CN115905787A - A high-precision indoor positioning method based on fuzzy transfer learning model - Google Patents

Abstract

The invention discloses a high-precision indoor positioning method based on a fuzzy transfer learning model. In the method, a fingerprint feature set of a target area is firstly collected by a target area data acquisition module; and then data processing is performed on the collected fingerprint feature set by a target area data processing module; Then, the indoor positioning module based on the fuzzy migration model receives the data processed by the data processing module, and analyzes and calculates the target result data according to the data, which is the positioning result data; this method not only realizes high-precision indoor Positioning also reduces various costs.

Description

A high-precision indoor positioning method based on fuzzy transfer learning model

Technical Field

The present invention belongs to the technical field of indoor positioning, and in particular relates to a high-precision indoor positioning method based on a fuzzy transfer learning model.

Background Art

In recent years, with the rapid development of the Internet of Things (IOT) and the popularization of wireless networks, location-based services (LBS) have become more and more deeply integrated into all aspects of people's lives. LBS involves calculating the location of target objects in different environments. The key to LBS is to accurately obtain the user's location. According to the application scenario, LBS is divided into outdoor positioning and indoor positioning.

In outdoor environments, the global navigation satellite system (GNSS) can achieve accurate positioning. However, the positioning granularity of GNSS in indoor environments is large and cannot meet the needs of high-precision positioning in outdoor environments. In indoor environments, since wireless networks are widely deployed in indoor scenes, the method of using wireless signals for indoor positioning has received more and more attention and application. The method of using wireless signals for indoor positioning includes a positioning method based on signal fingerprints. The positioning method based on signal fingerprints does not require the deployment of additional equipment, so it is convenient to apply and more scalable, and has always been a research hotspot in the field of indoor positioning technology. The positioning method based on signal fingerprints generally includes a fingerprint map construction process and a fingerprint matching positioning process. Among them, the fingerprint map construction stage requires offline sampling of the fingerprint database for the entire perception area. If the entire perception area needs to be sampled offline, it will take a lot of manpower and time. In this way, when the sampling area is large, such as when there are multiple floors, the sampling time required for the entire perception area will increase sharply. In particular, when some floors or rooms are not allowed to be opened due to some special reasons during the sampling process, it will lead to the inability to obtain the fingerprint information of these areas manually, which seriously affects the indoor positioning results. In summary, it is a practical problem to be solved urgently to reasonably control the sampling cost (including time cost, labor cost and equipment cost) and ensure the positioning accuracy.

Transfer learning in machine learning methods can achieve reuse in similar or related areas by learning from source domain data, making target learning accumulative learning, thereby reducing the cost of building the target domain model and improving the learning effect. Building a learning model for the source domain is a simple yet effective transfer learning method. Building a learning model for the source domain is to use source domain samples to build an initial model and save the corresponding parameters, then apply the target domain data to the model and fine-tune the parameters to adapt to its own data set. However, in real application scenarios, due to the lack and difference of the sampling space of the target area, there is a difference in the data distribution of the sampled fingerprint and the source domain fingerprint, which results in the positioning model obtained based on the source domain fingerprint not being able to perform well on the target domain fingerprint, resulting in inaccurate indoor positioning.

Therefore, a new indoor positioning method needs to be proposed.

Summary of the invention

Purpose of the invention: In order to overcome the shortcomings of the prior art such as poor indoor positioning accuracy and high indoor positioning cost, the present invention provides a high-precision indoor positioning method based on a fuzzy transfer learning model, which not only achieves high-precision indoor positioning, but also reduces various costs, including time cost, labor cost and equipment cost.

Technical solution: To solve the above technical problems, the present invention provides a high-precision indoor positioning system based on a fuzzy transfer learning model, the system comprising a target area data acquisition module, a target area data processing module, an indoor positioning module based on a fuzzy transfer model and a data output module connected in sequence;

The target area data acquisition module is used to collect the fingerprint feature set of the target area;

The target area data processing module is used to process the fingerprint feature set of the target area collected above to obtain data suitable for processing by the next module;

The indoor positioning module based on the fuzzy migration model is used to receive the data processed by the data processing module, and analyze and calculate the data to obtain the target result data, which is the positioning result data;

The data output module is used to output the positioning result data for user reference;

The method comprises the following steps:

Step 1, firstly, the target area data acquisition module in the high-precision indoor positioning system based on the fuzzy transfer learning model collects the fingerprint feature set of the target area;

Step 2, the target area data processing module in the high-precision indoor positioning system based on the fuzzy transfer learning model then processes the fingerprint feature set of the target area collected above to obtain data suitable for processing by the next module;

Step 3, then the indoor positioning module based on the fuzzy transfer model in the high-precision indoor positioning system based on the fuzzy transfer learning model receives the data processed by the data processing module, and analyzes and calculates the data to obtain target result data, which is the positioning result data;

Step 4: Finally, the data output module in the high-precision indoor positioning system based on the fuzzy transfer learning model outputs the positioning result data for user reference.

Furthermore, generating the indoor positioning module based on the fuzzy migration model specifically includes the following steps:

Step S1, collecting data from a sample area through a sample area data collection module to obtain a fingerprint feature set of the sample area;

Step S2, performing data processing on the fingerprint feature set of the sample area collected above by the sample area data processing module to obtain data suitable for processing by the next module;

Step S3, dividing the fingerprint feature set of the sample area after the data processing by the sample area data sample division module to obtain training set sample data and test set sample data;

Step S4, training the basic model based on the training set sample data to obtain a trained model, which is used as an indoor positioning module based on the fuzzy migration model to analyze and calculate the data set collected in the target area to obtain positioning result data for user reference.

Furthermore, the step S2 comprises the following steps:

S21 Assume that the sampling time length of the sampling point P is Γ, and the short period is τ. Then, in the time period t, that is, in the short period τ after time t, the RSS vector sampled by the sampling point P is expressed by formula (1):

Then in the next time period (t+t→t+2τ), the RSS vector sampled at the sampling point P is expressed by formula (2):

Then, in the time period t, that is, in the short period τ after time t, the RSS vector sampled by the neighbor P _i of the sampling point P is expressed by formula (3):

S22 calculates the ISF of sampling point P at time t using formula (4):

ISF(P,t)=RSS(P,t→t+τ)+RSS(P,t+τ→t+2τ)=[ISf ₁ ,ISf ₂ ,…,ISf _n ] Formula (4)

In the above formula (4),

S23 calculates the CSF of sampling point P at time t by formula (5):

CSF(P,t)=RSS(P,t→t+τ)+RSS(P _i ,t→t+τ)=[CSf ₁ ,CSf ₂ ,…,CSf _n ] Formula (5)

In the above formula (5),

S24 thus obtains the short-term feature set of the sampling point P at time T: SFS(P, t) = ISF(P, t) ∪ CSF(P, t);

S25 thus obtains the short-term feature set of the sampling point P in the entire sampling period Γ, which is expressed by formula (6):

SFS(P, Γ) = ∑ _t∈Γ SFS(P, t) Formula (6)

Through the above steps, the short-term feature set SFS of the sampling points is obtained.

Furthermore, the step S4 comprises the following steps:

S41 First, by optimizing the long short-term memory neural network to train the short-term feature data SFS, we get the OptimizedLSTM, which is used to preliminarily build the Pre-Model of transfer learning;

S42 Then, the idea of attention mechanism is introduced, and the lightweight mechanism SENet for sparse data is used to optimize the Pre-Model of transfer learning preliminarily constructed above to obtain Optimization SE-LSTM, which is the optimized training model.

Furthermore, the S41 includes the following steps: assuming that the short-term feature set input to the LSTM network at time μ is sfs(μ), then the input value is obtained by passing through the Input Gate in the LSTM network and combining the value h(μ-1) at time μ-1 in the Memory Cell:

I(μ)＝σ(w _s，I *sfs(μ)+w _h，I *h(μ-1)) Formula (7)

In the above formula (7), σ represents the activation function; _ws,I and wh _,I are the parameters of function I; sfs(μ) is the short-term feature input at time μ, and h(μ-1) is the value of the Memory Cell at time μ-1;

At the same time, the Input Gate generates a candidate value vector:

的参数；σ表示激活函数；sfs(μ)为μ时刻输入的短时特征，h(μ-1)为Memory Cell中μ-1时刻数值；In the above formula (8), w _s,c and w _h,c are functions

; σ represents the activation function; sfs(μ) is the short-term feature input at time μ, and h(μ-1) is the value of μ-1 in the Memory Cell;

Then, Forget Gate reads the sfs at time μ and the value h at time μ-1, and uses the activation function to output a value in a given interval to indicate the degree of acceptance of the value:

F(μ)＝σ(w _s，f *sfs(μ)+w _h，f *h(μ-1)) Formula (9)

In the above formula (9), w _s,f and w _h,f are parameters of function F; σ represents the activation function; sfs(μ) is the short-term feature input at time μ, and h(μ-1) is the value of the Memory Cell at time μ-1;

Next, according to the above formulas (7), (8) and (9), the status information is updated:

Finally, the Output Gate outputs the result at time μ according to the status information and updates the Memory Cell at the same time. The Memory Cell is updated according to formula (11) as follows:

h(μ)＝O(μ)*tanh(c(μ)) Formula (11)

通过公式(12)计算得到O(μ)：In the above formula (11),

O(μ) is calculated by formula (12):

O(μ)＝σ(w _s，O *Δsfs(μ)+w _h，O *h(μ-1)) Formula (12)

In the above formula (12), the vector difference between adjacent moments Δsfs(μ)=sfs(μ)-sfs(μ-1);

In the above, the short-term feature data SFS is trained through the long short-term memory neural network LSTM to obtain the Optimized LSTM, which is used to preliminarily build the Pre-Model of transfer learning.

Furthermore, the ReLU function is used as the activation function σ.

Further, the S42 comprises the following steps: 7. The high-precision indoor positioning method based on the fuzzy transfer learning model according to claim 6,

First, a globally distributed embedded channel feature response is generated, allowing all layers of the training model to use it, and the feature data is globally averaged pooled to represent the global distribution of responses on the feature channel:

In the above formula (13), H represents the number of elements in SFS, and W represents the length of each element;

Then, the relationship between feature channels is explored, and two layers of nonlinear activation functions are used to learn the nonlinear interactions between channels to obtain appropriate weights:

S(μ)＝σ(w ₂ *δ(w ₁ *Z(μ))) Formula (14)

In the above formula (14), σ represents the sigmoid activation function, δ represents the ReLU activation function, and _w1 and _w2 represent scaling parameters;

Finally, channel-wise multiplication is used to achieve output:

In the above formula (15), S(μ) is the formula (14), O(μ) is the formula (12);

In the above, by introducing SENet, a lightweight mechanism for sparse data, the pre-model O-LSTM of transfer learning initially constructed above is optimized to obtain OSE-LSTM.

Furthermore, the OSE-LSTM is further optimized: first, the fuzzy clustering method is used to establish class labels for the sampling points in the target domain according to the distribution of the sampling points in the source domain area; then, transfer learning is performed using the target domain data marked with class labels and the obtained transfer learning pre-model PreModel, thereby constructing a fuzzy transfer learning model that meets different missing situations and different floors.

Beneficial effects: Compared with the prior art, the present invention has the following advantages:

1. The method of the present invention designs a fingerprint construction method based on short-time features by mining the features of sampling points in a fine-grained manner, which displays the features of fingerprints in a more fine-grained manner. At the same time, due to the sparsity and temporal nature of short-time features, LSTM and SENet are optimized, and the OSE-LSTM model is proposed as a pre-model for transfer learning, which not only reduces the sampling time of the fingerprint map construction process, but also ensures accurate positioning accuracy;

2. The method of the present invention proposes a feature transfer method based on fuzzy clustering. Aiming at the problem of different data distribution caused by the differences in different distributions, the similarity of feature distributions of the source domain and the target domain is ensured, the robustness of transfer learning is enhanced, and the accuracy of indoor positioning is improved;

3. The method of the present invention proposes a fuzzy transfer learning model (FTL), which achieves accurate positioning on different floors and at different sampling rates. The fuzzy transfer learning model is fully compared with the prior art on different floors, at different sampling rates and with different devices. The experimental results prove that the method of the present invention is effective and reliable.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a system structure diagram of the present invention based on the fuzzy transfer learning model.

FIG2 is a flowchart showing the steps of generating an indoor positioning module based on a fuzzy migration model.

Figure 3 is a diagram of the sample area structure.

FIG4 is a block diagram of the implementation of the pre-model Pre-Model for transfer learning.

FIG. 5 is a specific experimental result of migrating the positioning model to each floor in Example 9. ...

FIG. 6 is a diagram showing the experimental results when the sampling rate of the target domain is set to 80% in Example 10.

FIG. 7 is a diagram showing the experimental results when the sampling rate of the target domain is set to 30% in Example 10.

FIG8 is a graph showing the error experimental results of the positioning results at different sampling intervals in Example 11.

DETAILED DESCRIPTION

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

Example 1

The high-precision indoor positioning method based on the fuzzy transfer learning model (Fuzzy Transfer Learning Model for Accuracy Localization) of this embodiment provides a high-precision indoor positioning system based on the fuzzy transfer learning model. Referring to FIG. 1 , the high-precision indoor positioning system based on the fuzzy transfer learning model includes a target area data acquisition module, a target area data processing module, an indoor positioning module based on the fuzzy transfer model, and a data output module that are connected in sequence;

The target area data acquisition module is used to acquire the fingerprint feature set of the target area;

The target area data processing module is used to process the fingerprint feature set of the target area collected above to obtain data suitable for processing by the next module;

The indoor positioning module based on the fuzzy migration model is used to receive the data processed by the data processing module, and analyze and calculate the data to obtain the target result data, which is the positioning result data;

The above data output module is used to output positioning result data for user reference.

When the high-precision indoor positioning system based on the fuzzy transfer learning model is working, the target area data acquisition module first collects the fingerprint feature set of the target area; then the target area data processing module performs data processing on the fingerprint feature set of the target area collected above to obtain data suitable for processing by the next module; then, the indoor positioning module based on the fuzzy transfer model receives the data processed by the data processing module, and analyzes and calculates the data to obtain the target result data, which is the positioning result data; finally, the data output module outputs the positioning result data for user reference.

The high-precision indoor positioning method based on the fuzzy transfer learning model of this embodiment specifically includes the following steps:

Step 1: First, the target area data acquisition module in the high-precision indoor positioning system based on the fuzzy transfer learning model collects the fingerprint feature set of the target area;

Step 2, then, the target area data processing module in the high-precision indoor positioning system based on the fuzzy transfer learning model processes the fingerprint feature set of the target area collected above to obtain data suitable for processing by the next module;

Step 3: Then, the indoor positioning module based on the fuzzy transfer model in the high-precision indoor positioning system based on the fuzzy transfer learning model receives the data processed by the data processing module, and analyzes and calculates the data to obtain target result data, which is the positioning result data;

Step 4: Finally, the data output module in the high-precision indoor positioning system based on the fuzzy transfer learning model outputs the positioning result data for user reference.

The above high-precision indoor positioning system based on the fuzzy transfer learning model can achieve indoor positioning of the target area with high positioning accuracy.

Example 2

The high-precision indoor positioning method based on the fuzzy transfer learning model of this embodiment is based on Example 1. The indoor positioning module based on the fuzzy transfer model proposed in Example 1 generates an indoor positioning module based on the fuzzy transfer model, and specifically includes the following steps, with reference to FIG. 2:

Step S1, collecting data from a sample area through a sample area data collection module to obtain a fingerprint feature set of the sample area;

Step S2, performing data processing on the fingerprint feature set of the sample area collected above by the sample area data processing module to obtain data suitable for processing by the next module;

Step S3, dividing the fingerprint feature set of the sample area after the data processing by the sample area data sample division module to obtain training set sample data and test set sample data;

Step S4, training the basic model based on the training set sample data to obtain a trained model, which is used as an indoor positioning module based on the fuzzy migration model to analyze and calculate the data set collected in the target area to obtain positioning result data for user reference.

Example 3

The high-precision indoor positioning method based on the fuzzy transfer learning model of this embodiment is based on Embodiment 2, wherein in step S1, data is collected from the sample area through the sample area data collection module to obtain a fingerprint feature set of the sample area;

Generally speaking, the feature of each sampling point is the fingerprint composed of the RSS value of the received AP. In the process of constructing the fingerprint map, each sampling point is sampled for a certain period of time to obtain a signal set. Each term in the set records information such as the AP name, RSS value, and timestamp. In order to simplify the representation and suppress the influence of outliers, the RSS value of the same AP obtained at each sampling point is usually averaged as the RSS value of the AP at that point, thereby obtaining a one-dimensional fingerprint vector. Obviously, this method of averaging all sampled signals ignores the real-time characteristics of each AP, resulting in the "filtering" of the signal's detailed fluctuations, and therefore the fingerprint features of the sampling points cannot be obtained in a fine-grained manner.

Therefore, in summary, it is necessary to obtain the fingerprint features of the sampling points in a fine-grained manner. In order to obtain the features of the sampling points in a more fine-grained manner, this embodiment constructs a feature set of the sampling points by extracting the short-time features of the fingerprint signal set of the sampling points, that is, the short-time feature set of the sampling points, which shows the features of the fingerprint in a more fine-grained manner.

In this embodiment, the sample area data processing module performs data processing on the fingerprint feature set of the sample area collected above to obtain a short-term feature set of the sampling point, which specifically includes the following steps:

其中1≤i≤4，在图3中圆环点的位置均为采样点P的近邻P_i的位置，设采样点P的采样时长为Γ，短时段设为τ，则在t时间段内即从时间t后的短时段τ内，采样点P采样到的RSS向量用公式(1)表示为：S21 refers to FIG3 , and assumes that there are n APs in the sample area, and the position coordinates of the sampling point P in the sample area are (x, y). The position of the five-pointed star in FIG3 is the position of the sampling point P, and the coordinates of the neighbor P _i of the sampling point P are

Where 1≤i≤4, the positions of the circular points in Figure 3 are all the positions of the neighbors P _i of the sampling point P. Assume that the sampling time of the sampling point P is Γ, and the short time period is τ. Then, in the time period t, that is, in the short time period τ after time t, the RSS vector sampled by the sampling point P is expressed by formula (1):

Then in the next time period (t+τ→t+2τ), the RSS vector sampled at the sampling point P is expressed by formula (2):

Then, in the time period t, that is, in the short period τ after time t, the RSS vector sampled by the neighbor P _i of the sampling point P is expressed by formula (3):

S22 calculates the ISF of sampling point P at time t using formula (4):

ISF(P, t)=RSS(P, t→t+τ)+RSS(P, t+τ→t+2τ)=[ISf ₁ , ISf ₂ ,…, ISf _n ] Formula (4)

In the above formula (4),

S23 calculates the CSF of sampling point P at time t by formula (5):

CSF(P,t)=RSS(P,t→t+τ)+RSS(P _i ,t→t+τ)=[CSf ₁ , CSf ₂ ,…, CSf _n ] Formula (5)

In the above formula (5),

S24 thus obtains the short-term feature set of the sampling point P at time T: SFS(P, t) = ISF(P, t) ∪ CSF(P, t);

S25 thus obtains the short-term feature set of the sampling point P in the entire sampling period Γ, which is expressed by formula (6):

SFS(P, Γ) = ∑ _t∈Γ SFS(P, t) Formula (6)

The short-term feature set SFS of the sampling point is obtained through the above steps, and the features of the sampling point are obtained in a more fine-grained manner. The model trained based on the short-term feature set further improves the accuracy of the positioning model.

As a preferred embodiment, when calculating ISF and CSF, the same AP may appear in adjacent time periods or adjacent sampling points. Of course, the probability of such a situation is extremely small. In order to unify the RSS vector length, this embodiment averages the RSS values that appear twice, suppressing the impact of the abnormality of the AP in a shorter time period, and further improving the accuracy of the positioning model.

Example 4

The high-precision indoor positioning method based on the fuzzy transfer learning model of this embodiment is based on Embodiment 3, wherein step S3 divides the fingerprint feature set of the sample area after the above data processing by the sample area data sample division module to obtain training set sample data and test set sample data;

This embodiment divides the short-term feature set SFS of the sampling points obtained by Example 2, and the division result is not limited to one. For example, the division result can be: 70% is divided into training set sample data, and the remaining 30% is divided into test set sample data; the division result can also be: 100% is all divided into training set sample data to obtain an effective positioning model.

In this embodiment, all sampled SFSs are trained to obtain an effective positioning model.

In the above step S4, the basic model is trained based on the training set sample data to obtain a trained model. The model is used as an indoor positioning module based on the fuzzy migration model to analyze and calculate the data set collected in the target area to obtain positioning result data for user reference. In this embodiment, in order to more effectively process the time series data SFS, a long short-term memory neural network (Long Short-term Memory, LSTM) model is selected as the basic model, and the basic model is trained based on all sampled SFS to construct a pre-model Pre-Model for transfer learning, which specifically includes the following steps:

S41 First, by optimizing the long short-term memory neural network (LSTM) to train the SFS data, we get the Optimized LSTM (O-LSTM) for the preliminary construction of the Pre-Model for transfer learning.

S42Then, due to the different construction methods of ISF and CSF in SFS, the idea of attention mechanism is further introduced, and a lightweight mechanism SENet (Squeeze-and-Excitation Network) for sparse data is used to optimize the above-mentioned preliminarily constructed transfer learning Pre-Model to obtain Optimization SE-LSTM (OSE-LSTM) to enhance the positioning accuracy of Pre-Model.

The optimized training model obtained through the above steps is used as the pre-model Pre-Model for transfer learning. In actual application, the pre-model Pre-Model for transfer learning is migrated to the target area. Regardless of whether the target area is adjacent to the sample area, high-precision indoor positioning can be achieved.

Example 5

The high-precision indoor positioning method based on the fuzzy transfer learning model of this embodiment is based on Embodiment 4, wherein step S41: First, by optimizing the long short-term memory neural network (Long Short-term Memory, LSTM) to train the SFS data, an Optimized LSTM (O-LSTM) is obtained, which is used to preliminarily construct the Pre-Model of transfer learning; specifically, the following steps are included: Since LSTM uses Memory Cell to selectively store information, and uses three gating states of Input Gate, Forget Gate and Output Gate to control the transmitted data in real time, this embodiment mainly includes the following three steps: 1) Input the short-term feature set sfs, first generate the input value through the Input Gate; 2) Then, the Forget Gate chooses to forget the information of the previous moment in the Memory Cell; 3) Finally, the Output Gate determines whether to output the information at this moment. Specifically: In the process of model training, suppose the short-term feature sfs (μ) is input at the moment μ, then it passes through the InputGate in the LSTM network, and is combined with the value h (μ-1) at the moment μ-1 in the Memory Cell to obtain the input value:

I(μ)＝σ(w _s，I *sfs(μ)+w _h，I *h(μ-1)) Formula (7)

In the above formula (7), σ represents the activation function; _ws,I and wh _,I are the parameters of function I; sfs(μ) is the short-term feature input at time μ, and h(μ-1) is the value of the Memory Cell at time μ-1;

At the same time, the Input Gate generates a candidate value vector:

的参数；σ表示激活函数；sfs(μ)为μ时刻输入短时特征，h(μ-1)为Memory Cell中μ-1时刻数值；In the above formula (8), w _s,c and w _h,c are functions

Parameters; σ represents the activation function; sfs(μ) is the short-term feature input at time μ, and h(μ-1) is the value of the Memory Cell at time μ-1;

Then, Forget Gate reads the sfs at time μ and the value h at time μ-1, and uses the activation function to output a value in a given interval to indicate the degree of acceptance of the value:

F(μ)＝σ(w _s，f *sfs(μ)+w _h，f *h(μ-1)) Formula (9)

In the above formula (9), w _s,f and w _h,f are parameters of function F; σ represents the activation function; sfs(μ) is the short-term feature input at time μ, and h(μ-1) is the value of the Memory Cell at time μ-1;

Next, according to the above formulas (7), (8) and (9), the status information is updated:

为公式(8)；c(μ)为当前状态信息，c(μ-1)为上一时刻状态信息；In the above formula (10), F(μ) is the formula (9); I(μ) is the formula (7);

is formula (8); c(μ) is the current state information, c(μ-1) is the state information at the previous moment;

Finally, the Output Gate outputs the result at time μ according to the status information and updates the Memory Cell at the same time. The Memory Cell is updated according to formula (11) as follows:

h(μ)＝O(μ)*tanh(c(μ)) Formula (11)

通过公式(12)计算得到O(μ)：In the above formula (11),

O(μ) is calculated by formula (12):

When SFS is used to train the positioning model, the sparsity of h _μ-1 and x _μ is large, so the similarity of the input data at adjacent moments is large. According to the time dependency of the input vector, the vector difference Δsfs(μ) at adjacent moments is used to generate the output data:

O(μ)＝σ(w _s，O *Δsfs(μ)+w _h，O *h(μ-1)) Formula (12)

In the above formula (12), Δsfs(μ)=sfs(μ)-sfs(μ-1);

This embodiment uses the above method for updating. By optimizing the long short-term memory neural network (Long Short-term Memory, LSTM) to train the SFS data, an Optimized LSTM is obtained for the preliminary construction of the Pre-Model for transfer learning, which greatly reduces the multiplication operations and data loading in the training process and reduces the training time of the model.

In this embodiment, the ReLU function is used as the activation function σ because it is more suitable for sparse representation and effectively suppresses the gradient vanishing problem of SFS data during training, making the entire network easier to converge and further improving the positioning accuracy.

Example 6

The high-precision indoor positioning method based on the fuzzy transfer learning model of this embodiment is based on Example 5. In view of the sparsity of SFS, ISF represents its own fingerprint features, while CSF represents the associated features with the surrounding neighbors and has a large amount of data. According to these differences, in order to improve the positioning accuracy, this embodiment introduces the idea of the attention mechanism and uses a lightweight mechanism SENet (Squeeze-and-Excitation Network) for sparse data to optimize the above-mentioned preliminarily constructed transfer learning Pre-Model to obtain Optimization SE-LSTM (OSE-LSTM), which better expresses the relationship between features to enhance the positioning accuracy of the Pre-Model. Specifically:

First, a globally distributed embedded channel feature response is generated, allowing all layers of the training model to use it, and global average pooling is performed on the feature data to represent the global distribution of responses on the feature channel:

In the above formula (13), H represents the number of elements in SFS, and W represents the length of each element;

Then, the relationship between feature channels is explored, and two layers of nonlinear activation functions are used to learn the nonlinear interactions between channels to obtain appropriate weights:

S(μ)＝σ(w ₂ *δ(w ₁ *Z(μ))) Formula (14)

In the above formula (14), σ represents the sigmoid activation function, δ represents the ReLU activation function, and _w1 and _w2 represent scaling parameters;

Finally, channel-wise multiplication is used to achieve output:

In the above formula (15), S(μ) is the same as formula (14), and O(μ) is the same as formula (12).

它能够兼顾短时数据的时序性和稀疏性，同时考虑了SFS中ISF和CSF两类生成策略不同的数据，实现了准确的定位精度。Referring to Figure 4, Figure 4 is a block diagram of the implementation of the pre-model Pre-Model of transfer learning. Based on the short-term feature set SFS, LSTM is trained to obtain O(μ), and then SENet (Squeeze-and-Excitation Network) is introduced to obtain S(μ), and finally

It can take into account the temporal sequence and sparsity of short-term data, and at the same time consider the two types of data with different generation strategies, ISF and CSF in SFS, to achieve accurate positioning accuracy.

Example 7

The high-precision indoor positioning method based on the fuzzy transfer learning model of this embodiment is based on Example 6. Transfer learning can transfer the knowledge learned when solving tasks in the source domain to the target domain, and use the small amount of target domain data to construct a model with generalization ability. Since the source domain interval is complete, the PreModel constructed using the sampled data in this area can be used to migrate and reuse to the target domain interval where samples are missing. The effect of transfer learning depends greatly on the distribution similarity between the source domain data and the target domain data. However, the degree of missingness (position, size, etc.) of the target domain is uncertain, resulting in different differences in the distribution of the sampled data and the source domain data. Therefore, it is difficult to achieve accurate positioning of the target domain by simply using fine-tuning.

This embodiment further optimizes the fuzzy transfer learning model. First, using the fuzzy clustering method, class labels are established for the sampling points in the target domain according to the distribution of the sampling points in the source domain area, so that the source domain data and the target domain data are divided by class, which avoids the difference in the overall data distribution between the two affecting the effect of transfer learning; then, transfer learning is performed using the target domain data marked with class labels and the obtained transfer learning pre-model PreModel, so as to construct a fuzzy transfer learning model that meets different missing situations and different floors. Specifically, the following steps are included:

(1) Use traditional clustering methods to cluster the sampling points in the source domain and mark the class labels. The features of the sampling points are short-term feature sets (SFS);

(2) According to the number of categories in the source domain, the target domain data is clustered using the fuzzy clustering method, and each sampling point is given a cluster label TL _f ;

(3) Since the sampling spaces of the target domain data and the source domain data are similar (or even identical), the sampling points at the corresponding positions in the target domain are labeled according to the class labels of the sampling points in the source domain to obtain a theoretical category label TL _t ;

(4) If the TL _f and TL _t of the element sfs(P) of the short-term feature set SFS of the sampling point P in the source domain data are the same or adjacent, the category label of sfs(P) is set to TL _f ;

(5) If TL _f and TL _t are not adjacent, then:

When there is a class TL _x adjacent to them, the class label of sfs(P) is set to TL _x ;

If there is no class adjacent to them, the record is deleted, that is, SFS(P) = SFS(P)-{sfs(P)};

Through the above steps (1)-(5), a category label is added to the target domain data. The target domain data marked with the class label and PreModel are used for transfer learning to build a fuzzy transfer learning model that meets different missing situations and different floors, thereby further optimizing the fuzzy transfer learning model.

In actual application, the optimized fuzzy transfer learning model is migrated to the target area, and high-precision indoor positioning can be achieved regardless of whether the target area is adjacent to the sample area.

Example 8

The high-precision indoor positioning method based on the fuzzy transfer learning model of this embodiment is based on Example 7. This embodiment selects a teaching building as a positioning experiment scene. It has five floors with similar structures, and the area of each floor is ^1450m2 . The experiment uses three types of devices for sampling, namely 5 smart phones, 2 tablets and 1 PDA. In order to fully verify the positioning effect of the present invention, this experiment has completely sampled all 5 floors, with a sampling point density of 1.2*1.2 meters and a sampling height of 1 meter. The acquisition time of each sampling point is 3 minutes. This experiment compares the performance of different positioning models (LSTM, SE-LSTM and OSE-LSTM), better verifies the robustness of the positioning model proposed in this article, and conducts comparative experiments on different positioning methods. The sampling interval is fixed to 1.2 meters, and a floor is randomly selected as the experimental area. Five devices, HUAWEI Mate7, HUAWEI Mate8, Honor, VIVO x6 and MI 6, are used to compare the model performance, and they are numbered as devices 1-5 respectively. The average error distance AED is used as the positioning performance evaluation standard. The experimental results are shown in Table 1 below. Table 1 shows the average positioning error distance of different positioning methods (unit: m):

Table 1

As shown in Table 1, the positioning result finally obtained by the OSE-LSTM method of the present invention has the lowest error, which further illustrates that the OSE-LSTM method proposed by the present invention can achieve high-precision indoor positioning.

Example 9

The high-precision indoor positioning method based on the fuzzy transfer learning model of this embodiment is based on Example 7. In order to better compare the positioning performance differences of the migration of the positioning model, this embodiment selects the first-floor corridor of the teaching building as the experimental area. In order to avoid the negative impact of device heterogeneity on the current experiment, the average value of the fingerprint features of the data collected by five devices, HUAWEI Mate7, HUAWEI Mate8, Honor, VIVO x6 and MI 6 on this floor is used as the initial data for model training and positioning accuracy test, and the sampling interval is 1.2 meters. Among them, the training set is 70% and the test set is 30%. The average error distance AED is used as the positioning performance evaluation standard. The basic positioning model OSE-LSTM trained with all fingerprint point data on the first floor is selected, and then the positioning model OSE-LSTM is migrated to other floors. And fine-tuning training is performed by using different sampling rates for the target domain data, and experimental verification is performed using the traditional transfer learning method and the fuzzy migration method proposed in this article. Finally, the experiment compares the positioning results of the traditional WKNN algorithm at different sampling rates on different floors. The specific experimental results of migrating the positioning model to each floor are shown in Table 2 and Figure 5, where AF represents adjacent floors, NAF represents non-adjacent floors (third, fourth, and fifth floors), TTL represents traditional transfer learning, and FTL represents fuzzy transfer learning. Table 2 shows the average positioning error distance (unit: m) of traditional transfer learning and fuzzy transfer learning using different sampling rates on different floors:

Table 2

From the above experimental results, it can be seen that with the decrease of sampling density, the positioning accuracy of the model also declines to varying degrees. However, compared with the traditional transfer learning method, the fuzzy transfer learning method proposed in this paper has higher accuracy, and its performance when migrating to adjacent floors is significantly better than that of non-adjacent floors, and it can better adapt to complex and changing environments.

Example 10

The high-precision indoor positioning method based on the fuzzy transfer learning model of this embodiment is based on Example 7. This embodiment verifies the positioning performance results of different devices when fuzzy migration to different floors through the OSE-LSTM model. The data of the pre-trained model comes from all the fingerprint features of the first floor, using five devices: HUAWEI Mate7, MI 6, HUAWEI HonorT1-823, MI Pad 3, and UROVO i6300A, numbered as devices 1-5 respectively. The sampling rates of the target domains are set to 80% and 30% respectively. The training method uniformly uses fine-tuning training, and the average error distance AED is used as the positioning performance evaluation standard. When the sampling rate of the target domain is 80%, the experimental results are shown in Table 3 and Figure 6 below. Table 3 shows the average positioning error distance (unit: m) of different floors at a sampling rate of 80%:

Table 3

The experimental results are shown in Table 4 and Figure 7 when the sampling rate of the target domain is 30%. Table 3 shows the average positioning error distance of different floors when the sampling rate is 80% (unit: m):

Table 4

It can be seen from the experimental results that the positioning error is large when migrating to non-adjacent floors. The reason is that the farther the floors are from each other, the greater the difference in their spatial structure distribution, resulting in a larger data distribution between the source domain and the target domain. At the same time, when the sampling rate is low, the positioning error between different devices after the positioning model is migrated to different floors has a certain jitter. The fuzzy transfer learning method proposed in this embodiment has higher robustness when migrating to other floors.

Embodiment 11

The high-precision indoor positioning method based on the fuzzy transfer learning model of this embodiment is based on Example 7. In order to verify the positive and negative effects of different sampling densities on the positioning model, this embodiment selects data from sampling points of different intervals collected by five devices on a certain floor of a teaching building as the original data, and uniformly uses the OSE-LSTM positioning model. The minimum sampling rectangles are 1.2*1.2, 1.2*2.4, and 2.4*2.4 (unit: meter). The training set is 70% and the test set is 30%. The experimental results are shown in Figure 8. As can be seen from Figure 8, as the sampling interval expands, the error tends to increase. But overall, the influence of different sampling intervals on the error of the positioning result is small.

From the above experiments, it can be seen that the present invention has high robustness between different devices. When the sampling rate is 80%, the positioning error of migration to the adjacent layer is only 1.38 meters, and when the sampling rate is 30%, the positioning error of migration to the adjacent layer is only 1.92 meters, which can greatly reduce the sampling work of fingerprint data while ensuring the positioning accuracy. At the same time, compared with traditional migration, the present invention improves by 18.1% and 12.6% respectively.

The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims (8)

1. A high-precision indoor positioning method based on a fuzzy transfer learning model, characterized in that a high-precision indoor positioning system based on a fuzzy transfer learning model is provided, the system comprising a target area data acquisition module, a target area data processing module, an indoor positioning module based on a fuzzy transfer model, and a data output module connected in sequence; The target area data acquisition module is used to collect the fingerprint feature set of the target area; The target area data processing module is used to process the fingerprint feature set of the target area collected above to obtain data suitable for processing by the next module; The indoor positioning module based on the fuzzy migration model is used to receive the data processed by the data processing module, and analyze and calculate the data to obtain the target result data, which is the positioning result data; The data output module is used to output the positioning result data for user reference; The method comprises the following steps: Step 1, firstly, the target area data acquisition module in the high-precision indoor positioning system based on the fuzzy transfer learning model collects the fingerprint feature set of the target area; Step 2, the target area data processing module in the high-precision indoor positioning system based on the fuzzy transfer learning model then processes the fingerprint feature set of the target area collected above to obtain data suitable for processing by the next module; Step 3, then the indoor positioning module based on the fuzzy transfer model in the high-precision indoor positioning system based on the fuzzy transfer learning model receives the data processed by the data processing module, and analyzes and calculates the data to obtain target result data, which is the positioning result data; Step 4: Finally, the data output module in the high-precision indoor positioning system based on the fuzzy transfer learning model outputs the positioning result data for user reference. 2. The high-precision indoor positioning method based on the fuzzy transfer learning model according to claim 1, characterized in that generating the indoor positioning module based on the fuzzy transfer model specifically comprises the following steps: Step S1, collecting data from a sample area through a sample area data collection module to obtain a fingerprint feature set of the sample area; Step S2, performing data processing on the fingerprint feature set of the sample area collected above by the sample area data processing module to obtain data suitable for processing by the next module; Step S3, dividing the fingerprint feature set of the sample area after the data processing by the sample area data sample division module to obtain training set sample data and test set sample data; Step S4, training the basic model based on the training set sample data to obtain a trained model, which is used as an indoor positioning module based on the fuzzy migration model to analyze and calculate the data set collected in the target area to obtain positioning result data for user reference. 3. The high-precision indoor positioning method based on the fuzzy transfer learning model according to claim 2, characterized in that step S2 comprises the following steps: S21 Assume that the sampling time length of the sampling point P is Γ, and the short period is τ. Then, in the time period t, that is, in the short period τ after time t, the RSS vector sampled by the sampling point P is expressed by formula (1):

Then in the next time period (t+τ→t+2τ), the RSS vector sampled at the sampling point P is expressed by formula (2):

Then, in the time period t, that is, in the short period τ after time t, the RSS vector sampled by the neighbor P _i of the sampling point P is expressed by formula (3):

S22 calculates the ISF of sampling point P at time t using formula (4): ISF(P,t)=RSS(P,t→t+τ)+RSS(P,t+τ→t+2τ)=[ISf ₁ ,ISf ₂ ,…,ISf _n ] Formula (4) In the above formula (4),

S23 calculates the CSF of sampling point P at time t by formula (5): CSF(P,t)=RSS(P,t→t+τ)+RSS(P _i ,t→t+τ)=[CSf ₁ ,CSf ₂ ,…,CSf _n ] Formula (5) In the above formula (5),

S24 thus obtains the short-term feature set of the sampling point P at time T: SFS(P, t) = ISF(P, t) ∪ CSF(P, t); S25 thus obtains the short-term feature set of the sampling point P in the entire sampling period Γ, which is expressed by formula (6): SFS(P,Γ)＝∑ _t∈Γ SFS(P,t) Formula (6) Through the above steps, the short-term feature set SFS of the sampling points is obtained. 4. The high-precision indoor positioning method based on the fuzzy transfer learning model according to claim 3, characterized in that step S4 comprises the following steps: S41 First, by optimizing the long short-term memory neural network to train the short-term feature data SFS, we get the Optimized LSTM, which is used to preliminarily build the Pre-Model for transfer learning; S42 Then, the idea of attention mechanism is introduced, and the lightweight mechanism SENet for sparse data is used to optimize the Pre-Model of transfer learning constructed above to obtain Optimization SE-LSTM, which is the optimized training model. 5. The high-precision indoor positioning method based on the fuzzy transfer learning model according to claim 4 is characterized in that the S41 comprises the following steps: assuming that the short-term feature set input to the LSTM network at the μ moment is sfs(μ), then passing through the Input Gate in the LSTM network, and combining the μ-1 moment value h(μ-1) in the Memory Cell, the input value is obtained: I(μ)＝σ(w _s,I *sfs(μ)+w _h,I *h(μ-1)) Formula (7) In the above formula (7), σ represents the activation function; w _s,I and w _h,I are the parameters of function I; sfs(μ) is the short-term feature input at time μ, and h(μ-1) is the value of the Memory Cell at time μ-1; At the same time, the Input Gate generates a candidate value vector:

In the above formula (8), w _s,c and w _h,c are functions

; σ represents the activation function; sfs(μ) is the short-term feature input at time μ, and h(μ-1) is the value of μ-1 in the Memory Cell; Then, Forget Gate reads the sfs at time μ and the value h at time μ-1, and uses the activation function to output a value in a given interval to indicate the degree of acceptance of the value: F(μ)＝σ(w _s,f *sfs(μ)+w _h,f *h(μ-1)) Formula (9) In the above formula (9), _ws,f and wh _,g are the parameters of the function F; σ represents the activation function; sfs(μ) is the short-term feature input at time μ, and h(μ-1) is the value of the Memory Cell at time μ-1; Next, according to the above formulas (7), (8) and (9), the status information is updated:

Finally, the Output Gate outputs the result at time μ according to the status information and updates the Memory Cell at the same time. The Memory Cell is updated according to formula (11) as follows: h(μ)＝O(μ)*tanh(c(μ)) Formula (11) In the above formula (11),

O(μ) is calculated by formula (12): O(μ)＝σ(w _s,O *Δsfs(μ)+w _h,O *h(μ-1)) Formula (12) In the above formula (12), the vector difference between adjacent moments Δsfs(μ)=sfs(μ)-sfs(μ-1); In the above, the short-term feature data SFS is trained through the long short-term memory neural network LSTM to obtain the Optimized LSTM, which is used to preliminarily build the Pre-Model of transfer learning. 6. The high-precision indoor positioning method based on the fuzzy transfer learning model according to claim 5 is characterized in that a ReLU function is used as the activation function σ. 7. The high-precision indoor positioning method based on the fuzzy transfer learning model according to claim 6, characterized in that the S42 comprises the following steps: First, a globally distributed embedded channel feature response is generated, allowing all layers of the training model to use it, and the feature data is globally averaged pooled to represent the global distribution of responses on the feature channel:

In the above formula (13), H represents the number of elements in SFS, and W represents the length of each element; Then, the relationship between feature channels is explored, and two layers of nonlinear activation functions are used to learn the nonlinear interactions between channels to obtain appropriate weights: S(μ)＝σ(w ₂ *δ(w ₁ *Z(μ))) Formula (14) In the above formula (14), σ represents the sigmoid activation function, δ represents the ReLU activation function, and _w1 and _w2 represent scaling parameters; Finally, channel-wise multiplication is used to achieve output:

In the above formula (15), S(μ) is the formula (14), O(μ) is the formula (12); In the above, by introducing SENet, a lightweight mechanism for sparse data, the pre-model O-LSTM of transfer learning initially constructed above is optimized to obtain OSE-LSTM. 8. According to claim 7, the high-precision indoor positioning method based on the fuzzy transfer learning model is characterized in that the OSE-LSTM is further optimized: first, a fuzzy clustering method is used to establish class labels for the sampling points in the target domain according to the distribution of the sampling points in the source domain area; then, transfer learning is performed using the target domain data marked with class labels and the obtained transfer learning pre-model PreModel, thereby constructing a fuzzy transfer learning model that meets different missing situations and different floors.

Images