into clear, interpretable class prediction probabilities.

[0086] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for removing outliers from Bluetooth AOA signals based on deep learning, characterized in that, include Step 1: Initialize system configuration, then use multiple Bluetooth base stations to collaboratively collect Bluetooth AOA signals; acquire IQ data from the multiple Bluetooth base stations and perform data preprocessing; calculate the preliminary positioning result of the Bluetooth terminal based on the IQ data; construct the input dataset of the deep learning model and design the network structure of the deep learning model based on the preprocessed data and the preliminary positioning result; train the deep learning model based on the input dataset. Step 2: Determine the training accuracy of the deep learning model. If the training accuracy meets the requirements, input the real-time collected IQ data into the trained deep learning model to obtain the real-time positioning result of the Bluetooth terminal; otherwise, return to adjust the network structure. Step 3: Optimize the real-time positioning results using the Kalman filter algorithm to obtain the final positioning result.

2. The Bluetooth AOA signal outlier removal method based on deep learning according to claim 1, characterized in that, The initialization of system configuration specifically includes: Configure the basic parameters of the multiple Bluetooth base stations, including center frequency, bandwidth, and modulation scheme; initialize the state vector of the Kalman filter; set the training parameters of the deep learning model, including learning rate, batch size, and maximum number of iterations.

3. The Bluetooth AOA signal outlier removal method based on deep learning according to claim 1, characterized in that, The process of acquiring the Bluetooth AOA signal specifically includes: The multiple Bluetooth base stations collaboratively collect the AOA signals sent by the Bluetooth terminal, and a master node Bluetooth base station and a passive node Bluetooth base station are set up; the master node Bluetooth base station and the passive node Bluetooth base station adopt different antenna array structures.

4. The Bluetooth AOA signal outlier removal method based on deep learning according to claim 1, characterized in that, Step one also includes a process for processing IQ data, specifically including: The IQ data acquired from multiple Bluetooth base stations is synchronized using either a master-slave base station time synchronization method or a GPS synchronization method.

5. The Bluetooth AOA signal outlier removal method based on deep learning according to claim 1, characterized in that, The data preprocessing specifically includes: The IQ data is denoised using a threshold removal method based on median and standard deviation or a threshold removal method based on PCA to remove outliers. Calculate the phase difference of the IQ data, and use cross-correlation or subspace methods to eliminate multipath effects in the IQ data.

6. The Bluetooth AOA signal outlier removal method based on deep learning according to claim 1, characterized in that, The network structure of the deep learning model specifically includes: Configure the input layer, hidden layer, and output layer of the network structure; design the number of nodes in the input layer, hidden layer, and output layer of the deep learning model's network structure, respectively.

7. The Bluetooth AOA signal outlier removal method based on deep learning according to claim 1, characterized in that, The process of training the deep learning model also includes: When constructing the input dataset, the input dataset is divided into training samples and test samples; and the number of training samples and test samples is determined respectively; the parameters of the deep learning model are optimized using the Adam optimization algorithm or the SGD optimization algorithm. After the real-time collected IQ data is input into the trained deep learning model, the deep learning model extracts features from the IQ data, generates feature vectors, and sends them to a multilayer perceptron. The multilayer perceptron analyzes the feature vectors, outputs a score representing whether the IQ data is an outlier, and sends it to the Softmax function for processing. The Softmax function ultimately gives the probability distribution of whether the IQ data is an outlier.

8. The Bluetooth AOA signal outlier removal method based on deep learning according to claim 1, characterized in that, The determination of the training accuracy of the deep learning model specifically includes: Set a training accuracy threshold; if the training accuracy of the deep learning model is less than or equal to the training accuracy threshold, then return the parameters for adjusting the network structure; if the accuracy is greater than the training accuracy threshold, then input the real-time collected IQ data into the trained deep learning model.

9. The Bluetooth AOA signal outlier removal method based on deep learning according to claim 1, characterized in that, Step two also includes: After the collected IQ data is input into the trained deep learning model, the deep learning model outputs the real-time positioning coordinates of the Bluetooth terminal, including x, y and z axis coordinates. When optimizing the real-time positioning results using the Kalman filter algorithm, a specific filtering time constant is set.

10. A Bluetooth AOA signal outlier removal system based on deep learning, characterized in that, It includes a memory and a processor, the memory storing a computer program, the processor invoking the computer program to perform the steps of the method as described in any one of claims 1 to 9.

Images