CN116861993A

CN116861993A - Distributed indoor positioning method and device based on federal learning and storage medium thereof

Info

Publication number: CN116861993A
Application number: CN202310703786.2A
Authority: CN
Inventors: 刘颖之; 魏凤生; 秦爽; 李晓倩; 冯钢
Original assignee: University of Electronic Science and Technology of China
Current assignee: University of Electronic Science and Technology of China
Priority date: 2023-06-14
Filing date: 2023-06-14
Publication date: 2023-10-10

Abstract

The invention relates to the technical field of indoor positioning, in particular to a distributed indoor positioning method and equipment based on federal learning and a storage medium thereof, comprising the steps of constructing an indoor positioning scene, and generating a training data set by a measurement sample based on DL PRS received in a time slot; based on the training data set, carrying out AI/ML deep learning through a Resnet model; the UE position in the indoor positioning scene is predicted through the advanced learning AI/ML based on the federal learning strategy, distributed multipoint co-positioning is adopted, and distributed training is carried out on a data set, so that a personalized deep neural network model can be provided for a user, and the overall performance of an algorithm is improved; the federal average algorithm plus element learning strategy is adopted, personalized federal learning is used, the aim is to find an initial sharing model, and the current or new user can adapt to the initial sharing model by performing gradient descent of one or more steps, so that a more personalized model is provided for distributed users.

Description

Distributed indoor positioning method and device based on federal learning and storage medium thereof

Technical Field

The invention belongs to the technical field of indoor positioning, and particularly relates to a distributed indoor positioning method and device based on federal learning and a storage medium thereof.

Background

The internet of things (IoT), machine communication (MTC), and other services layers based on the location of mobile devices are endless, and thus precise positioning technology has been attracting attention in recent years. Indoor positioning techniques include algorithms based on information such as wireless signals, visible light, sound, etc. Among these technologies, wireless-based approaches are most popular because wireless communication technologies are relatively mature and do not require the addition of additional components to the mobile device.

Indoor positioning plays a fundamental role in a wide range of internet of things (IoT) -based applications such as indoor emergency rescue, mall precision marketing, intelligent factory asset management and tracking, ambulatory medical services, virtual reality games, location-based social media, and the like. Despite the growing enormous market demand, in many cases it is not easy to provide a viable indoor positioning solution. Global satellite navigation systems (GNSS) have been a popular positioning technology with great success in positioning outdoor open scenes, and positioning accuracy can reach sub-meter level under the enhancement of various technologies. However, GNSS signals cannot be received well indoors because of the low power of GNSS, and continuous and reliable positioning cannot be provided. In many cases, especially in deeper indoor areas, the GNSS signals may be completely shielded.

Indoor positioning has been widely studied and various technologies based on WiFi, bluetooth, ultra Wideband (UWB), pseudolite, geomagnetism, sound/ultrasound, or Pedestrian Dead Reckoning (PDR) have been developed. Although different technologies have advantages, due to constraints of factors such as complex indoor space layout, complex topological structure, complex signal transmission environment and the like, realizing an accurate, effective, reliable and real-time positioning solution for indoor application still has high challenges.

Generally, the positioning methods can be classified into two major categories, namely, a geometry-based method and a feature matching-based method, and the geometry-based method can be classified into a triangulation method, a trilateration method and a joint estimation method, and the feature matching-based method is mainly called a fingerprint identification method.

The Chinese patent with publication number of CN116095600A discloses an indoor positioning method based on 5G space-time big data collaboration, wherein the 5G space-time big data is used as a benchmark, the data collaboration positioning is carried out by the characteristics of wide space coverage of a 5G network, and the like, but when machine learning is carried out to establish a positioning and position model, a large amount of marked training data exists, serious data privacy problems can be caused in the collecting process, and meanwhile, the method is limited by indoor environments and is easy to reflect, refract and scatter by various indoor objects, so that signals sent from one transmitter can reach a receiver from a plurality of different propagation paths, and the positioning precision is not ideal.

Disclosure of Invention

The invention aims at: aiming at the problems, the distributed indoor positioning method, the distributed indoor positioning equipment and the storage medium based on federal learning are provided.

In order to achieve the above purpose, the invention adopts the following technical scheme: a distributed indoor positioning method based on federal learning comprises the following steps:

constructing an indoor positioning scene, and generating a training data set by a measurement sample based on DL PRS received in one time slot;

based on the training data set, carrying out AI/ML deep learning through a Resnet model;

and predicting the position of the UE in the indoor positioning scene through the AI/ML after deep learning based on the federal learning strategy.

Further, the indoor positioning scene comprises a client and a server, and the client uploads the trained model to the server;

and the server averages the parameters and then issues model parameters to each client.

Further, the training data set includes CIR samples.

Further, estimating the CIR sample through two paths of LoS and NLoS;

in the NLoS path, the CIR samples from the transmitting antenna element s to the receiving antenna element u are

Wherein τ, τ _n ，τ _n,i Is a delay parameter;

delta (·) is a dirac delta function;

Is the NLoS channel coefficient of cluster N e {3,4, …, N } from transmit antenna s to receive antenna u time t;

is the NLoS channel coefficient of the cluster N epsilon {1,2} on the m-th ray from the transmitting antenna unit s to the receiving antenna unit u time t;

in the LoS path, the two terms are scaled according to the desired K-factor by adding the LoS channel factor to the NLoS channel impulse response

Wherein K is _R Is the rice K factor;

is the LoS channel coefficient at time t from transmit antenna element s to receive antenna element u.

Alternatively, the Resnet model extracts three-dimensional data features of the CIR sample, which are expressed as A X256X 2;

wherein A is CIR information of the number of base stations in a client received by one UE;

256 is the number of fast fourier transform samples of the impulse response waveform;

"2" is two data contents of real domain complex domain.

Further, adopting convolution layer and pooling layer operations in a Resnet model to reduce the dimension of the input three-dimensional data characteristics, and adding normalization operation and an activation function Relu after each convolution layer;

performing AI/ML training by a residual block after deformation, wherein a weight layer in the residual block is expressed by a convolution layer, performing dimension lifting operation after a second residual network and a fourth residual network, and under the condition that the dimension between the residual blocks is changed, performing dimension lifting on input cross-layer propagation in the residual block to directly add with output;

In the output layer operation, the average pooling operation is adopted to reduce the data dimension, the channel number is converted from 256 to 2 through linear layer conversion, and the model output is changed into the dimension identical to the UE coordinate, so that the predicted UE coordinate is obtained.

Further, the client receives global model parameters sent by the server, trains each client through the Resnet model, can acquire corresponding user data closest to each base station for training according to RSRP data sets generated in scene simulation, acquires the same number of data sets for training by each client, and feeds back the parameters of the Resnet model to the server after training;

the server distributes the global model parameters to the client, receives the parameters of the Resnet model fed back by the client, and updates the global model parameters.

Further, the client receives the global model parameter w from the server _t And let the client model parameters w _k ＝w _t And performs a random gradient descent of two local iterations:

b is local data divided in the training data set;

k is the kth client node;

then the front and back weight difference g _k ＝w _k -w ^t Transmitting to a server, wherein t is the t-th iteration;

the server receives the front-back weight difference value g sent by k client nodes ₁ ,g ₂ ,...,g _K And updating global model parameters by weighted averaging

Meanwhile, the invention also provides distributed indoor positioning equipment based on federal learning, which comprises:

the system comprises a processor and a memory, wherein the memory is used for storing a computer program, and the processor is used for calling and running the computer program stored in the memory and executing a distributed indoor positioning method based on federal learning.

The invention further provides a storage medium for distributed indoor positioning based on federal learning, which comprises the following components: for storing a computer program that causes a computer to perform a distributed indoor positioning method based on federal learning.

In summary, due to the adoption of the technical scheme, the beneficial effects of the invention at least comprise one of the following:

1. the invention adopts distributed multipoint co-location to perform distributed training on the data set, and can effectively prevent over fitting and/or under fitting, thereby improving the overall performance of the algorithm;

2. the invention adopts a policy of federal average plus element learning, uses personalized federal learning, aims at finding an initial sharing model, can be easily adapted to local data sets of current or new users by executing gradient descent of one or more steps on own data, can keep all benefits of a federal learning system structure, provides more personalized models for each user through a structure, and enables the federal framework to realize the same positioning performance as a large-scale data-driven learning model under the condition of protecting the privacy of the users;

3. In the present invention, each mobile user/intelligent agent is allowed to collect a smaller-scale local data set and approximate the global machine learning model in a collaborative manner. Compared with centralized training, the data collection mode can reduce excessive path loss and multipath fading caused by too far distance when data is transmitted to a data processing center, and reduce time delay in the transmission process. Meanwhile, as only parameter information is transmitted between each distributed node, the number of information bits transmitted mutually is far smaller than that of centralized training, so that the transmission speed is improved, and meanwhile, the memory burden at the data processing node is reduced;

4. according to the invention, the distributed nodes collect small-scale user data near the geographic position for training and positioning, so that the interference condition easily caused by long-distance transmission is avoided, the multipath effect is reduced, and the positioning precision is improved;

5. according to the invention, the distributed node model parameter aggregation is adopted, so that the model has better generalization performance, in the aspect of positioning accuracy, the data set with regional characteristics is used for distributed training, and a personalized solution can be provided for the coordinate prediction of each user, so that smaller positioning error is obtained, and better model generalization, positioning accuracy and convergence speed are obtained.

Drawings

FIG. 1 is a schematic view of an indoor positioning scenario;

FIG. 2 is a diagram of a channel coefficient generation process;

FIG. 3 is a Resnet model;

FIG. 4 is a residual block schematic;

FIG. 5 is a table of indoor positioning scene parameters;

FIG. 6 is a graph of a distributed algorithm positioning error;

FIG. 7 is a graph of a centralized algorithm positioning error;

FIG. 8 is a graph of alignment error versus centralized algorithm versus distributed algorithm;

FIG. 9 is a diagram of a distributed algorithm versus a centralized algorithm;

FIG. 10 is a graph of sustained low prediction error for a plurality of epochs;

FIG. 11 is a graph of positioning error for each epoch;

FIG. 12 is a flow chart of a distributed indoor positioning method based on federal learning.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments. The components of the embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.

Thus, the following detailed description of the embodiments of the invention, as presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, based on the embodiments of the invention, which are apparent to those of ordinary skill in the art without inventive faculty, are intended to be within the scope of the invention.

In addition, the embodiments of the present invention and the features of the embodiments may be combined with each other without collision.

The invention discloses a distributed indoor positioning method based on federal learning, which comprises the following steps:

The design aims at adopting distributed multipoint co-location to perform distributed training on a data set, and can effectively prevent over-fitting and/or under-fitting, thereby improving the overall performance of an algorithm; the adoption of the strategy of federal average plus element learning, the use of personalized federal learning, the goal is to find an initial sharing model, the current or new users can easily adapt to their local data sets by performing gradient descent of one or more steps to their own data, all benefits of the federal learning architecture can be preserved, more personalized models are provided for each user through the structure, and the federal framework makes it possible to realize the same positioning performance as a large-scale data-driven learning model under the condition of protecting the privacy of the users; each mobile user/intelligent agent is allowed to collect a smaller-scale local data set and approximate the global machine learning model in a collaborative manner. Compared with centralized training, the data collection mode can reduce excessive path loss and multipath fading caused by too far distance when data is transmitted to a data processing center, and reduce time delay in the transmission process. Meanwhile, as only parameter information is transmitted between each distributed node, the number of information bits transmitted mutually is far smaller than that of centralized training, so that the transmission speed is improved, and meanwhile, the memory burden at the data processing node is reduced; the distributed nodes collect small-scale user data near the geographic position for training and positioning, so that the interference condition easily caused by long-distance transmission is avoided, the multipath effect is reduced, and the positioning accuracy is improved; the distributed node model parameter aggregation is adopted, the model has better generalization performance by realizing multi-point cooperation, in the aspect of positioning accuracy, a data set distributed training with regional characteristics is used, a personalized solution can be provided for coordinate prediction of each user, so that smaller positioning errors are obtained, better model generalization performance, positioning accuracy and convergence speed are achieved, and a specific use flow can refer to fig. 12.

In specific implementation, the indoor positioning scene comprises a client and a server, and the client uploads the trained model to the server;

It should be noted that 18 base stations are typically selected to form the client.

In particular implementations, the training data set includes CIR samples and RSRP data sets generated from scene simulations.

It should be noted that, where CIR samples are used for depth model training, CIR channel impulse response is generated by generic channel model simulation.

Specifically, in a general channel model, taking a channel in a frequency domain as an example, the frequency domain space is divided into N samples with the size of Δf, and the total bandwidth w=n Δf occupied by the samples. Such a representation is typically present in an Orthogonal Frequency Division Multiplexing (OFDM) signal, but is not limited to use in such a signal.

Band N _Tx Transmitter (Tx) and band N of an antenna _Rx The model over the channel between the receivers (Rx) of the antennas, at frequency N e {0,..

Where L is the number of physical propagation paths (e.g., given by a ray tracker);

Alpha is the complex channel gain;

is the Rx array response as the angle of arrival (AoA) θεR ² Acquiring functions of azimuth and elevation;

is taken as the deviation angle +>As a function of acquiring azimuth and elevation;

τ represents the time of arrival (ToA);

v denotes the Doppler shift;

T _s representing the duration of a sequence;

AoA is defined in the reference standard of Rx and AoD is defined in the reference standard of Tx, so these angles depend on the respective direction. Below 6ghz explicit geometry information in the 1-channel is difficult to use due to limited delay and angular resolution, plus weak connection of the path to the ambient geometry. In contrast, at mmWave and above, the path is more closely related to the environment geometry, and can be more easily resolved. Therefore, we assume that each path in equation 1 corresponds to a physical object.

The signal observations at Rx can be expressed as:

wherein, the liquid crystal display device comprises a liquid crystal display device,is an orthogonal analog Rx synthesizer, satisfying +.>Using M _Rx ≤N _Rx A radio frequency link f _n,k For the kth Tx signal passing through the Tx array, satisfy +.>Is the combined noise. Wherein P is _tx For average transmit power, N ₀ Is the noise power spectral density. Transmitting signal f _n,k Is generally known (pilot in positioning or bi-static sensing or known data in single static sensing), but may be partially unknown in semi-blind estimation.

As shown in the scenario in fig. 1, in the indoor scenario positioning problem, each UE has an unknown state s, which can be inferred from the observation formula 2, and the state information includes the position x e R ³ Azimuth angle o e R ³ The time delay τ e R, the power p e R, etc., and is typically represented using parameters.

Basic communication node (BSs i e {1, …, N) _B }) has some known state information, namely position x ⁱ ∈R ³ And direction o ⁱ ∈R ³ And is time synchronized. The positioning in the indoor scenario is a user-centric Downlink (DL) positioning. In DL, each BS i transmits signals through orthogonal subcarriers, generating a channelUE observations of upper Rx->

Equation 1 is widely used in communication systems, but channels are generally used when positioning problems are consideredDividing into a line-of-sight path and a non-line-of-sight path, i.e.

And in the acknowledgement of NLoS channel coefficients:

as shown in fig. 2, the state information and geometry information at the UE may be obtained from parameters in the channel coefficients, for N-2 weakest clusters, such as n=3, 4. The channel coefficients can be written as:

wherein F is _rx,u,θ ，F _rx,u,φ Is the field pattern of the receiving antenna, F in the direction of the spherical basis vector _rx,s,θ ，F _rx,s,φ Respectively the field patterns of the transmitting antenna elements s in the direction of the spherical basis vector. It is noted that the diagram is given in the Global Coordinate System (GCS), so that depending on the transformation in the setup with respect to the antenna direction, Is provided with an azimuth angle of arrival phi _n、m、AOA And elevation angle of arrival θ _n、m、ZOA Is given by equation 5

Where n represents a cluster, m represents a ray n within the cluster,is the azimuth angle phi _n、m、AOD And elevation angle of departure angle theta _n、m、ZOD Is given by equation 6

Where n represents a cluster and m represents a ray within the cluster n.Is the position vector of the receiving antenna unit u,

is the position vector of the transmitting antenna element s, κ _m,n Is the cross polarization power ratio on a linear scale, lambda ₀ Is the wavelength of the carrier frequency. If polarization is not considered, the 2x2 polarization matrix can be scalar exp (j phi) _n,m ) Instead, and only vertical polarization field patterns are applied.

Meanwhile, the path loss expression of NLoS in InF-DH scene is PL=18.6+35.7log ₁₀ (d _3D )+20log ₁₀ (f _c ) Shadow fading is sigma _SF =7.2, by applying path loss and channel fading to the channel parameters, a channel model in indoor scenarios can be obtained

And in the validation of LoS channel coefficients:

the general expression of the LoS channel model can be written as

Wherein channel complex power gainCan be written as a general mathematical expression of (c)

Where lambda is the wavelength at the carrier wave,and->Representing the antenna element response at Tx and Rx, respectively, the power gain is determined by the path loss.

The mathematical expression is explained in more detail when applied to a real scene. In an indoor positioning scenario, i.e., an InF-DH scenario, channel parameters in the channel model are expressed as:

path loss and shadow fading are denoted as PL, respectively _LOS ＝31.84+21.50log ₁₀ (d _3D )+19.00log ₁₀ (f _c )，σ _SF =4.3, the LoS path channel model can be obtained by applying path loss and shadow fading to the channel coefficients

After the general channel model is established, the channel impulse response in the positioning problem needs to be calculated for two paths of LoS and NLoS respectively:

in the NLoS path, it is assumed that there are N clusters, each cluster having 20 rays. For the two strongest clusters, e.g., n=1 and 2, the ray is propagated delay to three sub-clusters with a fixed delay offset. The number of rays and the ray power of the three sub-clusters are different.

For the sub-cluster i ε {1,2,3}, mapping to rays uses R _i And (3) representing. For N-2 weakest clusters, let n=3, 4 …, N, the power of the different rays in each cluster is equal. In the case of NLoS, the CIR from the transmitting antenna element s to the receiving antenna element u is

Wherein τ, τ _n ，τ _n,i Is the delay parameter, delta (·) is the dirac delta function, i.e. the unit impulse response,is the NLoS channel coefficient of cluster N e {3,4, …, N } from transmit antenna s to receive antenna u time t. / >Is the NLoS channel coefficient at time t from the transmitting antenna element s to the receiving antenna element u on the m-th ray of cluster N e {1,2 }.

Wherein K is _R Is the rice K factor, which is the value of the rice,is the LoS channel coefficient at time t from transmit antenna element s to receive antenna element u.

After the above procedure is completed, measurement samples are generated based on DL PRS received in one slot. For each reception of DLPRS, the UE will make measurements corresponding to each measurement, such as DL-RSTD, CIR, and RSRP. The tag associated with each sample is known location information of the target UE. To evaluate each AI/ML based scheme, 80000 samples were generated with associated tags. The evaluation protocol for each protocol is as follows:

because of the requirement of Release 16 on indoor user Release, drop (i.e. user Release times) and user Release distribution need to be considered in the generation of the data set, the condition (1 drop) that 80000 users are released indoors at one time in the use of the data set is considered, and the distribution state is random distribution:

when indoor positioning is performed, a user in a scene is represented as u= { U _i } _{i＝1，...，N} For user u at kth drop _k To the point that it objectively has a real global coordinate relative to the indoor sceneWherein the method comprises the steps ofRespectively representing the east coordinate and the north coordinate in the horizontal space under the indoor scene, and P ^k Only when u _k E U is valid. Assuming that the predicted coordinates of each user through the depth model are valid, the problem can be translated into how to use the environmental information, convolved neural system { P ] ^k |u _k E U } to approximate the predicted global coordinates +.>

In AI/ML deep learning through a Resnet model based on CIR samples, since the structure of CIR input is similar to that of image input, the Resnet model is adopted as an AI/ML positioning scheme, and is often used in the field of image processing and is used for extracting the characteristics of CIR three-dimensional data. For each CIR sample we use complex values in the time domain to represent its complete feature, with its input dimension "18 x 256 x 2". The three-dimensional input is generated by 18 base station CIR information received by one UE and 256 FFT sampling points, 2 representing two data contents of a real domain complex domain. The structure of Resnet is shown in FIG. 4. In the first four reshape layers, the CIR input is converted to a size of 18 x 64, then convolved with 12 Con2D layers, and a convolution kernel of 3 x 3 size, with the shortcut operation performed at the dashed line, i.e., the data is reduced in dimension with a step size equal to 2.

AI/ML deep learning by Resnet model is based mainly on residual network and convolutional neural network, wherein convolutional neural network is defined as the convolution between two functions in mathematics

(f.times.g) (x) = ≡f (z) g (x-z) dz (formula 14)

That is, convolution is the calculation of the overlap between f and g when a function is "flipped" and shifted by x. In the case of discrete objects, the integral becomes a summation. For example, for vectors extracted from an infinite set of dimensional vectors that are square and directed as Z.

(f*g)(i)＝∑ _a f (a) g (i-a) (equation 15)

For a two-dimensional tensor, then the corresponding sum over the indices (a, b) of f and the indices (i-a, j-b) of g:

for convolutional layers in a neural network, the operation it expresses is actually a cross-correlation operation, not a convolutional operation. In the convolution layer, the input tensor and the kernel tensor generate an output tensor through a cross-correlation operation. The convolution layer performs a cross-correlation operation on the input and the convolution kernel weights and produces an output after adding a scalar offset. The two trained parameters in the convolutional layer are the convolutional kernel weights and scalar offsets. Just as the fully connected layer is initialized, the convolution kernel weights are also randomly initialized when training the convolution-layer based model.

Whereas in the residual network the original input is x as shown in fig. 4, and the ideal mapping that it is desired to train is f (x) (as input to the upper activation function in the figure). The part of the left half of the dashed box in fig. 4 needs to be fitted directly to the mapping f (x), while the part of the right half of the dashed box in fig. 4 needs to be fitted to the residual mapping f (x) -x. Residual mapping tends to optimize performance more easily in reality. The right graph is the infrastructure-residual block (residual block) of ResNet. In the residual block, the input may propagate more quickly forward through the cross-layer data channel.

In extracting the characteristic of the CIR three-dimensional data, the dimension may be expressed as 2xKx256, where K is the number of base stations, 256 is the number of fast fourier transform sampling points of the impulse response waveform, and 2 is real part data and imaginary part data, that is, single CIR environment information indicates real part and imaginary part information of channel impulse responses received by a single UE in an indoor environment from 18 base stations. The Resnet book is used to process multi-channel picture information, and in this design, the input data dimension can be regarded as Kx 256-sized 2-channel picture data.

In the first four-layer network part, the convolution layer and pooling layer operations are used to dimension the input information into 64-channel data with the size of KxK, so that the input information can be processed by convolution kernels with the size of 3×3 when the input information subsequently enters the residual block. Meanwhile, a normalization operation is added after each convolution layer, that is, samples input per batch are normalized using a batch norm2d operation, and an activation function Relu, that is, σ (·) =relu (x) =max (x, 0), is added after the normalization operation to increase the linear relationship between neural network layers.

After the deformation operation, the network is trained by 6 residual blocks, the weight layers in the residual blocks are expressed by convolution layers, and the number of channels and the convolution kernel size of the convolution operation are marked in detail in fig. 4. The second and fourth residual networks are followed by an up-scaling operation, i.e. doubling the number of channels by setting the convolution step size to 2. In the case of dimensional changes, the input cross-layer propagation in the residual block requires a downsampling operation to be upscaled to be directly added to the output, this operation being marked with a dashed line in the design drawing.

In the output layer operation, the data dimension is reduced by adopting an average pooling operation, the output data dimension is converted into 1 x 1,

finally, the channel number is converted from 256 to 2 through linear layer conversion, so that the model output is changed into the same dimension as the UE coordinate, and the predicted UE coordinate is obtained

When the federal learning strategy is performed, the transverse federal learning model of the rice is used for distributing training tasks to the local terminal and the central node. The local terminal is mainly responsible for updating local model parameters, calculating loss values and gradient of model weights. The central node gathers the locally uploaded gradient information and uses an aggregation algorithm to fuse all gradients to update the global model. The central node then passes the new model weights to the local terminal. These processes will be repeated until the model converges.

In the initial stage of training, the weights of the global model are assigned to random values. The global model then transmits the weights to the local model, which updates the local model.

Wherein during the local training, each local terminal inputs local private data into the local model to predict error and calculates an error value from a Mean Square Error (MSE) loss function as described by the following equation:

θ＝{W，b}

the weight gradient of the last convolutional layer is then calculated as follows, where u, v is denoted as the input and output of the convolutional layer:

wherein +. ^l J can be expressed as

Thus generalizing to all convolution layer cases, the gradient of each parameter can be expressed as

Finally, the gradients of all parameters can be expressed as Representing common->In the transverse federation, the parameter input of the local model is the average value of the parameter gradients of each local virtual link issued by the global model +.>

In the global process, the global model collects gradients from each local model for model weight update. For each iteration of the global model, the local model provides updated gradients, i.e., gradient sets Each parameter gradient is of length +.>Is a list of (3).

The global model then uses an aggregation algorithm to integrate all gradient information to update itself. After the global model is trained in the S-wheel mode, parameters are sent to the local model. The local model parameters are updated and the next iteration begins.

In performing federal learning, the objective function can be written generally as

The goal of federal meta-learning is to exploit the basic ideas behind the meta-learning (MAML) framework to design personalized variants of the FL problem. The basic concept of federal element learning can be set forth as follows: in MAML, it is assumed that there is a limited computational budget to update our model after a new task arrives, in which an initialization state is found that makes it possible for the model to perform well after updating the new task through one or several gradient descent steps. The meta-learned objective function can be written as:

where α.gtoreq.0 is a step size, the advantage of this formula is that it not only maintains the advantages of FL, but it also allows for differences between users, since the user can take this objective function as an initial point and update it slightly with its own data relative to the objective function. This means that the user can obtain and update the initialization result by checking their own data, performing one or several gradient descent steps to obtain a model that fits their own data set.

The client receives global model parameters w from the server _t And let w _k ＝w _t And performs a random gradient descent of the E local iterations:

b is local data divided in the training data set;

k is the kth client node;

In performing the simulation test, the positioning evaluation is performed by:

in indoor positioning scene, P _pre Is a predicted inaccurate coordinate value, which contains errors. Training a deep learning model F to predict coordinate errors in an areaThe coordinate error is the true coordinate P ^k And predicted coordinates P _pre I.e. the optimization objective is to minimize the MSE value between the two,/i>

For the evaluation index of accuracy, the MSE error distribution function (CDF) of 90%, 80%, 67%, 50% percentile is used to measure the accuracy level of the algorithm.

Indoor positioning scene parameter setting is shown in fig. 5, and for a model using a single-frequency data set, the performance effect is as follows, and positioning errors of different positioning schemes when the positioning accuracy is 50%, 67%, 80% and 90% percentile are recorded respectively, as shown in table one:

Method	50％	67％	80％	90％
					Conventional method	11.89	13.62	14.78	16.36
AI+CIR	0.27	0.36	0.43	0.54

(Table I)

The AI/ML based approach can significantly improve positioning accuracy. When the traditional positioning algorithm is adopted for the coordinate prediction of the UE, the positioning precision exceeds 10m under the CDF percentile of 90 percent. AI/ml based methods can reduce positioning errors to within 1 meter.

In the multi-point co-location federation learning simulation, under the condition that a single-frequency data set is used as input, the parameter setting of the personalized federation learning algorithm is shown in a table II:

total number of clients	18
		Global iteration number(epoch)	20
Local iteration number (epoch)	30
		The number of clients participating in training is selected for each round	10
Number of samples per round in local training (batch size)	32
		Learning rate of local training	0.0001
Parameters of SGD optimizer (momentum)	0.0001

(Meter two)

The distributed algorithm performance, 10 clients are selected in each round of global training, and in 50 epochs, the positioning error corresponding to the achieved 90% accuracy is 0.205m as shown in fig. 6. At 44 epochs, while the positioning error stabilizes below 0.3m after 25 epochs, it is considered to converge after 25 epochs.

The performance of the centralized algorithm is shown in fig. 7, the centralized algorithm converges after 600 iterations of positioning performance according to a positioning error curve, the optimal positioning error is 0.54m, and compared with the convergence speed, the convergence speed of the distributed algorithm is obviously faster than that of the centralized algorithm.

As can be seen from comparison of the performance of the distributed algorithm with that of the centralized algorithm in 50 epochs, the distributed algorithm is superior to the centralized algorithm in terms of convergence, positioning accuracy and positioning performance stability, as shown in FIG. 8.

For the distributed algorithm and the centralized algorithm, the two optimal positioning error distributions are compared, as shown in fig. 9, the results are shown in table three,

Method	50％	67％	80％	90％
					conventional method	11.89	13.62	14.78	16.36
CIR input+centralized algorithm	0.27	0.36	0.43	0.54
					CIR input+distributed algorithm	0.13	0.16	0.18	0.21

(Table III)

As can be seen from the cdf curve, for most users, the positioning error of the distributed algorithm is smaller than that of the centralized algorithm, and it can be considered that the positioning effect is significantly improved in accuracy by the distributed algorithm, and it is pointed out that the conventional method is the conventional positioning method used in the background technology.

For global training, as shown in fig. 10, different numbers of clients are selected for each round of global training, positioning errors are observed, and the numbers of clients participating in training are respectively set as follows: 10. 12, 14, 16, 18. By observing the difference of the positioning errors under the condition of different client numbers, the condition that the low prediction errors are continuously achieved under a plurality of epochs and the prediction errors are stable can be used as a standard for judging the positioning accuracy. It can be seen that, when the number of clients is 16, the positioning error obtained by the predicted coordinates is continuously minimum, and when the number of clients is 10, the positioning error obtained by the predicted coordinates is continuously maximum. When the number of the clients is 12, 14 and 18, the obtained positioning errors are not different, and the performance is in the middle position. Therefore, the number of the clients and the positioning performance are judged not to be in a linear relation, the performance reaches a peak value at a certain value, and the positioning accuracy is best when the number of the clients is 16 under the condition of the experiment.

As shown in FIG. 11, since the positioning errors under each epoch are different, we use the positioning errors stably reaching below 0.3m as the standard for judging convergence, choose different numbers of clients to train, observe the iteration times when the clients reach the convergence condition, observe that the larger the number of clients is, the fewer the iteration times when the clients reach the convergence condition is, and consider that in the simulation, the larger the number of clients participating in the training is, and the faster the convergence speed is.

There is also provided in this embodiment a federally learning-based distributed indoor positioning apparatus, comprising:

The processor may be a central processing unit (Central Processing Unit, CPU), but also other general purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), field programmable gate arrays (FieldProgrammable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or the like. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

The memory may be an internal memory unit or an external memory device, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital Card (SD), a Flash Card (Flash Card), etc. Further, the memory may also include both internal storage units and external storage devices. The memory is used for storing the computer program and other programs and data, and may also be used for temporarily storing data that has been or is to be output.

It will be apparent to those skilled in the art that, for convenience and brevity of description, only the above-described division of the functional units and modules is illustrated, and in practical application, the above-described functional distribution may be performed by different functional units and modules according to needs, i.e. the internal structure of the apparatus is divided into different functional units or modules to perform all or part of the above-described functions. The functional units and modules in the embodiment may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit, where the integrated units may be implemented in a form of hardware or a form of a software functional unit. In addition, the specific names of the functional units and modules are only for distinguishing from each other, and are not used for limiting the protection scope of the present application. The specific working process of the units and modules in the above system may refer to the corresponding process in the foregoing method embodiment, which is not described herein again.

In the foregoing embodiments, the descriptions of the embodiments are emphasized, and in part, not described or illustrated in any particular embodiment, reference is made to the related descriptions of other embodiments.

Those of ordinary skill in the art will appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, or combinations of computer software and electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

In the embodiments provided in the present invention, it should be understood that the disclosed apparatus/terminal device and method may be implemented in other manners. For example, the apparatus/terminal device embodiments described above are merely illustrative, e.g., the division of the modules or units is merely a logical function division, and there may be additional divisions in actual implementation, e.g., multiple units or components may be combined or integrated into another system, or some features may be omitted or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection via interfaces, devices or units, which may be in electrical, mechanical or other forms.

The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in the embodiments of the present invention may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit. The integrated units may be implemented in hardware or in software functional units.

The integrated modules/units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable storage medium. Based on such understanding, the present invention may implement all or part of the flow of the method of the above embodiment, or may be implemented by a computer program to instruct related hardware, where the computer program may be stored in a computer readable storage medium, and when the computer program is executed by a processor, the computer program may implement the steps of each of the method embodiments described above. Wherein the computer program comprises computer program code which may be in the form of source code, object code, executable files or some intermediate form or the like. The computer readable medium may include any entity or device capable of carrying the computer program code, a recording medium, a U disk, a removable hard disk, a magnetic disk, an optical disk, a computer Memory, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), an electrical carrier signal, a telecommunications signal, a software distribution medium, and so forth.

In this embodiment, a computer storage medium is further provided, where a computer program is stored, where the computer storage medium may be one of a magnetic random access memory, a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory, a flash memory, a magnetic surface memory, and an optical disc, and may also be various devices including one or any combination of the foregoing, such as a mobile phone, a computer, a tablet device, etc., where the computer program can drive a system for resolving log data conflicts in different formats, and where the computer program processor can perform a distributed indoor positioning method based on federal learning.

Finally, it should be noted that: the foregoing description of the preferred embodiments of the present invention is not intended to be limiting, but rather, it will be apparent to those skilled in the art that the foregoing description of the preferred embodiments of the present invention can be modified or equivalents can be substituted for some of the features thereof, and any modification, equivalent substitution, improvement or the like that is within the spirit and principles of the present invention should be included in the scope of the present invention.

Claims

1. The distributed indoor positioning method based on federal learning is characterized by comprising the following steps of:

2. The federally learned distributed indoor positioning method according to claim 1, wherein: the indoor positioning scene comprises a client and a server, and the client uploads the trained model to the server;

3. The federally learned distributed indoor positioning method according to claim 2, wherein: the training data set includes CIR samples.

4. A distributed indoor positioning method based on federal learning according to claim 3, wherein: estimating the CIR sample through two paths of LoS and NLoS;

Wherein τ, τ _n ，τ _n, ⁱ Is a delay parameter;

delta (·) is a dirac delta function;

Wherein K is _R Is the rice K factor;

5. A distributed indoor positioning method based on federal learning according to claim 3, wherein: the Resnet model extracts three-dimensional data features of CIR samples, which are expressed as A×256×2;

"2" is two data contents of real domain complex domain.

6. The federally learned distributed indoor positioning method according to claim 5, wherein: adopting convolution layer and pooling layer operations in a Resnet model to reduce the dimension of the input three-dimensional data characteristics, and adding normalization operation and an activation function Relu after each convolution layer;

7. The federally learned distributed indoor positioning method according to claim 6, wherein: the client receives global model parameters sent by the server, trains each client through a Resnet model, can acquire corresponding user data closest to each base station for training according to RSRP data sets generated in scene simulation, acquires the same number of data sets for training by each client, and feeds back the parameters of the Resnet model after training to the server;

8. The federally learned distributed indoor positioning method according to claim 7, wherein: the client receives global model parameters w from the server _t And let the client model parameters w _k ＝w _t And performs a random gradient descent of two local iterations:

b is local data divided in the training data set;

k is the kth client node;

9. A federally learned distributed indoor positioning apparatus, comprising:

a processor and a memory for storing a computer program, the processor being adapted to invoke and run the computer program stored in the memory, to perform the method according to any of claims 1 to 8.

10. A storage medium for distributed indoor positioning based on federal learning, characterized in that: for storing a computer program that causes a computer to perform the method of any one of claims 1 to 8.