CN117222003A - Wi-Fi and Bluetooth signal fusion method applied to indoor positioning - Google Patents

Abstract

The invention belongs to the field of indoor positioning and discloses a Wi-Fi and Bluetooth signal fusion method applied to indoor positioning, which mainly focuses on extracting and merging the deep features of Wi-Fi and Bluetooth signals. This method aims at the problem of missing signals in part of the collected signals. It uses the Kalman filter algorithm to compensate for missing values; it uses the long short-term memory network to extract features of Wi‑Fi and Bluetooth signals to capture the time series dependence of the signals; it introduces The multi-head self-attention model captures the relationship between signal features, assigns weights to each signal feature, and combines it with the original input to obtain a weighted feature representation; a cascade strategy is used to efficiently integrate weighted Wi‑Fi and Bluetooth signal characteristics. The present invention provides a new fusion strategy for Wi-Fi and Bluetooth signals, with the purpose of capturing their key information more effectively, thereby achieving error reduction in a cost-effective positioning solution.

Description

A Wi-Fi and Bluetooth signal fusion method applied to indoor positioning

Technical field

The invention belongs to the field of indoor positioning, and in particular relates to a Wi-Fi and Bluetooth signal fusion method applied to indoor positioning.

Background technique

Location-Based Services (LBS) have significant social and commercial significance in determining the precise location of objects or people, and have gradually become the focus of academia and industry. Although navigation and positioning technology based on satellite signals has achieved significant success in outdoor positioning, in indoor environments, the positioning effect of satellite signals is still unsatisfactory due to severe attenuation and multipath effects.

In order to improve the accuracy of indoor positioning, researchers began to explore indoor positioning technology using other signals. Examples include Wi-Fi and pedestrian dead reckoning. Wi-Fi positioning mainly relies on the existing wireless local area network infrastructure and performs triangulation positioning by measuring signal strength or propagation delay. Pedestrian dead reckoning mainly relies on wearable sensors, such as accelerometers and gyroscopes, to detect and estimate pedestrians' steps and walking directions to achieve positioning. While each of these positioning technologies has unique advantages, they also suffer from some common limitations. For example, Wi-Fi positioning may cause errors due to signal obstruction or reflection. Due to instability of Wi-Fi signals, some signals may be missing during the data collection process. The accuracy of pedestrian dead reckoning will gradually accumulate errors over time. In order to overcome these limitations, many researchers are working hard to explore methods to fuse multiple signals, hoping to improve the accuracy and robustness of indoor positioning through this fusion. For example, technologies that integrate Wi-Fi, geomagnetism and pedestrian dead reckoning, and technologies that integrate Wi-Fi and images have all shown good positioning effects in experiments.

Positioning technology that integrates multiple signal sources reduces indoor positioning errors to a certain extent. However, these convergence solutions often involve the integration of multiple hardware and software components, and the overall deployment and maintenance costs are high. This increase in cost may limit its widespread use in commercial applications and daily life.

Contents of the invention

The purpose of the present invention is to provide a Wi-Fi and Bluetooth signal fusion method for indoor positioning to solve the technical problem of high deployment and maintenance costs of positioning technology that integrates multiple signal sources.

In order to solve the above technical problems, the present invention aims to propose a new positioning solution that integrates Wi-Fi and Bluetooth signals. The goal of this solution is to reduce the deployment cost while ensuring the accuracy of the positioning system, thereby achieving economical, efficient and robust indoor positioning. The specific technical solutions of the present invention are as follows:

A Wi-Fi and Bluetooth signal fusion method applied to indoor positioning, including the following steps:

Step 1: Set reference points according to certain rules, and collect Wi-Fi signals and Bluetooth signals at predetermined collection points; organize the collected data in chronological order;

Step 2: For the signal missing problem in the data, use the Kalman filter algorithm to calculate the missing signal value through the iteration of the prediction step and the update step;

Step 3: For the signal after filling in missing values, find the maximum and minimum values in each column of data, and use the maximum and minimum normalization method to normalize each element;

Step 4: Use two long short-term memory networks to extract the features of the two types of data respectively, that is, create a separate long short-term memory network for each type of data, one network is used to process Wi-Fi signals, and the other network is used to Process the Bluetooth signal and finally output the extracted features;

Step 5: The output of each long short-term memory network will be further passed into a multi-head self-attention model, and the corresponding attention weight will be calculated for each element based on the similarity within the sequence;

Step 6: Fusion of the weighted Wi-Fi and Bluetooth signals through cascade to obtain a comprehensive and rich feature representation.

Further, the step one includes the following steps:

Select a space as the experimental area and divide the space into a two-dimensional plane grid. Each grid is a square with a side length of 0.6 meters. The intersection of the grid is the signal collection point, and the distance between two adjacent collection points is 1.2 meters.

Furthermore, the Kalman filter in step 2 is a recursive estimation algorithm, which consists of a prediction step and an update step. The collected data is a time series, and the position of the missing value is left blank. The algorithm will predict based on the state of the previous moment. , get the predicted value of the signal strength at the current moment, and then compare the actual signal strength measurement value with the predicted value, and then correct the predicted value. After repeated iterations, the algorithm finally predicts the predicted value of the missing signal, ensuring that it is close to True signal strength.

Further, the specific steps of step two are as follows:

Let x _y represent the signal strength at the y-th position, let F represent the state transition matrix, and w _y-1 represent the process noise, then the state transition model is obtained:

x _y ＝Fx _y-1 +w _y-1 (1)

Let z _y represent the observed value of the signal strength value at the y-th position, H represent the observation matrix, and u _y represent the observation noise, then the observation model is obtained:

z _y =Hx _y +u _y (2)

Initialize the Kalman filter parameters, x _0|0 is the initial state estimate, P _0|0 is the error covariance of the initial state estimate, O process noise covariance, and R observation noise covariance.

Predict the next signal strength value:

x _y|y-1 =Fx _y-1|y-1 (3)

Let F ^T represent the transpose of the state transition matrix, and the error covariance prediction is:

P _y|y-1 =FP _y-1|y-1 F ^T +O(4)

Let H ^T represent the transpose of the observation matrix, and the Kalman gain is:

Y _y ＝P _y|y-1 H ^T (HP _y|y-1 H ^T +R) ^-1 (5)

Update the signal strength value, and the status update equation is:

x _y|y ＝x _y|y-1 +Y _y (z _y -Hx _y|y-1 )(6)

Let I be the identity matrix, and the error covariance update equation is:

P _y|y =(I-HY _y )P _y|y-1 (7)

Using the above formula to calculate, x _y|y-1 will fill in the signal value at the missing position.

Further, the specific steps of step three are as follows:

For the normalization of Wi-Fi signals, first find the maximum value x _max and the minimum value x _min for each column of data, and use them to normalize each data x _i in the column through formula (8) Unification:

Among them, x _i ^′ is the normalized data;

The normalization of Bluetooth signals is the same, normalizing it to the range of [0,1].

Further, the specific steps of step four are as follows:

One network is used to process Wi-Fi signals, and the other network is used to process Bluetooth signals. The two types of data are in the long short-term memory network. The data undergoes a complex internal structure process, filtering through input gates, forgetting gates, and unit states. The update enables the network to learn and remember long-term and short-term dependencies, and finally obtains meaningful hidden states and outputs extracted features.

Further, the multi-head self-attention model in step 5 initializes three weight parameter matrices w _Q , w _K , w _V ∈R ^d×d , and multiplies them with the input to obtain Q, K, V∈R ^d×d respectively. Among them, Q, K and V respectively represent the query matrix, key matrix and value matrix. The attention module uses QKV to iteratively capture the dependencies between different features from the feature vector sequence, and then obtain different Attention weights, through a multi-head mechanism, ensure that each head focuses on different information. Finally, these attention weights are combined with the output of the original long short-term memory network to obtain a weighted output that reflects the various internal aspects of the sequence. Dependencies.

Further, the specific steps of step five are as follows:

Use the Softmax function to calculate the attention weight:

Among them, Q is the query matrix, K is the key matrix, V is the value matrix, d _k is the dimension of the key matrix, and A(Q,K,V) is the calculated result.

Further, the specific steps of step six are as follows:

The new, fused data set contains all key information from both data sources:

Among them, after calculating the multi-head attention, Represents the eigenvector of the Wi-Fi signal, /> Represents the feature vector of Bluetooth signal, [·;·] indicates feature cascade.

A Wi-Fi and Bluetooth signal fusion method applied to indoor positioning of the present invention has the following advantages:

1. Using Kalman filtering to process data can effectively estimate the state of the noise and uncertainty existing in the data, thereby filling in missing values and providing a more accurate data representation. Through this method, the noise, uncertainty and dynamic changes of the data are captured and corrected by the Kalman filter, ensuring that the data is processed accurately and stably without interference from other possible data, thus supplementing missing values and optimizing the data Integrity and quality.

2. Use two long short-term memory networks to extract the characteristics of two types of data respectively, which can effectively deal with each type of data having its own unique characteristics and structure, and capture its inherent information. In this strategy, all characteristics and patterns of Wi-Fi signals are learned and extracted by the first long short-term memory network without any interference from Bluetooth signals. Likewise, the Bluetooth signal is processed by a second dedicated long-short-term memory network to ensure its unique characteristics are fully exploited.

3. Use a multi-head self-attention mechanism to fuse the characteristics of the two data to further optimize this strategy. Give higher weight to those features that are important and lower the weight to those features that are not important. This approach allows the model to better capture the internal structure and relationships of the data and automatically select those important features. In this way, in addition to each capturing the characteristics of their respective data, the correlation between the two data can also be learned.

Description of drawings

Figure 1 is a flow chart of the Wi-Fi and Bluetooth signal fusion method applied to indoor positioning according to the present invention;

Figure 2 is a schematic diagram of the two-dimensional plane mesh division of the present invention.

Detailed ways

In order to better understand the purpose, structure and function of the present invention, a Wi-Fi and Bluetooth signal fusion method for indoor positioning of the present invention will be described in further detail below with reference to the accompanying drawings.

As shown in Figure 1, a Wi-Fi and Bluetooth signal fusion method for indoor positioning of the present invention includes the following steps:

Step 1: According to the research needs and the characteristics of the indoor environment, in order to ensure that Wi-Fi signals and Bluetooth signals can be collected comprehensively and systematically, it is necessary to set reference points according to certain rules, and measure Wi-Fi signals at predetermined collection points. Collect Fi signals and Bluetooth signals; organize the collected data in chronological order. In this embodiment, a classroom is selected as the experimental area, as shown in Figure 2. This space is divided into two-dimensional plane grids. Each grid is a square with a side length of 0.6 meters, and the intersection points of the grids are signal collection points. The circles in Figure 2 represent signal collection points, and the distance between two adjacent collection points is 1.2 meters.

Step 2: For the signal missing problem in the data, use the Kalman filter algorithm to calculate the missing signal value through the iteration of the prediction step and the update step. Kalman filtering is a recursive estimation algorithm consisting of a prediction step and an update step. The collected data is a time series, and the positions of missing values are left blank. The algorithm will predict based on the status of the previous moment and obtain the predicted value of the signal strength at the current moment. The actual signal strength measurement value is then compared with the predicted value, and then the predicted value is corrected. After repeated iterations, the algorithm finally predicts the predicted value of the missing signal to ensure that it is close to the true signal strength. The specific steps are as follows:

Let x _y represent the signal strength at the y-th position, let F represent the state transition matrix (usually 1 in a stable environment), and w _y-1 represent the process noise, then the state transition model is obtained:

x _y ＝Fx _y-1 +w _y-1 (1)

Let z _y represent the observed value of the signal strength value at the y-th position, H represent the observation matrix, and u _y represent the observation noise, then the observation model is obtained:

z _y =Hx _y +u _y (2)

Initialize the Kalman filter parameters, x _0|0 is the initial state estimate, P _0|0 is the error covariance of the initial state estimate, O process noise covariance, and R observation noise covariance.

Predict the next signal strength value:

x _y|y-1 =Fx _y-1|y-1 (3)

Let F ^T represent the transpose of the state transition matrix, and the error covariance prediction is:

P _y|y-1 =FP _y-1|y-1 F ^T +O(4)

Let H ^T represent the transpose of the observation matrix, and the Kalman gain is:

Y _y ＝P _y|y-1 H ^T (HP _y|y-1 H ^T +R) ^-1 (5)

Update the signal strength value, and the status update equation is:

x _y|y ＝x _y|y-1 +Y _y (z _y -Hx _y|y-1 )(6)

Let I be the identity matrix, and the error covariance update equation is:

P _y|y =(I-HY _y )P _y|y-1 (7)

Using the above formula to calculate, x _y|y-1 will fill in the signal value at the missing position.

Step 3: For the signal after filling in missing values, find the maximum and minimum values in each column of data, and use the maximum and minimum normalization method to normalize each element. The max-min normalization method, often referred to as normalization, is a technique that rescales data to a fixed interval (usually [0,1]). It can ensure that the data is on a unified scale while keeping the data structure and relationships unchanged, thus helping to improve the performance of the algorithm.

For the normalization of Wi-Fi signals, first find the maximum value x _max and the minimum value x _min for each column of data, and use them to normalize each data x _i in the column through formula (8) Unification:

Among them, x _i ^′ is the normalized data.

The normalization of Bluetooth signals is the same, normalizing it to the range of [0,1].

Step 4: Use two long-short-term memory networks to extract the features of the two types of data respectively, that is, create a separate long-short-term memory network for each type of data. This strategy allows the network to focus more on learning the characteristics of each type of data. Specificity. One network is used to process Wi-Fi signals, and the other network is used to process Bluetooth signals. The two kinds of data are in the long short-term memory network. The data undergoes a complex internal structure process, and passes through the filtering and units of input gates, forgetting gates, etc. The update of the state enables the network to learn and remember long-term and short-term dependencies to ultimately obtain meaningful hidden states and output extracted features.

Step 5: The output of each long short-term memory network will be further passed into a multi-head self-attention model, and the corresponding attention weight will be calculated for each element based on the similarity within the sequence. The model initializes three weight parameter matrices w _Q , w _K , w _V ∈R ^d×d , and multiplies them with the input to obtain Q, K, V∈R ^d×d respectively, where Q, K, and V respectively represent the query (Query ) matrix, key matrix and value matrix. The attention module iteratively captures the dependencies between different features from the feature vector sequence through QKV, and then obtains different attention weights. Through the multi-head mechanism, it is ensured that each head focuses on different information, which further enhances the representation ability of the model. Finally, these attention weights are combined with the output of the original long short-term memory network to obtain a weighted output that better reflects the various internal dependencies in the sequence.

Use the Softmax function to calculate the attention weight:

Among them, Q is the query matrix, K is the key matrix, V is the value matrix, d _k is the dimension of the key matrix, and A(Q,K,V) is the calculated result.

Step 6: Fusion of the weighted Wi-Fi and Bluetooth signals through cascade to obtain a comprehensive and rich feature representation. The new, fused data set contains all key information from both data sources.

Among them, after calculating the multi-head attention, Represents the eigenvector of the Wi-Fi signal, /> Represents the feature vector of Bluetooth signal, [·;·] indicates feature cascade.

The present invention uses a recurrent neural network to perform indoor positioning on original data and fused data, thereby predicting coordinate positions. In order to evaluate the performance of the model, the present invention uses Euclidean distance as an evaluation index to measure the error between predicted coordinates and real coordinates. Euclidean distance is a commonly used calculation method to measure the distance between two points, which can effectively reflect the gap between the predicted value and the true value. By comparing the positioning effect of fused data with the effect of using Wi-Fi and Bluetooth signals alone, the superiority of the fusion strategy can be further evaluated. The detailed positioning error results of the experiment are shown in Table 1.

Table 1 Positioning error results of three types of data

type Positioning error (unit: meters) Wi-Fi signal 3.07 Bluetooth signal 1.33 Fusion signal 1.22

It can be seen from the experimental results that the positioning error of the fused data is significantly smaller than the positioning error of single data, indicating that the fusion method can achieve higher accuracy positioning.

It is understood that the present invention has been described through some embodiments. Those skilled in the art know that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, the features and embodiments may be modified to adapt a particular situation and material to the teachings of the invention without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed here, and all embodiments falling within the scope of the claims of the present application are within the scope of protection of the present invention.

Claims (9)

1. The Wi-Fi and Bluetooth signal fusion method applied to indoor positioning is characterized by comprising the following steps of:

step one: setting reference points according to a certain rule, and respectively collecting Wi-Fi signals and Bluetooth signals on a preset acquisition point; the collected data are arranged according to the time sequence;

step two: for the problem of signal missing in the data, a Kalman filtering algorithm is used, and a signal missing value is calculated through iteration of a prediction step and an updating step;

step three: for the signal after filling the missing value, finding out the maximum value and the minimum value in each column of data, and carrying out normalization calculation on each element by using a maximum and minimum normalization method;

step four: respectively extracting the characteristics of two types of data by using two long-short-period memory networks, namely respectively creating a single long-short-period memory network for each type of data, wherein one network is used for processing Wi-Fi signals, the other network is used for processing Bluetooth signals, and finally outputting the extracted characteristics;

step five: the output of each long-short-term memory network is further transmitted into a multi-head self-attention model, and corresponding attention weights are calculated for each element according to the similarity in the sequence;

step six: and fusing the weighted Wi-Fi signals and the weighted Bluetooth signals in a cascading mode, so that a comprehensive and rich characteristic representation is obtained.

2. The Wi-Fi and bluetooth signal fusion method for indoor positioning of claim 1, wherein the first step comprises the steps of:

a space is selected as an experimental area, the space is divided into two-dimensional plane grids, each grid is a square with a side length of 0.6 meter, the intersection point of the grids is a signal acquisition point, and the interval between two adjacent acquisition points is 1.2 meters.

3. The Wi-Fi and bluetooth signal fusion method for indoor positioning according to claim 1, wherein the kalman filtering in the second step is a recursive estimation algorithm, and comprises a prediction step and an updating step, the collected data is a time sequence, the position of the missing value is left blank, the algorithm predicts based on the state of the previous moment to obtain a predicted value of the signal intensity at the current moment, then the predicted value is corrected by comparing the actual signal intensity measured value with the predicted value, and the algorithm finally predicts the predicted value of the missing signal through repeated iteration to ensure that the predicted value is close to the actual signal intensity.

4. The Wi-Fi and bluetooth signal fusion method for indoor positioning of claim 3, wherein the specific steps of step two are as follows:

let x _y Signal intensity representing the y-th position, let F represent the state transition matrix, w _y-1 Representing process noise, then a state transition model is obtained:

x _y ＝Fx _y-1 +w _y-1 (1)

let z _y An observation value representing the intensity value of the y-th position signal, H representing the observation matrix, u _y Representing observation noise, then an observation model is obtained:

z _y ＝Hx _y +u _y (2)

initializing Kalman filtering parameters, x _0|0 For initial state estimation, P _0|0 And (3) as an error covariance of the initial state estimation, O process noise covariance and R observation noise covariance.

Predicting the next signal strength value:

x _y|y-1 ＝Fx _y-1|y-1 (3)

let F ^T Representing the transpose of the state transition matrix, the error covariance prediction is:

P _y|y-1 ＝FP _y-1|y-1 F ^T +O (4)

let H ^T Representing the transpose of the observation matrix, the Kalman gain is:

Y _y ＝P _y|y-1 H ^T (HP _y|y-1 H ^T +R) ^-1 (5)

updating the signal intensity value, and the state updating equation is as follows:

x _y|y ＝x _y|y-1 +Y _y (z _y -Hx _y|y-1 ) (6)

let I be the identity matrix and the error covariance update equation be:

P _y|y ＝(I-HY _y )P _y|y-1 (7)

calculation using the above formula, x _y|y-1 The signal value of the missing position will be padded.

5. The Wi-Fi and bluetooth signal fusion method for indoor positioning according to claim 1, wherein the specific steps of the third step are as follows:

for the normalization of Wi-Fi signals, the maximum value x of each column of data is found out _max And a minimum value x _min And uses them for each data x in the column _i Normalization is performed by equation (8):

wherein x is _i ^′ Is normalized data;

the normalization of the bluetooth signal is the same as that of the bluetooth signal, and the bluetooth signal is normalized to be in the range of [0,1 ].

6. The Wi-Fi and bluetooth signal fusion method for indoor positioning according to claim 1, wherein the specific steps of the fourth step are as follows:

one network is used for processing Wi-Fi signals, the other network is used for processing Bluetooth signals, two kinds of data undergo a complex internal structure process in a long-term and short-term memory network, and the network can learn and memorize long-term and short-term dependency relationships through screening of input gates and forget gates and updating of unit states, so that meaningful hidden states and output extracted features are finally obtained.

7. The Wi-Fi and bluetooth signal fusion method for indoor positioning according to claim 1, wherein the multi-headed self-attention model of step five initializes three weight parameter matrices w _Q ,w _K ,w _V ∈R ^d×d Multiplying the input to obtain Q, K, V e R ^d×d Wherein Q, K, V represent Query (Query) matrix, key (Key) matrix and Value (Value) matrix respectively, the attention module captures the dependency relationship between different features from the feature vector sequence iteratively through Q-K-V, then obtains different attention weights, ensures that each head focuses on different information through a multi-head mechanism, finally, the attention weights are combined with the output of the original long-short term memory network to obtain weighted output, and the output reflects various internal dependency relationships in the sequence.

8. The Wi-Fi and bluetooth signal fusion method for indoor positioning according to claim 1, wherein the specific steps of the fifth step are as follows:

the attention weight was calculated using a Softmax function:

wherein Q is a query matrix, K is a key matrix, V is a value matrix, d _k Is key momentThe dimensions of the array are such that,

a (Q, K, V) is the result after calculation.

9. The Wi-Fi and bluetooth signal fusion method for indoor positioning according to claim 1, wherein the specific steps of the sixth step are as follows:

the new, fused dataset contains all the key information of both data sources:

wherein after the multi-head attention calculation,feature vector representing Wi-Fi signal, < >>Representing the eigenvector of the bluetooth signal, [; carrying out]Representing a cascade of features.