CN114897063A - Indoor positioning method based on-line pseudo label semi-supervised learning and personalized federal learning - Google Patents

Indoor positioning method based on-line pseudo label semi-supervised learning and personalized federal learning Download PDF

Info

Publication number
CN114897063A
CN114897063A CN202210467468.6A CN202210467468A CN114897063A CN 114897063 A CN114897063 A CN 114897063A CN 202210467468 A CN202210467468 A CN 202210467468A CN 114897063 A CN114897063 A CN 114897063A
Authority
CN
China
Prior art keywords
model
local
user
data set
initial
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210467468.6A
Other languages
Chinese (zh)
Inventor
伍哲舜
吴晓萍
龙云亮
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Sun Yat Sen University
Original Assignee
Sun Yat Sen University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Sun Yat Sen University filed Critical Sun Yat Sen University
Priority to CN202210467468.6A priority Critical patent/CN114897063A/en
Publication of CN114897063A publication Critical patent/CN114897063A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
    • H04W4/02—Services making use of location information
    • H04W4/021—Services related to particular areas, e.g. point of interest [POI] services, venue services or geofences
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
    • H04W4/02—Services making use of location information
    • H04W4/023—Services making use of location information using mutual or relative location information between multiple location based services [LBS] targets or of distance thresholds
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
    • H04W4/30—Services specially adapted for particular environments, situations or purposes
    • H04W4/33—Services specially adapted for particular environments, situations or purposes for indoor environments, e.g. buildings
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W64/00—Locating users or terminals or network equipment for network management purposes, e.g. mobility management
    • H04W64/006—Locating users or terminals or network equipment for network management purposes, e.g. mobility management with additional information processing, e.g. for direction or speed determination

Abstract

The invention relates to the technical field of indoor positioning, and discloses an indoor positioning method based on-line pseudo label semi-supervised learning and personalized federal learning, which comprises the following steps: s1, a user constructs a local data set, and a server constructs a cloud data set; s2, the server is provided with machine learning models, the server respectively issues the machine learning models to each user, the machine learning model located at the user side is called a local model, and the machine learning model located at the user side is called a local model; s3, training a local model by a user through data with labels in a local data set to obtain an initial local model; s4, the server trains a global model through data with labels in the cloud data set to obtain an initial global model; s5, obtaining a trained local model and an updated global model through federal learning; and S6, the user carries out personalized positioning through the mixed expert model according to the trained local model and the updated global model. The invention solves the problem that the prior art ignores the difference of the height dynamic state and the positioning requirement of the local data and can not carry out personalized positioning, and has the characteristics of high efficiency and high precision.

Description

Indoor positioning method based on-line pseudo label semi-supervised learning and personalized federal learning
Technical Field
The invention relates to the technical field of federal learning, in particular to an indoor positioning method based on-line pseudo label semi-supervised learning and personalized federal learning.
Background
With the continuous development of the information communication technology field, the communication perception integrated technology is widely concerned as a representative application and scheme of 6G, and the information processing flow of the service shows a trend that wireless communication and wireless perception tend to overlap. Based on ubiquitous wireless communication signals in the environment, wireless sensing applications such as wireless fingerprint positioning and wireless human behavior identification are developed vigorously. Meanwhile, the method is beneficial to the development of an artificial intelligence algorithm represented by deep learning, and the wireless indoor fingerprint positioning technology which can be effectively modeled as a supervised learning task can train a deep learning model by utilizing the fingerprint matching characteristic between the characteristics and the position coordinates of a wireless signal so as to efficiently and accurately complete the positioning task.
However, the wireless indoor positioning technology enabled by artificial intelligence faces the bottleneck of further development. For artificial intelligence technologies represented by deep learning, training a high-performance model with high recognition accuracy requires high-density computing resources and massive labeled data. Thus, current deep learning based wireless location technologies typically need to be deployed in cloud computing data centers with powerful computing capabilities and rely on service providing platforms to hire workers to acquire and annotate wireless signal data offline. With the high popularity and development of the current mobile terminal devices, the organic combination of the edge intelligence technology and the wireless sensing technology has attracted high attention in academia and industry. The edge intelligent technology represented by the federal learning provides a technology which allows a deep learning model to be deployed on a terminal device, and further can sink tasks of calculating and perceiving signals to an edge side and an end side together. The technology of 'last kilometer' for solving artificial intelligence landing can effectively promote the further development of the wireless positioning technology.
In order to solve the problem, a federal learning-based building floor indoor positioning method is provided, wherein a radio frequency fingerprint positioning model is constructed by using a distributed deep learning technology which is participated by an edge server and a plurality of mobile clients together. The server firstly initializes the model and carries out centralized pre-training by using a small amount of fingerprint data, each client uses local fingerprint data to carry out further model training, then the trained local model is transmitted to the server, and the server aggregates the local models collected from each client to obtain a global model for radio frequency fingerprint positioning
However, the prior art has the problem that personalized positioning cannot be performed by neglecting the difference between the local data height dynamic and the positioning requirement, so how to invent an indoor positioning scheme which can cover the difference between the local data height dynamic and the positioning requirement and perform personalized positioning is a problem to be solved in the technical field.
Disclosure of Invention
The invention provides an indoor positioning method based on-line pseudo label semi-supervised learning and personalized federal learning, which aims to solve the problem that personalized positioning cannot be carried out due to neglect of difference of local data height dynamics and positioning requirements in the prior art, and has the characteristics of high efficiency and high precision.
In order to achieve the purpose of the invention, the technical scheme is as follows:
an indoor positioning method based on-line pseudo label semi-supervised learning and personalized federal learning comprises the following steps:
s1, a user constructs a local data set, and a server constructs a cloud data set;
s2, the server is provided with machine learning models, the server respectively issues the machine learning models to each user, the machine learning model located at the user side is called a local model, and the machine learning model located at the user side is called a local model;
s3, training a local model by a user through data with labels in a local data set to obtain an initial local model;
s4, training a global model by the server through data with labels in the cloud data set to obtain an initial global model;
s5, obtaining a trained local model and an updated global model through federal learning;
and S6, the user carries out personalized positioning through the mixed expert model according to the trained local model and the updated global model.
Optimally, a user constructs a local data set, specifically: supposing that K users are provided, K wireless routers are arranged in an area to be detected, the wireless routers transmit and receive signals, received signal strength obtained by the wireless routers and position labels obtained by the inertial navigation technology are used as input, and K received signal strength vectors xi ═ r are constructed i1 ,r i2 ,…,r ik ]And simultaneously recording a position coordinate vector yi, wherein components xi and yi are respectively an abscissa and an ordinate of the indoor plan, integrating the signal intensity vector and the position coordinate vector into a local data set, and calling the received signal intensity as RSS.
Further, the machine learning model is a multilayer perceptron model with fixed hyper-parameters, and the multilayer perceptron model with fixed hyper-parameters is called an MLP model.
Furthermore, the trained local model and the updated global model are obtained through federal learning, and the method specifically comprises the following steps:
s401, regarding the local data set as label-free, performing on-line pseudo label semi-supervised learning on the initial local model and the initial global model to obtain a pseudo label, and labeling the local data set by the pseudo label;
s402, further training an initial local model by combining the local data set marked by the pseudo label through a knowledge distillation technology to obtain a trained local model;
and S403, the server performs federated learning on the trained local model to obtain an average value of the local model weight, and updates the global model according to the average value of the local model weight.
Furthermore, regarding the local data set as label-free, obtaining a pseudo label by performing online pseudo label semi-supervised learning on the initial local model and the initial global model, and labeling the local data set by the pseudo label, specifically comprising the following steps:
K1. labeling RSS data in the local data set through the initial global model and the initial local model to obtain a pseudo label estimated by the initial global model and a pseudo label estimated by the initial local model:
Figure BDA0003624950180000031
Figure BDA0003624950180000032
wherein, the first and the second end of the pipe are connected with each other,
Figure BDA0003624950180000033
the pseudo-label estimated for the initial global model,
Figure BDA0003624950180000034
a pseudo label estimated for the initial local model; m is an MLP model, and M is an MLP model,
Figure BDA0003624950180000035
characteristics of RSS data for user k in the t-th round,
Figure BDA0003624950180000036
Weight parameter, ω, for global model of the t-th round user k k A local model weight for the kth user;
K2. by mixing
Figure BDA0003624950180000037
And
Figure BDA0003624950180000038
weighted sum to obtain final pseudoLabel (R)
Figure BDA0003624950180000039
Figure BDA00036249501800000310
Wherein
Figure BDA00036249501800000311
Parameters for balancing two pseudo tag estimates;
K3. adaptive adjustment by the following formula
Figure BDA00036249501800000312
Figure BDA00036249501800000313
Wherein the sim function is the reciprocal of the euclidean distance;
K4. if it is
Figure BDA00036249501800000314
And if the local data set is larger than the set threshold value, labeling the local data set by using the pseudo label of the global model, otherwise labeling the local data by using the pseudo label of the local model.
Furthermore, the initial local model is further trained by combining the local data set labeled with the pseudo label through a knowledge distillation technology to obtain a trained local model, and the specific steps are as follows:
A1. taking the initial global model as a teacher network and the initial local model as a student network, and modifying a loss function used for training the local model;
Figure BDA0003624950180000041
wherein the content of the first and second substances,
Figure BDA0003624950180000042
a federal knowledge distillation loss function used for training a local model, wherein beta is a knowledge distillation factor, y is a position label, Mk (x) is position prediction output by the local model, and MG (x) is position prediction output by the global model;
A2. further modifying the function used for training the local model to obtain the modified function used for training the local model:
Figure BDA0003624950180000043
wherein the content of the first and second substances,
Figure BDA0003624950180000044
loss of regularization term;
A3. and training the local model by combining the obtained modified objective function with the labeled local data set.
Further, the objective function is further modified, specifically: by the formula
Figure BDA0003624950180000045
To obtain
Figure BDA0003624950180000046
And will be
Figure BDA0003624950180000047
M in k (x)-M G (x) Modified to
Figure BDA0003624950180000048
Obtaining a modified objective function:
Figure BDA0003624950180000049
furthermore, the server performs federated learning on the trained local model to obtain a trained local model weight, and updates the global model according to the trained local model weight, and the specific steps are as follows:
B1. obtaining a weight parameter omega capable of minimizing the global model loss through a loss formula of the global model, wherein the loss formula of the global model is as follows:
Figure BDA0003624950180000051
wherein F () is a global model loss function, N is a total number of users, F k () Is a local loss function;
B2. according to the optimal weight omega, the server outputs local loss through the MAE loss function, and obtains the trained local model weight capable of minimizing the local loss through a local loss formula, wherein the local loss formula is as follows:
Figure BDA0003624950180000052
wherein D is k Is the size of the kth user local data set, i is the ith sample in the user k local data set,
Figure BDA0003624950180000053
a location label for the ith sample of user k,
Figure BDA0003624950180000054
sample feature, ω, for the ith sample of user k k A local model weight for the kth user;
B3. aggregating the weights of the local models of the users through a federal averaging algorithm, and updating the global model weights through the weights of the local models, wherein the federal averaging algorithm is as follows:
Figure BDA0003624950180000055
wherein the content of the first and second substances,
Figure BDA0003624950180000056
for user k, the local dataset size of the t' th round, n t The sum of the sizes of the local data sets of all the users in the t round;
B4. the user receives the updated global model weight, and updates the local model weight according to the random gradient decrease of the updated global model weight:
Figure BDA0003624950180000057
furthermore, the user performs personalized positioning through the hybrid expert model according to the trained local model and the updated global model, specifically:
s601. input
Figure BDA0003624950180000058
Training in a neural network with a sigmoid activation function as an output layer to obtain a gate control network: the formula for the gated network is as follows:
Figure BDA0003624950180000059
wherein, the first and the second end of the pipe are connected with each other,
Figure BDA00036249501800000510
g () is a local gating network function, which is a probability weight of 0 to 1 for the gating network output;
s602, a user applies for positioning service;
s603, the user downloads the global model which is trained from the server;
s604, inputting the received RSS vector to the local model which is trained and the global model which is trained by the user, and respectively outputting position prediction;
s605, predicting probability weight through a gate control network;
s606, weighting the position prediction output by the local model after training and the position prediction output by the global model after training according to the probability weight to obtain a final position prediction value, wherein a final position prediction formula is as follows:
Figure BDA0003624950180000061
a computer system comprising a plurality of computer devices, each computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processors of said plurality of computer devices when executing said computer program collectively implement the steps of said block chain based secure data sharing and value transfer method.
The invention has the following beneficial effects:
the method comprises the steps that a user constructs a local data set, an initial local model is trained through a machine learning model distributed by a server, the server simultaneously trains an initial global model, the user and the server respectively train the local model and update the global model through a federal learning algorithm, and finally, the user obtains final position prediction through a mixed expert model to perform positioning.
Drawings
Fig. 1 is a schematic flow chart of an indoor positioning method based on-line pseudo label semi-supervised learning and personalized federal learning.
Detailed Description
The invention is described in detail below with reference to the drawings and the detailed description.
Example 1
As shown in fig. 1, an indoor positioning method based on-line pseudo tag semi-supervised learning and personalized federal learning includes the following steps:
s1, a user constructs a local data set, and a server constructs a cloud data set;
s2, the server is provided with machine learning models, the server respectively issues the machine learning models to all users, the machine learning model located at the user side is called a local model, and the machine learning model located at the user side is called a local model;
s3, training a local model by a user through data with labels in a local data set to obtain an initial local model;
s4, training a global model by the server through data with labels in the cloud data set to obtain an initial global model;
s5, obtaining a trained local model and an updated global model through federal learning;
and S6, the user carries out personalized positioning through the mixed expert model according to the trained local model and the updated global model.
Example 2
As shown in fig. 1, an indoor positioning method based on-line pseudo tag semi-supervised learning and personalized federal learning includes the following steps:
s1, a user constructs a local data set, and a server constructs a cloud data set;
s2, the server is provided with machine learning models, the server respectively issues the machine learning models to each user, the machine learning model located at the user side is called a local model, and the machine learning model located at the user side is called a local model;
s3, training a local model by a user through data with labels in a local data set to obtain an initial local model;
s4, training a global model by the server through data with labels in the cloud data set to obtain an initial global model;
s5, obtaining a trained local model and an updated global model through federal learning;
and S6, the user carries out personalized positioning through the mixed expert model according to the trained local model and the updated global model.
In a specific embodiment, the local data set is constructed, specifically: assuming K users, the user is to be detectedArranging K wireless routers in the area, transmitting and receiving signals by the wireless routers, taking the received signal strength obtained by the wireless routers and the position label obtained by the inertial navigation technology as input, and constructing K received signal strength vectors xi ═ r i1 ,r i2 ,…,r ik ]While recording the position coordinate vector y i ＝[x i y i ] T The components xi and yi are respectively the abscissa and ordinate of the indoor plan, the signal intensity vector and the position coordinate vector are integrated into a local data set, and the received signal intensity is called RSS.
In one embodiment, the machine learning model is a multi-layered sensor model with fixed hyper-parameters, and the multi-layered sensor model with fixed hyper-parameters is referred to as an MLP model.
In a specific embodiment, the trained local model and the updated global model are obtained through federal learning, and the specific steps are as follows:
s401, regarding the local data set as label-free, performing on-line pseudo label semi-supervised learning on the initial local model and the initial global model to obtain a pseudo label, and labeling the local data set by the pseudo label;
s402, further training an initial local model by combining the local data set marked by the pseudo label through a knowledge distillation technology to obtain a trained local model;
and S403, the server performs federated learning on the trained local model to obtain an average value of the local model weight, and updates the global model according to the average value of the local model weight.
In a specific embodiment, the local data set is regarded as label-free, a pseudo label is obtained by performing online pseudo label semi-supervised learning on an initial local model and an initial global model, and the local data set is labeled by the pseudo label, which specifically comprises the following steps:
K1. labeling RSS data in the local data set through the initial global model and the initial local model to obtain a pseudo label estimated by the initial global model and a pseudo label estimated by the initial local model:
Figure BDA0003624950180000081
Figure BDA0003624950180000082
wherein the content of the first and second substances,
Figure BDA0003624950180000083
the pseudo-label estimated for the initial global model,
Figure BDA0003624950180000084
a pseudo label estimated for the initial local model; m is an MLP model, and M is an MLP model,
Figure BDA0003624950180000085
characteristics of RSS data for user k in the t-th round,
Figure BDA0003624950180000086
Weight parameter, ω, for global model of the t-th round user k k A local model weight for the kth user;
K2. by mixing
Figure BDA0003624950180000087
And
Figure BDA0003624950180000088
weighted sum to obtain final pseudo label
Figure BDA0003624950180000089
Figure BDA00036249501800000810
Wherein
Figure BDA00036249501800000811
To balance the parameters of the two pseudo tag estimates;
K3. by the formulaAdaptive adjustment
Figure BDA00036249501800000812
Figure BDA0003624950180000091
Wherein the sim function is the reciprocal of the euclidean distance;
K4. if it is
Figure BDA0003624950180000092
And if the local data set is larger than the set threshold, the pseudo label of the global model is used for labeling the local data set, otherwise, the pseudo label of the local model is used for labeling the local data.
In a specific embodiment, the initial local model is further trained by combining the pseudo-label labeled local data set through a knowledge distillation technology to obtain a trained local model, and the specific steps are as follows:
A1. taking the initial global model as a teacher network and the initial local model as a student network, and modifying a loss function used for training the local model;
Figure BDA0003624950180000093
wherein the content of the first and second substances,
Figure BDA0003624950180000094
a Federal knowledge distillation loss function used for training a local model, wherein beta is a knowledge distillation factor, y is a position label, Mk (x) is position prediction output by the local model, and MG (x) is position prediction output by the global model;
A2. further modifying the function used for training the local model to obtain the modified function used for training the local model:
Figure BDA0003624950180000095
wherein the content of the first and second substances,
Figure BDA0003624950180000096
loss of regularization term;
A3. and training the local model by combining the obtained modified objective function with the labeled local data set.
In a specific embodiment, the objective function is further modified, specifically: by the formula
Figure BDA0003624950180000097
To obtain
Figure BDA0003624950180000098
And will be
Figure BDA0003624950180000099
M in k (x)-M G (x) I is modified to
Figure BDA00036249501800000910
Obtaining a modified objective function:
Figure BDA0003624950180000101
in a specific embodiment, the server performs federal learning on the trained local model to obtain a trained local model weight, and updates the global model according to the trained local model weight, which specifically includes the following steps:
B1. obtaining a weight parameter omega capable of minimizing the global model loss through a loss formula of the global model, wherein the loss formula of the global model is as follows:
Figure BDA0003624950180000102
wherein F () is a global model loss function, N is a total number of users, F k () Is a local loss function;
B2. according to the optimal weight omega, the server outputs local loss through the MAE loss function, and obtains the trained local model weight capable of minimizing the local loss through a local loss formula, wherein the local loss formula is as follows:
Figure BDA0003624950180000103
wherein D is k Is the size of the kth user local data set, i is the ith sample in the user k local data set,
Figure BDA0003624950180000104
a location label for the ith sample of user k,
Figure BDA0003624950180000105
sample feature, ω, for the ith sample of user k k A local model weight for the kth user;
B3. aggregating the weights of the local models of the users through a federal averaging algorithm, and updating the global model weights through the weights of the local models, wherein the federal averaging algorithm is as follows:
Figure BDA0003624950180000106
wherein the content of the first and second substances,
Figure BDA0003624950180000107
for user k, the local dataset size of the t' th round, n t The sum of the sizes of the local data sets of all the users in the t round;
B4. the user receives the updated global model weight, and updates the local model weight according to the random gradient decrease of the updated global model weight:
Figure BDA0003624950180000108
in a specific embodiment, the server obtains an optimal weight parameter ω by using a loss formula of the global model, specifically: obtaining a weight parameter omega capable of minimizing the global model loss through a loss formula of the global model, wherein the loss formula of the global model is as follows:
Figure BDA0003624950180000111
where F () is the global model loss function and N is the total number of users.
Example 3
As shown in fig. 1, an indoor positioning method based on-line pseudo tag semi-supervised learning and personalized federal learning includes the following steps:
s1, a user constructs a local data set, and a server constructs a cloud data set;
s2, the server is provided with machine learning models, the server respectively issues the machine learning models to each user, the machine learning model located at the user side is called a local model, and the machine learning model located at the user side is called a local model;
s3, training a local model by a user through data with labels in a local data set to obtain an initial local model;
s4, the server trains a global model through data with labels in the cloud data set to obtain an initial global model;
s5, obtaining a trained local model and an updated global model through federal learning;
and S6, the user carries out personalized positioning through the mixed expert model according to the trained local model and the updated global model.
In a specific embodiment, the local data set is constructed, specifically: supposing that K users are provided, K wireless routers are arranged in the area to be detected, the wireless routers transmit and receive signals, and the wireless routers receive the signalsThe signal intensity and the position label obtained by the inertial navigation technology are used as input to construct K received signal intensity vectors xi ═ r i1 ,r i2 ,…,r ik ]While recording the position coordinate vector y i ＝[x i y i ] T The components xi and yi are respectively the abscissa and ordinate of the indoor plan, the signal intensity vector and the position coordinate vector are integrated into a local data set, and the received signal intensity is called RSS.
In one embodiment, the machine learning model is a multi-layered sensor model with fixed hyper-parameters, and the multi-layered sensor model with fixed hyper-parameters is referred to as an MLP model.
In a specific embodiment, the trained local model and the updated global model are obtained through federal learning, and the specific steps are as follows:
s401, regarding the local data set as label-free, performing on-line pseudo label semi-supervised learning on the initial local model and the initial global model to obtain a pseudo label, and labeling the local data set by the pseudo label;
s402, further training an initial local model by combining the local data set marked by the pseudo label through a knowledge distillation technology to obtain a trained local model;
and S403, the server performs federated learning on the trained local model to obtain an average value of the local model weight, and updates the global model according to the average value of the local model weight.
In a specific embodiment, the local data set is regarded as label-free, a pseudo label is obtained by performing online pseudo label semi-supervised learning on an initial local model and an initial global model, and the local data set is labeled by the pseudo label, which specifically comprises the following steps:
K1. labeling RSS data in the local data set through the initial global model and the initial local model to obtain a pseudo label estimated by the initial global model and a pseudo label estimated by the initial local model:
Figure BDA0003624950180000121
Figure BDA0003624950180000122
wherein the content of the first and second substances,
Figure BDA0003624950180000123
the pseudo-label estimated for the initial global model,
Figure BDA0003624950180000124
a pseudo label estimated for the initial local model; m is an MLP model, and M is an MLP model,
Figure BDA0003624950180000125
characteristics of RSS data for the t-th round of user k,
Figure BDA0003624950180000126
Weight parameter, ω, for global model of the t-th round user k k A local model weight for the kth user;
K2. by mixing
Figure BDA0003624950180000127
And
Figure BDA0003624950180000128
weighted sum to obtain final pseudo label
Figure BDA0003624950180000129
Figure BDA00036249501800001210
Wherein
Figure BDA00036249501800001211
Parameters for balancing two pseudo tag estimates;
K3. adaptive adjustment by the following formula
Figure BDA00036249501800001212
Figure BDA00036249501800001213
Wherein the sim function is the reciprocal of the euclidean distance;
K4. if it is
Figure BDA00036249501800001214
And if the local data set is larger than the set threshold, the pseudo label of the global model is used for labeling the local data set, otherwise, the pseudo label of the local model is used for labeling the local data.
In a specific embodiment, the initial local model is further trained by combining the pseudo-label labeled local data set through a knowledge distillation technology to obtain a trained local model, and the specific steps are as follows:
A1. taking the initial global model as a teacher network and the initial local model as a student network, and modifying a loss function used by training the local model;
Figure BDA0003624950180000131
wherein the content of the first and second substances,
Figure BDA0003624950180000132
a federal knowledge distillation loss function used for training a local model, wherein beta is a knowledge distillation factor, y is a position label, Mk (x) is position prediction output by the local model, and MG (x) is position prediction output by the global model;
A2. further modifying the function used for training the local model to obtain the modified function used for training the local model:
Figure BDA0003624950180000133
wherein the content of the first and second substances,
Figure BDA0003624950180000134
loss of regularization term;
A3. and training the local model by combining the obtained modified objective function with the labeled local data set.
In a specific embodiment, the objective function is further modified, specifically: by the formula
Figure BDA0003624950180000135
To obtain
Figure BDA0003624950180000136
And will be
Figure BDA0003624950180000137
M in k (x)-M G (x) I is modified to
Figure BDA0003624950180000138
Obtaining a modified objective function:
Figure BDA0003624950180000139
the hybrid expert model and knowledge distillation techniques are both individualized federal learning techniques.
In a specific embodiment, the server performs federal learning on the trained local model to obtain a trained local model weight, and updates the global model according to the trained local model weight, which specifically includes the following steps:
B1. obtaining a weight parameter omega capable of minimizing the global model loss through a loss formula of the global model, wherein the loss formula of the global model is as follows:
Figure BDA0003624950180000141
wherein F () is a global model loss function, N is a total number of users, F k () Is a local loss function;
B2. according to the optimal weight omega, the server outputs local loss through the MAE loss function, and obtains the trained local model weight capable of minimizing the local loss through a local loss formula, wherein the local loss formula is as follows:
Figure BDA0003624950180000142
wherein D is k Is the size of the kth user local data set, i is the ith sample in the user k local data set,
Figure BDA0003624950180000143
a location label for the ith sample of user k,
Figure BDA0003624950180000144
sample feature, ω, for the ith sample of user k k A local model weight for the kth user;
B3. aggregating the weights of the local models of the users through a federal averaging algorithm, and updating the global model weights through the weights of the local models, wherein the federal averaging algorithm is as follows:
Figure BDA0003624950180000145
wherein the content of the first and second substances,
Figure BDA0003624950180000146
for user k, the local dataset size of the t' th round, n t The sum of the sizes of the local data sets of all the users in the t round;
B4. the user receives the updated global model weight, and updates the local model weight according to the random gradient decrease of the updated global model weight:
Figure BDA0003624950180000147
in a specific embodiment, the user performs personalized positioning through the hybrid expert model according to the trained local model and the updated global model, specifically:
s601. input
Figure BDA0003624950180000148
Training in a neural network with a sigmoid activation function as an output layer to obtain a gate control network: the formula for the gated network is as follows:
Figure BDA0003624950180000151
wherein the content of the first and second substances,
Figure BDA0003624950180000152
g () is a local gating network function, which is a probability weight of 0 to 1 for the gating network output;
s602, a user applies for positioning service;
s603, the user downloads the global model which is trained from the server;
s604, inputting the received RSS vector to the local model which is trained and the global model which is trained by the user, and respectively outputting position prediction;
s605, predicting probability weight through a gate control network;
s606, weighting the position prediction output by the local model after training and the position prediction output by the global model after training according to the probability weight to obtain a final position prediction value, wherein a final position prediction formula is as follows:
Figure BDA0003624950180000153
in the embodiment, a user constructs a local data set, trains an initial local model through a machine learning model distributed by a server, the server trains an initial global model, the user and the server train the local model and update the global model through a federal learning algorithm respectively, and finally, a final position prediction is obtained by a user mixed expert model for positioning. The invention solves the problem that the prior art ignores the difference of the height dynamic state and the positioning requirement of the local data and can not carry out personalized positioning, and has the characteristics of high efficiency and high precision.
Example 4
A computer system comprising a plurality of computer devices, each computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processors of the plurality of computer devices when executing the computer program collectively implement the steps of the blockchain-based secure data sharing and value transfer method of embodiment 1, or embodiment 2, or embodiment 3.
It should be understood that the above-described embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention should be included in the protection scope of the claims of the present invention.

Claims (10)

1. An indoor positioning method based on-line pseudo label semi-supervised learning and personalized federal learning is characterized in that: the method comprises the following steps:
s1, a user constructs a local data set, and a server constructs a cloud data set;
s2, the server is provided with machine learning models, the server respectively issues the machine learning models to each user, the machine learning model located at the user side is called a local model, and the machine learning model located at the user side is called a local model;
s3, training a local model by a user through data with labels in a local data set to obtain an initial local model;
s4, training a global model by the server through data with labels in the cloud data set to obtain an initial global model;
s5, obtaining a trained local model and an updated global model through federal learning;
and S6, the user carries out personalized positioning through the mixed expert model according to the trained local model and the updated global model.
2. The indoor positioning method based on-line pseudo-label semi-supervised learning and personalized federal learning of claim 1, wherein; the user constructs a local data set, specifically: the user constructs K received signal strength vectors xi ═ r by receiving the signal strength transmitted by K wireless routers i1 ,r i2 ,…,r ik ]And simultaneously integrating the signal intensity vector and the position coordinate vector into a local data set by recording the position coordinate vector yi.
3. The indoor positioning method based on-line pseudo tag semi-supervised learning and personalized federal learning of claim 1, wherein: the machine learning model is a multilayer perceptron model with fixed hyper-parameters, and the multilayer perceptron model with the fixed hyper-parameters is called an MLP model.
4. The indoor positioning method based on-line pseudo-label semi-supervised learning and personalized federal learning of claim 2, characterized in that; obtaining a trained local model and an updated global model through federal learning, and the method comprises the following specific steps:
s401, regarding the local data set as label-free, performing on-line pseudo label semi-supervised learning on the initial local model and the initial global model to obtain a pseudo label, and labeling the local data set by the pseudo label;
s402, further training an initial local model by combining the pseudo label labeled local data set through a knowledge distillation technology to obtain a trained local model;
and S403, the server performs federated learning on the trained local model to obtain an average value of the local model weight, and updates the global model according to the average value of the local model weight.
5. The indoor positioning method based on-line pseudo tag semi-supervised learning and personalized federal learning of claim 4, wherein: regarding the local data set as label-free, performing on-line pseudo label semi-supervised learning on the initial local model and the initial global model to obtain a pseudo label, and labeling the local data set by the pseudo label, wherein the specific steps are as follows:
K1. labeling RSS data in the local data set through the initial global model and the initial local model to obtain a pseudo label estimated by the initial global model and a pseudo label estimated by the initial local model:
Figure FDA0003624950170000021
Figure FDA0003624950170000022
wherein the content of the first and second substances,
Figure FDA0003624950170000023
the pseudo-label estimated for the initial global model,
Figure FDA0003624950170000024
a pseudo label estimated for the initial local model; m is an MLP model, and M is an MLP model,
Figure FDA0003624950170000025
characteristics of RSS data for the t-th round of user k,
Figure FDA0003624950170000026
Weight parameter, ω, for global model of the t-th round user k k A local model weight for the kth user;
K2. by mixing
Figure FDA0003624950170000027
And
Figure FDA0003624950170000028
weighted sum to obtain final pseudo label
Figure FDA0003624950170000029
Figure FDA00036249501700000210
Wherein
Figure FDA00036249501700000211
Parameters for balancing two pseudo tag estimates;
K3. adaptive adjustment by the following formula
Figure FDA00036249501700000212
Figure FDA00036249501700000213
Wherein the sim function is the reciprocal of the euclidean distance;
K4. if it is
Figure FDA00036249501700000214
And if the local data set is larger than the set threshold, the pseudo label of the global model is used for labeling the local data set, otherwise, the pseudo label of the local model is used for labeling the local data.
6. The block chain based secure data sharing and value transfer method of claim 4, wherein: combining the pseudo label labeled local data set, further training the initial local model by a knowledge distillation technology to obtain a trained local model, and specifically comprising the following steps:
A1. taking the initial global model as a teacher network and the initial local model as a student network, and modifying a loss function used for training the local model;
Figure FDA0003624950170000031
wherein the content of the first and second substances,
Figure FDA0003624950170000032
a federal knowledge distillation loss function used for training a local model, wherein beta is a knowledge distillation factor, y is a position label, Mk (x) is position prediction output by the local model, and MG (x) is position prediction output by the global model;
A2. further modifying the function used for training the local model to obtain the modified function used for training the local model:
Figure FDA0003624950170000033
wherein the content of the first and second substances,
Figure FDA0003624950170000034
loss of regularization term;
A3. and training a local model by combining the obtained modified target function with the labeled local data set.
7. The block chain based secure data sharing and value transfer method of claim 5, wherein: further modifying the objective function, specifically: by the formula
Figure FDA0003624950170000035
To obtain
Figure FDA0003624950170000036
And will be
Figure FDA0003624950170000037
In | | | M k (x)-M G (x) Modified to
Figure FDA0003624950170000038
Obtaining a modified objective function:
Figure FDA0003624950170000039
8. the block chain based secure data sharing and value transfer method of claim 6, wherein: the server performs federated learning on the trained local model to obtain a trained local model weight, and updates the global model through the trained local model weight, and the specific steps are as follows:
B1. obtaining a weight parameter omega capable of minimizing the global model loss through a loss formula of the global model, wherein the loss formula of the global model is as follows:
Figure FDA0003624950170000041
wherein F () is a global model loss function, N is a total number of users, F k () Is a local loss function;
B2. according to the optimal weight parameter omega, the server outputs local loss through the MAE loss function, and obtains the trained local model weight capable of minimizing the local loss through a local loss formula, wherein the local loss formula is as follows:
Figure FDA0003624950170000042
wherein D is k Is the size of the kth user local data set, i is the ith sample in the user k local data set,
Figure FDA0003624950170000043
a location label for the ith sample of user k,
Figure FDA0003624950170000044
sample feature, ω, for the ith sample of user k k A local model weight for the kth user;
B3. aggregating the weights of the local models of the users through a federal averaging algorithm, and updating the global model weights through the weights of the local models, wherein the federal averaging algorithm is as follows:
Figure FDA0003624950170000045
wherein the content of the first and second substances,
Figure FDA0003624950170000046
for user k, the local dataset size of the t' th round, n t The sum of the sizes of the local data sets of all the users in the t round;
B4. the user receives the updated global model weight, and updates the local model weight according to the random gradient decrease of the updated global model weight:
Figure FDA0003624950170000047
9. the block chain based secure data sharing and value transfer method of claim 5, wherein: the user carries out personalized positioning through the mixed expert model according to the trained local model and the updated global model, and the personalized positioning method specifically comprises the following steps:
s601. input
Figure FDA0003624950170000048
Training in a neural network with a sigmoid activation function as an output layer to obtain a gate control network: the formula for the gated network is as follows:
Figure FDA0003624950170000051
wherein the content of the first and second substances,
Figure FDA0003624950170000052
g () is a local gating network function, which is a probability weight of 0 to 1 for the gating network output;
s602, a user applies for positioning service;
s603, the user downloads the global model which is trained from the server;
s604, inputting the received RSS vector to the local model which is trained and the global model which is trained by the user, and respectively outputting position prediction;
s605, predicting probability weight through a gate control network;
s606, weighting the position prediction output by the local model after training and the position prediction output by the global model after training according to the probability weight to obtain a final position prediction value, wherein the final position prediction formula is as follows:
Figure FDA0003624950170000053
10. a computer system comprising a plurality of computer devices, each computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processors of the plurality of computer devices when executing the computer program collectively implement the steps of the blockchain based secure data sharing and value transfer method of any one of claims 1 to 10.
CN202210467468.6A 2022-04-29 2022-04-29 Indoor positioning method based on-line pseudo label semi-supervised learning and personalized federal learning Pending CN114897063A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210467468.6A CN114897063A (en) 2022-04-29 2022-04-29 Indoor positioning method based on-line pseudo label semi-supervised learning and personalized federal learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210467468.6A CN114897063A (en) 2022-04-29 2022-04-29 Indoor positioning method based on-line pseudo label semi-supervised learning and personalized federal learning

Publications (1)

Publication Number Publication Date
CN114897063A true CN114897063A (en) 2022-08-12

Family

ID=82719232

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210467468.6A Pending CN114897063A (en) 2022-04-29 2022-04-29 Indoor positioning method based on-line pseudo label semi-supervised learning and personalized federal learning

Country Status (1)

Country Link
CN (1) CN114897063A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115310130A (en) * 2022-08-15 2022-11-08 南京航空航天大学 Multi-site medical data analysis method and system based on federal learning

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115310130A (en) * 2022-08-15 2022-11-08 南京航空航天大学 Multi-site medical data analysis method and system based on federal learning
CN115310130B (en) * 2022-08-15 2023-11-17 南京航空航天大学 Multi-site medical data analysis method and system based on federal learning

Similar Documents

Publication Publication Date Title
Gharehchopogh et al. Advances in sparrow search algorithm: a comprehensive survey
US10451712B1 (en) Radar data collection and labeling for machine learning
US20180268292A1 (en) Learning efficient object detection models with knowledge distillation
Campbell et al. The explosion of artificial intelligence in antennas and propagation: How deep learning is advancing our state of the art
CN113807399B (en) Neural network training method, neural network detection method and neural network training device
CN110377587B (en) Migration data determination method, device, equipment and medium based on machine learning
Liu et al. Geomagnetism-based indoor navigation by offloading strategy in NB-IoT
CN114091554A (en) Training set processing method and device
Yadav et al. A systematic review of localization in WSN: Machine learning and optimization‐based approaches
CN113326826A (en) Network model training method and device, electronic equipment and storage medium
CN114241226A (en) Three-dimensional point cloud semantic segmentation method based on multi-neighborhood characteristics of hybrid model
Mallik et al. Paving the way with machine learning for seamless indoor–outdoor positioning: A survey
CN114897063A (en) Indoor positioning method based on-line pseudo label semi-supervised learning and personalized federal learning
Vahidnia et al. A hierarchical signal-space partitioning technique for indoor positioning with WLAN to support location-awareness in mobile map services
Kota et al. IOT‐HML: A hybrid machine learning technique for IoT enabled industrial monitoring and control system
Bai et al. UAV-supported intelligent truth discovery to achieve low-cost communications in mobile crowd sensing
US20240095539A1 (en) Distributed machine learning with new labels using heterogeneous label distribution
CN107124761B (en) Cellular network wireless positioning method fusing PSO and SS-ELM
Janabi et al. A new localization mechanism in IoT using grasshopper optimization algorithm and DVHOP algorithm
Gokulakrishnan et al. Maliciously roaming person's detection around hospital surface using intelligent cloud-edge based federated learning
Ijaz et al. A UAV assisted edge framework for real-time disaster management
US20230055079A1 (en) Method of load forecasting via attentive knowledge transfer, and an apparatus for the same
Barolli et al. Web, Artificial Intelligence and Network Applications: Proceedings of the Workshops of the 33rd International Conference on Advanced Information Networking and Applications (WAINA-2019)
KR102474974B1 (en) Method, device and system for analyzing brand based on artificial intelligence
Hensel et al. Object Detection and Mapping with Unmanned Aerial Vehicles Using Convolutional Neural Networks

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination