An indoor wireless positioning method based on federated transfer learning

Abstract

The present invention belongs to the field of indoor positioning technology, and discloses an indoor wireless positioning method based on federated transfer learning, comprising the following steps: S1, collecting RSS in the source domain and the target domain at the edge node, and constructing the source domain and the target domain fingerprint library; S2, the central server initializes the global model, performs federated learning training, and obtains the source domain positioning model after multiple rounds of training until convergence; S3, migrates the source domain positioning model to the target domain, adjusts the model, performs federated transfer learning training, and obtains the target domain positioning model after multiple rounds of training until convergence; S4, inputs the signal characteristics of the device to be located into the positioning model, and the positioning model outputs the location information of the device. The present invention adopts the above-mentioned indoor wireless positioning method based on federated transfer learning, which can quickly adapt to new scenarios, enable the new positioning model to converge quickly, and improve the performance of the model.

Description

Technical Field

[0001] The present invention relates to the field of indoor positioning technology, and in particular to an indoor wireless positioning method based on federated transfer learning. Background Art

[0002] The fingerprint signal positioning method is that the signal detection device collects Wi-Fi signals in the positioning area, and forms data pairs with the location information of the Wi-Fi signals and the sampling points to build a fingerprint database. Since signal detection usually uses multiple devices, the centralized method is to send fingerprint data to the central server. The huge fingerprint database will bring large communication overhead and increase the training burden of the central server. The user's geographic location information will also be exposed during the fingerprint upload process, causing privacy data leakage. Federated learning uses a distributed training method, and the central server only has an aggregation function. This method can reduce the communication quota, reduce the training burden of the central server, and protect the user's privacy data.

[0003] Traditional indoor positioning methods use the data of the positioning area to train the positioning model of the positioning area. However, in the face of complex and changeable indoor positioning environments, traditional indoor positioning methods are not flexible enough. When there is a new positioning task, training a model from scratch will cause a lot of time overhead. Therefore, it is necessary to propose an indoor wireless positioning method that can quickly adapt to new scenarios, help train new models, and improve model performance. Summary of the invention

[0004] The purpose of the present invention is to provide an indoor wireless positioning method based on federated transfer learning, which can quickly adapt to new scenarios, enable new positioning models to converge quickly, and improve the performance of the model.

[0005] To achieve the above object, the present invention provides an indoor wireless positioning method based on federated transfer learning, comprising the following steps:

[0006] S1, collect RSS in the source domain and target domain at the edge node, and build the source domain and target domain fingerprint library;

[0007] S2: The central server initializes the global model and performs federated learning training. After multiple rounds of training until convergence, the source domain positioning model is obtained.

[0008] S3. Migrate the source domain positioning model to the target domain, adjust the model, and perform federated transfer learning training. After multiple rounds of training until convergence, the target domain positioning model is obtained.

[0009] S4. Input the signal characteristics of the device to be located into the positioning model, and the positioning model outputs the location information of the device.

[0010] Preferably, in step S1, the source domain and target domain fingerprint libraries are constructed, and the specific process is as follows:

[0011] S11, L Wi-Fi signal access points are distributed in the positioning area, the positioning area is divided into grids, the edge node signal detection device moves in the positioning area, and samples at the center of each grid; each edge node detection device performs multiple sampling at each sampling point, takes the median, and forms a data pair (x) with the positioning information of the sampling point. i ,y i );

[0012] The fingerprint is x i =[RSS i,1 ,RSS i,2 ,…,RSS i,L ], which is composed of L Wi-Fi signal strengths collected by edge node i at the sampling point, y i is the location coordinate information of the sampling point;

[0013] S12: Each edge node detection device collects data pairs of signal strength and location information to form a local fingerprint database Among them, the superscript indicates the number of samples;

[0014] S13, dividing the local fingerprint database into a source domain database and a target domain database;

[0015] S14. The edge node draws the convex hull of the location information group set of the local fingerprint database through the Melkman algorithm and calculates the convex hull area, and sends it to the central server.

[0016] Preferably, in step S2, the central server initializes the global model, performs federated learning training, and obtains the source domain positioning model after multiple rounds of training until convergence. The specific process is as follows:

[0017] S21. The central server initializes a multi-layer perceptron MLP model as the global model; selects the mean absolute error MAE as the loss function, selects the adaptive moment estimate Adam as the optimizer, sets the learning rate, and distributes the global model to the edge nodes;

[0018] S22, each edge node trains a local model using the local source domain fingerprint database, and obtains a local model after multiple local iterations;

[0019] S23, the edge node uploads the local model parameters to the central server, and the central server aggregates the models uploaded by the edge nodes by weighted convex hull area. The aggregation formula is as follows:

[0020]

[0021] Among them, ωt+1 is the updated global model, N represents the number of participating edge nodes, denote the convex hull areas of edge nodes i and j respectively, Indicates that edge node i downloads the global model ω from the central server t The local model is then trained with local data;

[0022] S24, the edge node downloads the global model aggregated by the central server, and after multiple model aggregations until the global model converges, a source domain positioning model is obtained.

[0023] Preferably, in step S3, the source domain positioning model is migrated to the target domain, the model is adjusted, and federated transfer learning training is performed. After multiple rounds of training until convergence, the target domain positioning model is obtained. The specific process is as follows:

[0024] S31, migrate the source domain localization model to the target domain, select stochastic gradient descent SGD as the optimizer, and select MAE as the loss function;

[0025] S32, the central server reinitializes an MLP model with a structure similar to the source domain model;

[0026] S33, copying the parameters of the source domain localization model to the new MLP model, freezing some layers in the hidden layer of the new MLP model so that they do not participate in the update;

[0027] S34, setting different learning rates for the hidden layer and regression layer of the new MLP model respectively, and setting a lower learning rate for the hidden layer to make the training smoother;

[0028] S35, distributing the global model to edge nodes, each edge node uses the local target domain fingerprint database to train a local model, and obtains a local model after multiple rounds of local iterations;

[0029] S36, the edge node uploads the local model parameters to the central server. The central server uses a weighted aggregation model based on the convex hull area. The edge node downloads the global model aggregated by the central server. After multiple model aggregations until the global model converges, the target domain positioning model is obtained.

[0030] Therefore, the present invention adopts the above-mentioned indoor wireless positioning method based on federated transfer learning, and the beneficial effects are as follows:

[0031] (1) The present invention allows the source domain localization model to be flexibly migrated to the target localization area to quickly adapt to new scenarios;

[0032] (2) The present invention can learn based on the knowledge learned by the source domain localization model, thereby improving the performance of the target domain localization model, and by adjusting the parameters of the localization model, the target domain localization model can converge quickly;

[0033] (3) The positioning model trained by the method proposed in the present invention has higher positioning accuracy than the positioning model trained by traditional federated learning.

[0034] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a system block diagram of an indoor wireless positioning method based on federated transfer learning of the present invention;

[0036] Figure 2 It is a flow chart of an indoor wireless positioning method based on federated transfer learning of the present invention;

[0037] Figure 3 This is a performance simulation diagram of the traditional federated learning indoor wireless positioning method;

[0038] Figure 4 This is a performance simulation diagram of an indoor wireless positioning method based on federated transfer learning of the present invention. DETAILED DESCRIPTION

[0039] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0040] like Figure 1 As shown, the indoor wireless positioning method based on federated transfer learning of the present invention combines federated learning and transfer learning. Federated learning, as a training method, undertakes the training tasks of the source domain positioning model and the target domain positioning model. Transfer learning fine-tunes the source domain positioning model so that the source domain positioning model can be more adapted to the target domain fingerprint database.

[0041] like Figure 2 As shown, the present invention provides an indoor wireless positioning method based on federated transfer learning, comprising the following steps:

[0042] S1, collect RSS in the source domain and target domain at the edge node, and build the source domain and target domain fingerprint library;

[0043] S2: The central server initializes the global model and performs federated learning training. After multiple rounds of training until convergence, the source domain positioning model is obtained.

[0044] S3. Migrate the source domain positioning model to the target domain, adjust the model, and perform federated transfer learning training. After multiple rounds of training until convergence, the target domain positioning model is obtained.

[0045] S4. Input the signal characteristics of the device to be located into the positioning model, and the positioning model outputs the location information of the device.

[0046] Example

[0047] S1. Collect RSS in the source domain and target domain at the edge node and build the source domain and target domain fingerprint library.

[0048] S11, L Wi-Fi signal access points are distributed in the positioning area, the positioning area is divided into grids, the edge node signal detection device moves in the positioning area, and samples at the center of each grid. Each edge node detection device performs multiple sampling at each sampling point, takes the median, and forms a data pair (x) with the positioning information of the sampling point. i ,y i ).

[0049] The fingerprint is x i =[RSS i,1 ,RSS i,2 ,…,RSS i,L ], which is composed of L Wi-Fi signal strengths collected by edge node i at the sampling point, y i It is the location coordinate information of the sampling point.

[0050] S12: Each edge node detection device collects data pairs of signal strength and location information to form a local fingerprint database The superscript indicates the number of samples.

[0051] S13. Divide the local fingerprint database into a source domain database and a target domain database.

[0052] S14. The edge node draws the convex hull of the location information group set of the local fingerprint database through the Melkman algorithm and calculates the convex hull area, and sends it to the central server.

[0053] S2. The central server initializes the global model and performs federated learning training. After multiple rounds of training until convergence, the source domain positioning model is obtained.

[0054] S21. The central server initializes a multilayer perceptron (MLP) model as the global model; selects the mean absolute error (MAE) as the loss function, selects the adaptive moment estimation (Adam) as the optimizer, and sets the learning rate; distributes the global model to the edge nodes.

[0055] S22. Each edge node trains a local model using the local source domain fingerprint database, and obtains a local model after multiple local iterations.

[0056] S23, the edge node uploads the local model parameters to the central server, and the central server aggregates the models uploaded by the edge nodes by weighted convex hull area. The aggregation formula is as follows:

[0057]

[0058] Among them, ω t+1 is the updated global model, N represents the number of participating edge nodes, denote the convex hull areas of edge nodes i and j respectively, Indicates that edge node i downloads the global model ω from the central server t The local model is then trained using local data.

[0059] S24, the edge node downloads the global model aggregated by the central server, and after multiple model aggregations until the global model converges, a source domain positioning model is obtained.

[0060] S3. Migrate the source domain positioning model to the target domain, adjust the model, and perform federated transfer learning training. After multiple rounds of training until convergence, the target domain positioning model is obtained.

[0061] S31. Migrate the source domain localization model to the target domain, select Stochastic Gradient Descent (SGD) as the optimizer, and select MAE as the loss function.

[0062] S32: The central server reinitializes an MLP model with a structure similar to the source domain model.

[0063] S33, copying the parameters of the source domain localization model to the new MLP model, and freezing some layers in the hidden layer of the new MLP model so that they do not participate in the update.

[0064] S34. Different learning rates are set for the hidden layer and regression layer of the new MLP model respectively. A lower learning rate is set for the hidden layer to make the training smoother.

[0065] S35. Distribute the global model to edge nodes. Each edge node uses the local target domain fingerprint database to train the local model. After multiple rounds of local iterations, a local model is obtained.

[0066] S36, the edge node uploads the local model parameters to the central server. The central server uses a weighted aggregation model based on the convex hull area. The edge node downloads the global model aggregated by the central server. After multiple model aggregations until the global model converges, the target domain positioning model is obtained.

[0067] S4. Input the signal characteristics of the device to be located into the positioning model, and the positioning model outputs the location information of the device.

[0068] In order to verify the authenticity of the present invention, an experimental simulation was conducted, and the experimental data was based on a real Wi-Fi fingerprint database. Figure 3 and Figure 4 As shown in the figure, both figures are trained on data from the same area. The comparison shows that the federated transfer learning indoor wireless positioning method can adapt well to new scenarios, and the performance of the migrated positioning model is very good. Under the same training rounds, the indoor wireless positioning model of federated transfer learning can achieve higher positioning accuracy than the indoor wireless positioning model of traditional federated learning.

[0069] Therefore, the present invention adopts the above-mentioned indoor wireless positioning method based on federated transfer learning, which can quickly adapt to new scenarios, enable the new positioning model to converge quickly, and improve the performance of the model.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. An indoor wireless positioning method based on federated transfer learning, characterized in that: The following steps are involved: S1. Collect RSS in the source domain and target domain at the edge node to build the source domain and target domain fingerprint library. The specific process is as follows: S11, L Wi-Fi signal access points are distributed in the positioning area, the positioning area is divided into grids, the edge node signal detection device moves in the positioning area, and samples are taken at the center of each grid; Each edge node detection device performs multiple sampling at each sampling point, takes the median, and forms a data pair (x i ,y i ); The fingerprint is x i =[RSS i,1 ,RSS i,2 ,…,RSS i,L ], which is composed of L Wi-Fi signal strengths collected by edge node i at the sampling point, y i is the location coordinate information of the sampling point; S12: Each edge node detection device collects data pairs of signal strength and location information to form a local fingerprint database Among them, the superscript indicates the number of samples; S13, dividing the local fingerprint database into a source domain database and a target domain database; S14, the edge node draws the convex hull of the location information group set of the local fingerprint database through the Melkman algorithm and calculates the convex hull area, and sends it to the central server; S2. The central server initializes the global model and conducts federated learning training. After multiple rounds of training until convergence, the source domain positioning model is obtained. The specific process is as follows: S21. The central server initializes a multi-layer perceptron MLP model as the global model; selects the mean absolute error MAE as the loss function, selects the adaptive moment estimate Adam as the optimizer, sets the learning rate, and distributes the global model to the edge nodes; S22, each edge node trains a local model using the local source domain fingerprint database, and obtains a local model after multiple local iterations; S23, the edge node uploads the local model parameters to the central server, and the central server aggregates the models uploaded by the edge nodes by weighted convex hull area. The aggregation formula is as follows: Among them, ω t+1 is the updated global model, N represents the number of participating edge nodes, denote the convex hull areas of edge nodes i and j respectively, Indicates that edge node i downloads the global model ω from the central server t The local model is then trained with local data; S24, the edge node downloads the global model aggregated by the central server, and after multiple model aggregations until the global model converges, a source domain positioning model is obtained; S3: Migrate the source domain positioning model to the target domain, adjust the model, and perform federated transfer learning training. After multiple rounds of training until convergence, the target domain positioning model is obtained. The specific process is as follows: S31, migrate the source domain localization model to the target domain, select stochastic gradient descent SGD as the optimizer, and select MAE as the loss function; S32, the central server reinitializes an MLP model with a structure similar to the source domain model; S33, copying the parameters of the source domain localization model to the new MLP model, freezing some layers in the hidden layer of the new MLP model so that they do not participate in the update; S34, setting different learning rates for the hidden layer and regression layer of the new MLP model respectively, and setting a lower learning rate for the hidden layer to make the training smoother; S35, distributing the global model to edge nodes, each edge node uses the local target domain fingerprint database to train a local model, and obtains a local model after multiple rounds of local iterations; S36, the edge node uploads the local model parameters to the central server, the central server weights the aggregation model based on the convex hull area, and the edge node downloads the global model aggregated by the central server. After multiple model aggregations until the global model converges, the target domain positioning model is obtained; S4. Input the signal characteristics of the device to be located into the positioning model, and the positioning model outputs the location information of the device.

Images