Bluetooth AOA and RSSI fusion indoor positioning method based on deep learning

Abstract

The invention belongs to the technical field of Bluetooth indoor positioning, and relates to a Bluetooth AOA and RSSI fusion indoor positioning method based on deep learning, which comprises the following steps: step S10, a Bluetooth positioning base station receives a CTE signal sent by a signal source by using a BGAP protocol, and extracts I / Q data and RSSI data in the CTE signal; step S20, the Bluetooth positioning base station calculates the azimuth angle, the pitch angle and the distance of the signal by extracting the data, and uploads the azimuth angle, the pitch angle and the distance to the proxy server through an MQTT protocol; step S30, the PC subscribes to the proxy server, downloads azimuth angle, pitch angle and distance data, performs Kalman filtering on the angle data, and performs moving average filtering on the distance information; and S40, multiplying the filtered data by own transpose, inputting the data into a deep learning model for training, and finally outputting a label position coordinate. The distance information of the RSSI and the azimuth angle and pitch angle information of the AOA are fused, Kalman filtering and moving average filtering are used for processing data, and the specific position of the label is predicted through a deep learning network.

Description

Technical Field

[0001] The present invention belongs to the technical field of Bluetooth indoor positioning, and particularly relates to a Bluetooth AOA and RSSI fusion indoor positioning method based on deep learning. Background Art

[0002] With the rapid development of science and technology and the continuous improvement of people's living standards, the application demand for indoor positioning technology in various fields is increasing day by day. Traditional outdoor positioning technologies, such as the Global Navigation Satellite System (GNSS), can achieve precise positioning of outdoor targets. However, in indoor environments, due to factors such as signal occlusion and multipath effects, its positioning accuracy drops significantly. Therefore, indoor positioning technology has emerged as the times require. Among the many application fields of indoor positioning technology, the most prominent ones include retail and mall navigation, warehouse management, smart home, medical monitoring, and intelligent transportation, etc. Existing indoor positioning solutions usually rely on different wireless communication protocols, positioning algorithms, or working principles. This diversity results in a lack of compatibility and generality between systems. The development of indoor positioning systems usually focuses on commercial applications, and there are relatively few solutions for individuals or small users. This is because enterprises have an urgent need for positioning technology, while the demand and scale of the personal market are relatively limited. In addition, applications for individuals also face many challenges in terms of technical complexity and cost. Therefore, how to design an indoor positioning system with low cost, high accuracy, and wide applicability has become the key direction of current indoor positioning research.

[0003] In early 2019, the Bluetooth Special Interest Group released the Bluetooth 5.1 specification, which first introduced the Angle of Arrival (AOA) ranging technology that supports the "finding direction" function. Compared with traditional Bluetooth indoor positioning methods, AOA technology has the advantages of high positioning accuracy, strong anti-multipath interference, and low deployment cost, and is particularly suitable for complex indoor environments. With a significant improvement in accuracy and flexibility, this technology has injected new impetus into the development of indoor positioning and shown broad application prospects in the field of precise positioning. In AOA positioning, traditional subspace algorithms have high computational complexity, are sensitive to the Signal-to-Noise Ratio (SNR) and array scale, and their performance drops significantly in low SNR or multi-signal correlation scenarios, which is not conducive to real-time applications and the popularization in complex environments. In contrast, deep learning can automatically extract complex features in indoor positioning, improve positioning accuracy, and has strong robustness to multipath interference and noise. It can focus on global features and key information, fuse multiple data sources, support end-to-end optimization, and is suitable for complex environments and large-scale applications. Summary of the Invention

[0004] To solve the above problems, the present invention provides an indoor positioning method based on the fusion of Bluetooth AOA and Received Signal Strength Indicator (RSSI) using deep learning. The distance information of RSSI is fused with the azimuth and elevation angle information of AOA. The AOA data is optimized by Kalman filtering to reduce angular noise, and the sliding average filtering is used to smooth the distance data of RSSI. A CNN-MHA deep learning model is proposed, which uses a Convolutional Neural Network (CNN) to extract high-order features of angles and distances, and dynamically adjusts the feature weights through a Multi-Head Attention (MHA) mechanism, thereby achieving more accurate positioning of tags.

[0005] To achieve the above object of the present invention, the following technical solutions are adopted:

[0006] An indoor positioning method based on the fusion of Bluetooth AOA and RSSI using deep learning, comprising the following steps:

[0007] Step S10: The Bluetooth positioning base station uses the BGAP protocol to receive the CTE signal sent by the signal source, and extracts the I / Q data and RSSI data therein;

[0008] Step S20: The Bluetooth positioning base station calculates the azimuth, elevation angle and distance of the signal by extracting the data, and uploads them to the proxy server through the MQTT protocol;

[0009] Step S30: The PC subscribes to the proxy server, downloads the azimuth, elevation angle and distance data, uses Kalman filtering for the angle data, and performs sliding average filtering on the distance information;

[0010] Step S40: Multiply the filtered data by its own transpose and input it into the deep learning model for training, and finally output the tag position coordinates.

[0011] As a preferred technical solution of the present invention, the step S20 includes the following steps:

[0012] Step S21: Use the RTL library provided by Silicon Labs to calculate the azimuth and elevation angle of the signal, and calculate the RSSI value into the distance using the relationship between RSSI and the signal propagation distance; the relationship between RSSI and the signal propagation distance is:

[0013]

[0014] where A represents the signal strength value at 1 meter from the receiving end to the transmitting end, and n represents the exponent of the propagation loss;

[0015] Step S22: The three Bluetooth positioning base stations and the PC are connected to a local area network through a router and communicate with each other via MQTT. The base stations transmit the base station ID, timestamp, sampling sequence, calculated azimuth angle, elevation angle, and distance data to the proxy server through the MQTT protocol.

[0016] As a preferred technical solution of the present invention, the step S30 includes the following steps:

[0017] Step S31: The PC subscribes to the topic of the MQTT proxy server, downloads the data uploaded by the base stations, and arranges the data according to the base station ID and timestamp.

[0018] Step S32: Perform Kalman filtering on the azimuth angle and elevation angle. Taking one of the base stations as an example, the process of optimizing the azimuth angle and elevation angle by Kalman filtering is as follows:

[0019] The state vector is defined as:

[0020]

[0021] The state transition matrix F and the control matrix G are expressed as:

[0022]

[0023] Step S33: Calculate their accelerations through the azimuth angle and elevation angle of consecutive timestamps by the second-order central difference method. The formula is:

[0024]

[0025] In the formula: θ represents the azimuth angle or elevation angle, Δt represents the sampling interval between adjacent angle measurement values, t k-1 ,t k ,t k+1 represent three consecutive acquisition time points;

[0026] Use the measured acceleration information as the input u of the system k , and it is expressed as:

[0027]

[0028] Step S34: Obtain the state transition equation as:

[0029]

[0030] Step S35: The covariance prediction is expressed as:

[0031] P k|k-1 = F·P k-1 ·F T + Q (8)

[0032] Where: P k|k-1 is the predicted error covariance at the current time k, and P k-1|k-1 is the error covariance at the previous time k - 1; Q is the process noise covariance matrix. Since the signal source may be moving, Q is dynamically adjusted, and its expression is:

[0033]

[0034] Where: G is the control matrix, is a diagonal matrix composed of the variances of the real-time accelerations according to the azimuth angle and the elevation angle, and is expressed as:

[0035]

[0036] The Kalman gain is expressed as:

[0037] K k = P k|k-1 H T (HP k|k-1 H T + R) -1 (11)

[0038] Where: K k is the Kalman gain at the current time k, H is the observation matrix, and R is the observation noise covariance matrix;

[0039] To set an appropriate R, before positioning, the signal source is fixed at a certain position in the experimental site, the means of the azimuth angle and the elevation angle are measured multiple times, and the variances are calculated. The diagonal matrix R is set with the two variances, and R can be expressed as:

[0040]

[0041] Step S36. The updated state equation is expressed as:

[0042]

[0043] Where: is the predicted state at the current time, is the updated state estimate, and Z k is the observation value at the current time k;

[0044] The updated error covariance matrix is expressed as:

[0045] P k|k = (I - K k H)P k|k-1 (14)

[0046] Where: I is the identity matrix;

[0047] The above steps constitute the update process of the Kalman filter, and these equations ensure that when each new measurement arrives, the estimation of the angle state can be improved based on the state at the previous moment and the new observation value;

[0048] Step S37: Process the distance using a moving average filter. Define the size of the moving window as N, and set an array for storing distance data. When the number of data in the group is equal to the window size, calculate the average value of the distance data in the group, and this value is the distance value between the filtered and smoothed tag and the base station. The mathematical formula is expressed as:

[0049]

[0050] where D t is the t-th distance value after filtering, d t-i is the (t - i)-th distance value in the original signal, and N is the size of the filtering window.

[0051] As a preferred technical solution of the present invention, step S40 includes the following steps:

[0052] Step S41: Multiply the nine data of the three azimuth angles, pitch angles, and distances of the tag by its own transpose to generate a 9×9 outer product matrix, and input it into the deep learning model for training;

[0053] Step S42: Use a Convolutional Neural Network (CNN) to extract features from the AOA and RSSI data; adopt 3 convolutional layers, and connect a batch normalization layer and a Leaky ReLU activation function after each convolutional layer to improve the training speed and the expression ability of the model; finally, connect a max pooling layer to retain important feature information;

[0054] Step S43: Adjust the output data of the CNN and input it into the Multi-Head Attention (MHA). MHA can adaptively allocate attention weights to different parts of the input data. When a certain feature input is distorted, its weight is reduced to reduce the impact on the positioning result; input the output of MHA into a fully connected layer to predict the final position of the tag; use the Dropout regularization technique in the MHA layer to improve the generalization ability of the model;

[0055] Step S44: In the output layer, the features output by the fully connected layer are further processed, and linear regression is used to predict the position information of the signal source; the mean square error is used as the loss function to calculate the error between the actual position and the predicted position of the tag. The mean square error formula is:

[0056]

[0057] Among them, N is the number of samples, x i is the i-th actual x coordinate, y i is the i-th actual y coordinate, the i-th predicted x coordinate, is the i-th predicted y coordinate.

[0058] The RMSprop optimizer is used, which can dynamically adjust the learning rate to effectively handle the sparse gradient problem, accelerate the convergence of the model, and improve the training stability.

[0059] A Bluetooth AOA and RSSI fusion indoor positioning method based on deep learning according to the present invention, compared with the prior art using the above technical solutions, has the following technical effects:

[0060] (1) The method of the present invention fuses the distance information of RSSI with the azimuth and elevation angle information of AOA, processes the data using Kalman filtering and moving average filtering, and predicts the specific position of the tag through a deep learning network.

[0061] (2) The present invention improves the accuracy and robustness of indoor positioning in complex environments and enhances the adaptability of the positioning system to environmental changes. Brief Description of the Drawings

[0062] Figure 1 is the overall flowchart of the Bluetooth AOA and RSSI fusion indoor positioning method based on deep learning of the present invention;

[0063] Figure 2 is the structural diagram of the Bluetooth positioning base station of the present invention;

[0064] Figure 3 is the network framework diagram of the present invention;

[0065] Figure 4 is the deep learning model framework diagram based on CNN and MHA of the present invention. Detailed Embodiments

[0066] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present application, some symbols in the embodiments of the present application are first explained to facilitate understanding by those skilled in the art.

[0067] Embodiment 1

[0068] See Figure 1 , this embodiment provides a Bluetooth AOA and RSSI fusion indoor positioning method based on deep learning, including the following steps:

[0069] S10. The Bluetooth positioning base station uses the BGAP protocol to receive the CTE signal sent by the signal source, and extracts the I / Q data and RSSI data therein.

[0070] S20. The Bluetooth positioning base station calculates the azimuth angle, elevation angle and distance of the signal by extracting data, and uploads them to the proxy server through the MQTT protocol.

[0071] S30. The PC subscribes to the proxy server, downloads the azimuth angle, elevation angle and distance data, uses Kalman filtering for the angle data, and performs moving average filtering on the distance information.

[0072] S40. Multiply the filtered data by its own transpose and input it into the deep learning model for training, and finally output the label position coordinates.

[0073] Step S10 includes the following steps:

[0074] S11. The Bluetooth base station used in the present invention is equipped with a 4x4 uniform rectangular antenna array to receive the CTE signal emitted by the tag, as Figure 2 shown. When the base station receives the CTE signal broadcast by the tag, the I / Q data and RSSI values are collected in the specified order of the signal, and the CTE duration is set to 160 μs.

[0075] Step S20 includes the following steps:

[0076] S21. For the I / Q data, use the RTL library of Silicon Labs to calculate the azimuth angle and elevation angle. For the RSSI value, use the relationship between RSSI and propagation distance to calculate the distance.

[0077] S22. The present invention uses a router to connect three base stations and the PC in a local area network by pinging the IPs of the base stations and the PC, and uses the MQTT protocol to complete the communication between the base station and the PC, as Figure 3 shown. After the base station calculates the data, it uploads the azimuth angle, elevation angle, base station ID, sampling timestamp, and sampling sequence number to the proxy server in JSON format through the MQTT protocol.

[0078] Step S30 includes the following steps:

[0079] S31. The PC subscribes to the topic of the MQTT proxy server, downloads the data uploaded by the base station, and arranges the data according to the base station ID and timestamp for subsequent classification and filtering.

[0080] S32. Use Kalman filtering to process the azimuth angle and elevation angle. Taking one of the base stations as an example, the process of optimizing the azimuth angle and elevation angle by Kalman filtering is specifically as follows:

[0081] The state vector is defined as:

[0082]

[0083] The state transition matrix F and the control matrix G are expressed as:

[0084]

[0085] S33. Through the azimuth angle and elevation angle of consecutive timestamps, their accelerations can be calculated by the second-order central difference method, and the formula is:

[0086]

[0087] In the formula: θ represents the azimuth angle or elevation angle, Δt represents the sampling interval between adjacent angle measurement values, t k-1 , t k , t k+1 represent three consecutive acquisition time points.

[0088] Use the measured acceleration information as the input u of the system k , which is expressed as:

[0089]

[0090] S34. The state transition equation can be obtained as:

[0091]

[0092] S35. The covariance prediction is expressed as:

[0093] P k|k-1 = F·P k-1 ·F T + Q (7)

[0094] In the formula: P k|k-1 is the predicted error covariance at the current moment k, and P k-1|k-1 is the error covariance at the previous moment k - 1. Q is the process noise covariance matrix. In the system of this article, since the signal source may be moving, Q is dynamically adjusted, and its expression is:

[0095]

[0096] In the formula: G is the control matrix mentioned above, is a diagonal matrix composed of the variances of the real-time accelerations according to the azimuth angle and elevation angle, and is expressed as:

[0097]

[0098] The Kalman gain is expressed as:

[0099] K k = P k|k-1 H T (HP k|k-1 H T + R) -1 (10)

[0100] Where: K k is the Kalman gain at the current time k, H is the observation matrix, and R is the observation noise covariance matrix.

[0101] To set an appropriate R, before positioning, we fix the signal source at a certain position in the experimental site, measure the mean values of the azimuth angle and elevation angle multiple times, and calculate the variances. We set the diagonal matrix R with the two variances. R can be expressed as:

[0102]

[0103] S36. The updated state equation is expressed as:

[0104]

[0105] Where: is the predicted state at the current time, is the updated state estimate, Z k is the observation value at the current time k.

[0106] The updated error covariance matrix is expressed as:

[0107] P k|k = (I - K k H)P k|k-1 (13)

[0108] Where: I is the identity matrix.

[0109] The above steps constitute the update process of the Kalman filter. These equations ensure that when a new measurement arrives, the estimate of the angle state can be improved based on the state at the previous time and the new observation value.

[0110] S37. The distance between the base station and the tag is processed using a moving average filter. Define the size of the moving window as N, set an array for storing distance data. When the number of data in the group is equal to the window size, calculate the average value of the distance data in the group. This value is the distance value between the tag and the base station after filtering and smoothing. The mathematical formula can be expressed as:

[0111]

[0112] Where, D t is the t-th distance value after filtering, d t-iis the (t-i)-th distance value in the original signal, and N is the size of the filtering window.

[0113] Step S40 includes the following steps:

[0114] S41. First, sort the filtered data according to the base station ID and timestamp. Nine data, including the azimuth and elevation angles of the three base stations and the distance of the tag relative to each base station, are used as a dataset. Multiply each dataset by its own transpose to generate a 9×9 outer product matrix and input it into the deep learning model for training.

[0115] S42. The deep learning model architecture adopted in the embodiments of the present invention is composed of a combination of CNN and MHA. The model architecture is as Figure 4 shown. Convert the data into a two-dimensional outer product matrix so that CNN can extract features from the AOA and RSSI data just like it does for two-dimensional images. Three convolutional layers are used, with a convolutional kernel size of 3x3 and a stride of 1. After each convolutional layer, a batch normalization layer and a Leaky ReLU activation function are connected to improve the training speed and enhance the expressive ability of the model. Finally, a 2x2 max pooling layer is connected to retain important feature information.

[0116] S43. Adjust the output of CNN to conform to the input format of MHA. In the multi-head self-attention layer, the input feature dimension is 32, there are 4 attention heads in total, and the query, key, and value dimensions of each head are 8. Each head calculates the attention score through independent queries, keys, and values, and uses Softmax normalization to obtain the attention weights. These weights are used to weighted calculate the sum of the values, and the output of each head represents a different attention perspective. Finally, the outputs of all heads are concatenated through a linear transformation to generate a more comprehensive feature representation with a dimension of 64, and the dropout rate is set to 0.1 to prevent overfitting.

[0117] S44. In the output layer, the features output by the fully connected layer are further processed, and linear regression is used to predict the position information of the signal source. The loss function uses the mean square error to calculate the error between the actual position of the label and the predicted position. The mean square error formula is:

[0118]

[0119] where N is the number of samples, x i is the i-th actual x coordinate, y i is the i-th actual y coordinate, the i-th predicted x coordinate, is the i-th predicted y coordinate.

[0120] Use the RMSprop optimizer, which can dynamically adjust the learning rate, thus effectively dealing with the sparse gradient problem, accelerating the convergence of the model and improving the training stability.

[0121] After the PC uses the accurate position of the label output by deep learning, it can also upload the position information to the MQTT proxy server as a new topic for other users to subscribe to.

[0122] In the embodiment of the present invention, the angle information of AOA and the distance information of RSSI are fused, and the data is preprocessed by using Kalman filtering and sliding smoothing filtering. The fast mapping relationship between the angle and distance information of the label and the spatial coordinates is realized through a deep learning model based on CNN and MHA. Compared with the traditional indoor positioning method, the present invention has significant improvements in terms of accuracy and robustness, especially in static and dynamic positioning in complex environments, and thus designs an indoor positioning system with high accuracy, good stability, strong versatility and low cost.

[0123] The specific implementation schemes described above further elaborate on the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific implementation schemes of the present invention and are not intended to limit the scope of the present invention. Any equivalent changes and modifications made by those skilled in the art without departing from the concept and principles of the present invention shall fall within the scope of protection of the present invention.

Claims

1. A Bluetooth AOA and RSSI fusion indoor positioning method based on deep learning, characterized in that, It includes the following steps: Step S10: The Bluetooth positioning base station uses the BGAP protocol to receive the CTE signal sent by the signal source, and extracts the I / Q data and RSSI data therein; Step S20: The Bluetooth positioning base station calculates the azimuth angle, elevation angle and distance of the signal through the extracted data, and uploads them to the proxy server through the MQTT protocol; Step S30: The PC subscribes to the proxy server, downloads the azimuth angle, elevation angle and distance data, uses Kalman filtering for the angle data, and performs moving average filtering on the distance information; Step S40: Multiply the filtered data by its own transpose and input it into the deep learning model for training, and finally output the label position coordinates.

2. The indoor positioning method based on the fusion of Bluetooth AOA and RSSI using deep learning according to claim 1, characterized in that, The said Step S20 includes the following steps: Step S21: Use the RTL library provided by Silicon Labs to calculate the azimuth angle and elevation angle of the signal, and use the RSSI and signal propagation distance relationship formula to calculate the RSSI value into distance; The RSSI and signal propagation distance relationship formula is: Where, A represents the signal strength value at a distance of 1 meter from the transmitter at the receiver, and n represents the exponent of propagation loss; Step S22: Three Bluetooth positioning base stations and the PC are connected to a local area network through a router and communicate with each other through MQTT; The base station transmits the base station ID, timestamp, sampling sequence, calculated azimuth angle, elevation angle and distance data to the proxy server through the MQTT protocol.

3. The indoor positioning method based on the fusion of Bluetooth AOA and RSSI by deep learning according to claim 2, characterized in that, The said Step S30 includes the following steps: Step S31: The PC subscribes to the topic of the MQTT proxy server, downloads the data uploaded by the base station, and arranges the data according to the base station ID and timestamp; Step S32: Use Kalman filtering to process the azimuth angle and elevation angle. Taking one of the base stations as an example, the process of optimizing the azimuth angle and elevation angle by Kalman filtering is specifically as follows: The state vector is defined as: The state transition matrix F and the control matrix G are expressed as: Step S33: Through the azimuth angle and elevation angle of consecutive timestamps, calculate their acceleration through the second-order central difference method, and the formula is: Where: θ represents the azimuth angle or the elevation angle, Δt represents the sampling interval between adjacent angle measurement values, t k-1 ,t k ,t k+1 represent three consecutive acquisition time points; Use the measured acceleration information as the input u of the system k , expressed as: Step S34: Obtain the state transition equation as: The covariance prediction is expressed as: P k|k-1 = F·P k-1 ·F T + Q(8) where: P k|k-1 is the predicted error covariance at the current time k, and P k-1|k-1 is the error covariance at the previous time k - 1; Q is the process noise covariance matrix. Since the signal source may be moving, Q is dynamically adjusted, and its expression is: where: G is the control matrix, is a diagonal matrix composed of the variances of the real-time accelerations according to the azimuth and elevation angles, expressed as: The Kalman gain is expressed as: K k = P k|k-1 H T (HP k|k-1 H T + R) -1 (11) Where: K k is the Kalman gain at the current time k, H is the observation matrix, and R is the observation noise covariance matrix; In order to set a suitable R, before positioning, the signal source is fixed at a certain position in the experimental site, the mean values of the azimuth angle and elevation angle are measured multiple times, and the variance is calculated. The diagonal matrix R is set with the two variances, and R can be expressed as: Step S36: The updated state equation is expressed as: In the formula: is the predicted state at the current moment, is the updated state estimate, Z k is the observation value at the current moment k; The updated error covariance matrix is expressed as: P k|k =(I - K k H)P k|k-1 (14) In the formula: I is the identity matrix; The above steps constitute the update process of Kalman filtering. These equations ensure that when each new measurement arrives, the estimation of the angle state can be improved according to the state at the previous moment and the new observation value; Step S37: Perform moving average filtering on the distance. Define the size of the moving window as N, set an array for storing distance data. When the number of data in the group is equal to the window size, calculate the average value of the distance data in the group, and this value is the distance value between the filtered and smoothed label and the base station. The mathematical formula is expressed as: where D t is the t-th distance value after filtering, d t-i is the (t - i)-th distance value in the original signal, and N is the size of the filtering window.

4. The indoor positioning method based on the fusion of Bluetooth AOA and RSSI using deep learning according to claim 3, wherein The step S40 includes the following steps: Step S41: Multiply the nine data of the three azimuth angles, pitch angles, and distances of the tag by its own transpose to generate a 9×9 outer product matrix, and input it into the deep learning model for training; Step S42: Use a convolutional neural network CNN to extract features from the AOA and RSSI data; adopt three convolutional layers, and connect a batch normalization layer and a Leaky ReLU activation function after each convolutional layer to improve the training speed and the model expression ability; finally connect a max pooling layer to retain important feature information; Step S43: Adjust the output data of the CNN and input it into the multi-head attention mechanism MHA. MHA can adaptively allocate attention weights to different parts of the input data. When a certain feature input is distorted, its weight is reduced to reduce the impact on the positioning result; input the output of MHA into the fully connected layer to perform the final position prediction of the tag; use the Dropout regularization technique in the MHA layer to improve the generalization ability of the model; Step S44: In the output layer, the features output by the fully connected layer are further processed, and linear regression is used to predict the position information of the signal source; the loss function uses the mean square error to calculate the error between the actual position and the predicted position of the tag. The mean square error formula is: where N is the number of samples, x i is the i-th actual x coordinate, y i is the i-th actual y coordinate, the i-th predicted x coordinate, is the i-th predicted y coordinate.

Images