CN112637883A - Federal learning method with robustness to wireless environment change in power Internet of things - Google Patents

Federal learning method with robustness to wireless environment change in power Internet of things Download PDF

Info

Publication number
CN112637883A
CN112637883A CN202011432782.8A CN202011432782A CN112637883A CN 112637883 A CN112637883 A CN 112637883A CN 202011432782 A CN202011432782 A CN 202011432782A CN 112637883 A CN112637883 A CN 112637883A
Authority
CN
China
Prior art keywords
model
local
neural network
wireless environment
global
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.)
Granted
Application number
CN202011432782.8A
Other languages
Chinese (zh)
Other versions
CN112637883B (en
Inventor
付美明
刘庆扬
王祥
那辰星
王学良
李晓霞
汤志颖
汪志亮
李莉华
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.)
State Grid Information and Telecommunication Co Ltd
Beijing Smartchip Microelectronics Technology Co Ltd
China Gridcom Co Ltd
Shenzhen Zhixin Microelectronics Technology Co Ltd
Original Assignee
State Grid Information and Telecommunication Co Ltd
Beijing Smartchip Microelectronics Technology Co Ltd
China Gridcom Co Ltd
Shenzhen Zhixin Microelectronics Technology Co Ltd
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 State Grid Information and Telecommunication Co Ltd, Beijing Smartchip Microelectronics Technology Co Ltd, China Gridcom Co Ltd, Shenzhen Zhixin Microelectronics Technology Co Ltd filed Critical State Grid Information and Telecommunication Co Ltd
Priority to CN202011432782.8A priority Critical patent/CN112637883B/en
Publication of CN112637883A publication Critical patent/CN112637883A/en
Application granted granted Critical
Publication of CN112637883B publication Critical patent/CN112637883B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W24/00—Supervisory, monitoring or testing arrangements
    • H04W24/06—Testing, supervising or monitoring using simulated traffic
    • 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
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/08—Learning methods
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Abstract

The invention provides a federal learning method with robustness to wireless environment change in a power internet of things, which comprises the following steps: building a neural network model, initializing weight parameters of the neural network model, and distributing a data set to multiple User Equipment (UEs); training a neural network model; uploading the parameters of the trained neural network model to a wireless Access Point (AP), and simulating wireless environment factors to analyze the convergence performance of the global FL model; analyzing the convergence performance condition of the global FL model through simulating wireless environment factors; a client side executing the federated learning is selected at an access point AP by adopting a check mechanism and a scheduling strategy to weaken the influence of corresponding wireless environment factors on the convergence of the global model, and the federated learning method has robustness on wireless environment changes. The method uses a verification mechanism at the access point to detect the local FL model received by the access point and adopts a scheduling strategy to select the client side executing the federal learning to weaken the influence of infinite factors, so that the federal learning framework has robustness to the change of the wireless environment.

Description

Federal learning method with robustness to wireless environment change in power Internet of things
Technical Field
The disclosure relates to the technical field of wireless communication networks, in particular to a federal learning method with robustness to wireless environment changes in the power internet of things.
Background
The power internet of things is an application of the internet of things in an intelligent power grid, is a result obtained when an information communication technology develops to a certain stage, effectively integrates communication infrastructure resources and power system infrastructure resources, improves the informatization level of the power system, improves the utilization efficiency of the existing infrastructure of the power system, and provides important technical support for links such as power grid generation, transmission, transformation, distribution and power utilization.
The standard Machine learning (Machine learning) method requires training data to be focused on one Machine or data center. Due to latency, bandwidth, and privacy limitations, it is often not feasible to transmit a user's collected data set to a data center or cloud to implement a powerful machine learning solution. Meanwhile, as barriers which are difficult to break exist between data sources, the data required by artificial intelligence generally relates to multiple fields. In most industries, data exists in an isolated island form, and due to problems of industry competition, privacy safety, complex administrative procedures and the like, even if data integration is realized among different departments of the same company, important resistance is faced, and in reality, it is almost impossible or the required cost is huge to integrate data scattered in various places and various mechanisms.
Federal Learning (FL), which allows model training to be decoupled from the need for direct access to the original training data. Unlike centralized learning in a data center, the FL typically operates in a wireless edge network where the communication medium is resource constrained and unreliable. The iterative algorithm running on FL requires connections with very low delay and high throughput between computational units, but the AP usually needs to link a large number of UEs over a resource-constrained spectrum, so in each round of global aggregation only a limited number of multi-user device-side UEs are allowed to send their trained weight parameters over unreliable channels. Due to bandwidth limitations, only a portion of the UEs may be scheduled for updating in each iteration. Due to the shared nature of the wireless medium, the transmission is subject to interference and cannot be guaranteed.
Disclosure of Invention
The federal learning method which is robust to wireless environment changes in the power internet of things is provided by the disclosure aiming at the problems.
In order to solve at least one of the above technical problems, the present disclosure proposes the following technical solutions:
the federal learning method with robustness to wireless environment changes in the power internet of things is provided, and comprises the following steps:
the method comprises the following steps: building a neural network model for learning, initializing weight parameters of the neural network model, and distributing a data set to multiple User Equipment (UEs);
step two: carrying out federal learning through the built neural network model to obtain a trained neural network model;
step three: uploading the parameters of the trained neural network model to a wireless Access Point (AP), and simulating wireless environment factors to analyze the convergence performance of the global FL model at the moment;
step four: analyzing the convergence performance condition of the global FL model at the moment through simulating the wireless environment factors;
step five: and selecting a client executing the federated learning at the wireless access point AP by adopting a check mechanism and a scheduling strategy according to the acquired convergence performance condition to weaken the influence of the corresponding wireless environment factors on the convergence of the global model, and finally obtaining the federated learning method having robustness on the wireless environment change in the power Internet of things.
In some embodiments, building the neural network model for learning in the first step comprises:
determining a type of neural network used;
determining the number of layers of the used neural network and the number of ganglion points of each layer;
the weights and bias parameters of all neural networks are initialized.
In some embodiments, step two comprises:
determining parameters of a neural network model, training by using a data set based on the neural network model, and determining the parameters of the neural network model to comprise communication round times, client number K, client score C, local training generation E, local small batch B, test data small batch, learning rate lr and an optimization method;
distributing the original data set to K clients in a uniform distribution or non-uniform distribution mode;
the K clients learn by using local training data sets thereof until the total loss function reaches the minimum value, and upload the local parameter model of the neural network to the wireless access point AP through the wireless link;
after the wireless access point AP receives local model parameters from all the multi-user equipment ends UEs, the wireless access point AP starts to aggregate K local parameter models to obtain a global average model and transmits the model to all the multi-user equipment ends UEs, the multi-user equipment ends UEs start a new round of learning according to the updated local FL model, and the trained models are obtained through the learning of a plurality of communication rounds.
In some embodiments, in the step of distributing the original data set to the K clients in a uniformly distributed or non-uniformly distributed manner,
the uniform and uniform distribution is to directly divide the original data set into K equal parts, and the non-uniform and uniform distribution requires that the original data set is sorted according to the size of the label data and then distributed to K clients, namely the data set of each client is non-uniformly distributed.
In some embodiments, the uploading of the parameters of the trained neural network model to the wireless access point AP in the third step includes:
in the training process of the communication turn, each user equipment firstly trains the local FL model w by using the training data of each user equipmentk；
Each user equipment transmits w through a wireless cellular linkkSending the data to the AP;
the AP receives the local FL models of all the participating user equipment ends, updates the global FL model and updates the global FL model wt+1Sending the data to all user equipment terminals to optimize a local FL model; the global FL model is obtained from all local FL models of all ue participating in learning, and can be represented by the following formula:
Figure BDA0002825675790000031
wherein, wt+1Is a global FL model generated by the wireless access point AP, nkRepresenting the size of a local data set of a user device K, K representing a Kth client, t representing a t communication turn, wkRepresenting the local FL model.
In some embodiments, the convergence performance of the global FL model at this time is analyzed by simulating wireless environment factors in step four, wherein the wireless environment factors include bandwidth limitation, inter-cell interference and path loss,
bandwidth limitations in wireless environment factors are simulated by limiting the number of clients participating in federal learning,
inter-zone interference and path loss in wireless environmental factors are simulated by adding errors to local FL model parameters.
In some embodiments, the step of simulating inter-zone interference and path loss in the wireless environment factors by adding errors to the local FL model parameters comprises:
in each communication round, selecting K.C (C belongs to (0, 1)) clients from K clients to participate in training;
in K X client, testing the influence of error of model parameters of client selected from A X K X C, b X K X C and C X K X C on the identification precision of final polymerization model; wherein a, b, c is belonged to (0, 1) and different from each other;
the error mode of the client model parameter comprises the following steps:
setting L layers of the model neural network layer, selecting M layers to make mistakes, traversing all parameters of the M layers, and enabling each parameter to have the probability of making mistakes of Q, wherein M is less than L, and Q is the wrong setting of (0, 1):
W＝W+randomuniform(-1,1)
wherein randomniform (-1, 1) randomly generates a floating point number from-1 to 1, and w represents the local FL model.
In some embodiments, the scheduling policy in step five comprises:
allocating resource-limited radio channels to appropriate multi-user equipment UEs, including in each communication round, the wireless access point AP randomly selecting multiple relevant multi-user equipment UEs uniformly for parameter updating, and allocating a dedicated sub-channel to transmit training parameters for each selected UE.
In some embodiments, the checking mechanism is used to check for data errors in the local FL model received at the wireless access point AP;
the wireless access point AP does not use the received data error local FL model to update the global FL model;
the wireless access point AP will update the global FL model directly using the remaining correct local FL model.
The beneficial effects of this disclosure are: the invention uses a check mechanism at the access point to detect the received local FL model and adopts a scheduling strategy to select the client end for executing the federal learning to weaken the influence of infinite factors, so that the federal learning framework has robustness to the change of the wireless environment.
In addition, in the technical solutions of the present disclosure, the technical solutions can be implemented by adopting conventional means in the art, unless otherwise specified.
Drawings
In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the drawings used in the following description of the embodiments or the prior art will be briefly described, and it is obvious that the drawings in the following description are some embodiments of the present disclosure, and other drawings can be obtained by those skilled in the art without creative efforts.
Fig. 1 is a schematic structural diagram of a federal learning method with robustness to wireless environment changes in a power internet of things according to an embodiment of the present disclosure;
fig. 2 is a flow chart of a federal learning method robust to wireless environment changes in the power internet of things of the disclosure.
Detailed Description
In order to make the objects, technical solutions and advantages of the present disclosure more clearly understood, the present disclosure is further described in detail below with reference to the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are merely illustrative of some, but not all embodiments of the disclosure, and are not to be construed as limiting the disclosure. All other embodiments, which can be derived by a person skilled in the art from the embodiments disclosed herein without making any creative effort, shall fall within the protection scope of the present disclosure.
It should be noted that the terms "comprises" and "comprising," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or server that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.
The UEs in this application are user equipments, i.e. multi-user equipment terminals, access points: access point, FL, i.e. fed learning refers to federal learning, UE is here a client, i.e. user Equipment.
Example 1:
referring to the accompanying drawings 1-2 in the specification, a federal learning method for robustness to wireless environment changes in a power internet of things is provided by one embodiment of the application, and the method comprises the following steps:
the method comprises the following steps: s1, building a neural network model for learning, initializing weight parameters of the neural network model, and distributing a data set to multiple User Equipment (UEs);
step two: s2, carrying out federal learning through the built neural network model to obtain a trained neural network model;
step three: s3, uploading the parameters of the trained neural network model to a wireless access point AP, and simulating wireless environment factors to analyze the convergence performance of the global FL model at the moment;
step four: s4, analyzing the convergence performance condition of the global FL model at the moment through simulating the wireless environment factors;
step five: s5, selecting a client executing federated learning at the wireless access point AP by adopting a check mechanism and a scheduling strategy according to the acquired convergence performance condition to weaken the influence of corresponding wireless environment factors on the convergence of the global model, and finally obtaining the federated learning method having robustness on wireless environment changes in the power Internet of things.
As shown in fig. 1, fig. 1 (a) represents each UE computing a single update based on its locally stored data, fig. 1 (B) the access point AP aggregates the updates received from the UEs to build a new global model, fig. 1 (C) sends the new model back to the UEs, and the process is repeated.
Wherein, the step one of building the neural network model for learning comprises the following steps:
determining a type of neural network used; wherein, the type of the neural network can be MLP or CNN;
determining the number of layers of the used neural network and the number of ganglion points of each layer;
the weights and bias parameters of all neural networks are initialized.
Specifically, the second step comprises:
determining parameters of a neural network model, namely selecting a universal model, and training by using a data set based on the universal model, specifically, determining the number of communication rounds, the number K of clients, the client score C, a local training generation E, a local small batch B, a test data small batch, a learning rate lr and an optimization method; the number of communication rounds is also called communication rounds, and refers to one-time training and updating exchange between the AP and related UEs, the number K of the clients refers to the number of the UEs associated with the AP, and the client score C refers to the number of the UEs participating in federal learning compared with the number of the clients;
distributing the original data set to K clients in a uniform distribution or non-uniform distribution mode;
the K clients learn by using local training data sets thereof until the total loss function reaches the minimum value, and upload the local parameter model of the neural network to the wireless access point AP through the wireless link;
after the wireless access point AP receives local model parameters from all the multi-user equipment ends UEs, the wireless access point AP starts to aggregate K local parameter models to obtain a global average model and transmits the model to all the multi-user equipment ends UEs, the multi-user equipment ends UEs start a new round of learning according to the updated local FL model, and the trained models are obtained through the learning of a plurality of communication rounds.
For example, the following steps are carried out: a cellular network comprising an AP and a set of associated K multi-user equipment UEs that cooperate to perform FL algorithms for data analysis and inference, e.g., a wireless access point assumes that the AP has K associated multi-user equipment UEs evenly distributed within its Thiessen polygon, then in this network a fixed amount of frequency spectrum is divided evenly into N radio access channels, where N < K. For a generic user equipment k, it is considered to be equipped with nKA local data set of individual sample points. For a Mnist data set, the data set is disorganized, 60000/K label data and 10000/K test data are distributed to a client in sequence, the score of the client is C, namely the number of user equipment participating in federal learning is m-K C, the local training generation is local ep, the local small batch is local bs, the learning rate is lr, and the communication round is epochs. The task is to classify handwritten digits using the MNIST dataset or color images using the CIFAR-10 dataset.
The neural network model has two types, namely MLP and CNN, wherein the MLP model includes an input layer, a hidden layer and an output layer, the number of hidden layers and the number of nodes are selected according to the needs, the number of nodes of the input layer and the output layer is determined according to specific data sets and classification conditions, but for the above two classification problems, the number of nodes of the output layer can be both 10. The CNN model contains a shim layer that activates volumes using the Relu function, a fully-connected layer that activates using the Relu function, and one softmax output layer. The number of convolutional layers and channels, the number of pooling layers, the number of channels, and the pooling method used are selected to be appropriate values as required.
Specifically, the step of minimizing the overall loss function includes: at each of the APs, the AP is,the target is to learn a statistical model through data existing on K related multi-user equipment terminals (UEs), namely AP needs to fit a vector w epsilon RdIn order to minimize a specific loss function by using the entire data set of all multi-user equipment UEs under its service. In form, such tasks may be represented as:
Figure BDA0002825675790000081
Figure BDA0002825675790000082
the formula is a multi-client minimum loss function, is a loss function of an AP (access point), and shows that the aim of local training is to minimize the sum of the loss functions of all user equipment participating in learning, and the formula can be used as a termination condition of local model training.
Wherein n isk＝|DKL represents the size of the local data set of the user equipment k,
Figure BDA0002825675790000083
representing the size, w, of the entire data settIs formed by APGenerated global FL model, f (w)i) Is a loss function, different loss functions can be defined for different FL algorithms;
k represents the number of UEs, i.e., the number of UEs associated with the AP;
in particular, the constraints are to ensure that once the FL algorithm converges, all user devices and APs will share the same FL model for their learning task, i.e. the download of the global FL model to all clients;
the purpose of the formula expressing the local training is to minimize the sum of loss functions of all clients participating in learning, and the sum can be used as a termination condition of the local model training.
In the FL algorithm, the local FL model w for each user kkIs dependent on the global model wtAnd a global model wtThe update of (2) depends on the local FL models of all users. Local FL model wkIs dependent on the learning algorithm. Global model wtIs given by:
in the training process, each user first trains the local FL model w using its training datakThen w is transmitted over the wireless cellular linkkTo the AP. Once the AP receives the local FL models for all participating users, it will update the global FL model based on the above equation and will update the global FL model wt+1Sent to all users to optimize the local FL model. Over time, the AP and multi-user devices UEs can find their optimal FL models and use them to minimize the loss function.
Preferably, in the step of distributing the original data set to the K clients in a uniformly distributed or non-uniformly distributed manner,
the uniform and uniform distribution is to directly divide the original data set into K equal parts, and the non-uniform and uniform distribution requires that the original data set is sorted according to the size of the label data and then distributed to K clients, namely the data set of each client is non-uniformly distributed.
Preferably, the uploading the parameters of the trained neural network model to the wireless access point AP in the third step includes:
in the training process of the communication turn, each user equipment firstly trains the local FL model w by using the training data of each user equipmentk；
Each user equipment transmits w through a wireless cellular linkkSending the data to the AP;
the AP receives the local FL models of all the participating user equipment ends, updates the global FL model and updates the global FL model wt+1Sending the data to all user equipment terminals to optimize a local FL model; the global FL model is obtained from all local FL models of all ue participating in learning, and can be represented by the following formula:
Figure BDA0002825675790000101
wherein, wt+1Is a global FL model generated by the wireless access point AP, nkRepresenting the size of a local data set of a user device K, K representing a Kth client, t representing a t communication turn, wkRepresenting the local FL model.
Preferably, the convergence performance of the global FL model at this time is analyzed by simulating wireless environment factors in step four, wherein the wireless environment factors include bandwidth limitation, inter-cell interference and path loss,
bandwidth limitation in wireless environment factors is simulated by limiting the number of clients participating in federal learning;
inter-zone interference and path loss in wireless environmental factors are simulated by adding errors to local FL model parameters.
Specifically, for the limitation of the bandwidth, the wireless access point AP should select only a subset of the UEs on the multi-user device side and update its parameters at the same time, so as to keep the communication time within an acceptable range. For this reason, the scheduling strategy plays a key role in allocating resource-limited radio channels to appropriate multi-user equipment UEs. The invention adopts a random scheduling strategy to allocate wireless channels with limited resources to proper multi-user equipment ends (UEs), wherein in each communication round, a wireless Access Point (AP) uniformly and randomly selects a plurality of related multi-user equipment ends (UEs) to update parameters, and allocates a special sub-channel to each selected UE to send training parameters.
Wherein, in the step of simulating the inter-section interference and the path loss in the wireless environment factors by adding errors to the local FL model parameters, the method comprises the following steps:
in each communication round, selecting K.C (C belongs to (0, 1)) clients from K clients to participate in training;
in K X client, testing the influence of error of model parameters of client selected from A X K X C, b X K X C and C X K X C on the identification precision of final polymerization model; wherein a, b, c is belonged to (0, 1) and different from each other; specifically, set as User _ error _ rate _ list [ a, b, c ]; a, b, c is ∈ (0, 1) and is different from each other
The error mode of the client model parameter comprises the following steps:
setting L layers of the model neural network layer number, selecting M layers to be in error, traversing all parameters of the M layers, and enabling each parameter to have the probability of Q to be in error, wherein M is less than L and Q belongs to (0, 1), specifically setting the error as layer _ error _ rate to be M/L, M is less than L, value _ error _ prob to be Q, and setting the error as Q belongs to (0, 1):
W＝W+randomuniform(-1,1)
wherein randomniform (-1, 1) randomly generates a floating point number from-1 to 1, and w represents the local FL model.
Specifically, for inter-zone interference, path loss and instability, a checking mechanism is used to check for data errors in the local FL model received at the wireless access point AP; assume each local FL model wkWill be sent as a single packet in the uplink, where C (w)k) 0 indicates that the local FL model received by the AP contains data errors; otherwise, we have C (w)k)＝1。
In the system, as long as the received local FL model contains errors, the wireless access point AP does not use the received data error local FL model to update the global FL model;
the wireless access point AP will update the global FL model directly using the remaining correct local FL model.
When C (w)k) When the local FL model parameter is 0, i.e. the local FL model parameter contains an error, the AP discards the local FL model containing an error and updates the global model with the remaining correct local FL model, otherwise, C (w)k) And 1, the AP updates the global FL model by the integrated local FL model and performs the next communication round of training. Thus, the global model update formula can be given by:
Figure BDA0002825675790000111
this formula simulates the process of the checking mechanism at the AP point discarding erroneous packets, i.e. the global average of the local FL model with the correct parameters. K denotes K clients, i.e. also user equipment clients in the present application.
Note that there is a constant alternation between communication and computation during the training phase. In this respect, the retransmission of a failed packet may not be beneficial because each uplink transmission of a local update will be followed by a globally averaged downlink transmission, and upon receiving this transmission the multi-user equipment UEs will refresh their reference parameters and start using local data to solve the new sub-problem. Therefore, when retransmission is requested in the presence of an error, random queuing delay and retransmission delay may be increased, thereby increasing learning delay and further affecting FL convergence performance.
The principle of the invention is as follows:
the FL system in the present invention optimizes the global model by repeating the following process: the method comprises the steps of firstly, performing local calculation by the multi-user equipment ends UEs by using local data thereof to minimize a predefined empirical risk function, namely a loss function, and updating training weights to the APs, secondly, collecting the updating by the APs from the multi-user equipment ends UEs and consulting FL units to generate an improved global model, thirdly, redistributing the output of the FL model to the multi-user equipment ends UEs, and finally, performing further local training by the multi-user equipment ends UEs by using the global model as a reference.
In the uploading stage of the local FL model, the bandwidth, the interval interference, the path loss and the instability of wireless environment factors are simulated, and finally the global aggregation model is influenced, so that the global model is distributed to the multi-User Equipment (UEs), and the influence is caused on the local learning of the multi-User Equipment (UEs). By adopting a scheduling strategy and a verification mechanism at the AP, the constructed Federal learning framework has robustness to wireless environment changes.
As shown in the figure, according to the simulation result, with the increase of the communication round, the loss function value in the training process is continuously reduced, the identification precision obtained by testing the model parameters after global aggregation is continuously increased, and it can be seen that the results all have certain fluctuation. This is a signal that the federal learning is proceeding normally, indicating that the global FL model derived from the local FL model is gradually fitting the data, i.e., the federal learning framework is learning, with ongoing parameter updates between the AP and the UEs.
As shown in fig. 2, after the influence of the wireless factor is simulated, in the communication round just started, the difference between the recognition accuracies obtained by testing the globally aggregated model by the loss function values in the training process under different packet error rates is very small, even better than that of the general federal learning algorithm, because the small change is equivalent to the role of adjusting the parameters of the weight matrix, so that the weight matrix has better performance. The performance of the federal learning with a large error rate gradually deteriorates or even fails along with the increase of communication rounds, and the more wrong parameters, the larger the fluctuation range of the curve, and the larger the influence on the performance of the aggregation model.
Specifically, a verification mechanism is adopted to discard two measures, namely a parameter packet with errors and a random scheduling strategy, in the model, the performance of the model after global aggregation is greatly improved, and a loss function value in the training process and the identification precision obtained by testing the model after global aggregation are converged according to a certain trend.
The larger the error rate is, the more the number of clients containing errors is, so the more local model parameters are updated, i.e. the fewer the number of local models participating in global aggregation, which does not increase retransmission delay or queuing delay. Therefore, the larger the error rate, the more the model convergence performance is affected, i.e. compared with the general federal learning algorithm, it needs more communication turns to achieve the same recognition accuracy, and secondly, it can be observed that the larger the error rate, the larger the fluctuation of the three curves is, which is caused by the instability of the global FL model due to the smaller number of local models participating in the global aggregation. The more these results and participating users, the better the learning effect tends to be.
Therefore, the method comprises the steps of firstly constructing a federal learning method performed by K clients; on the basis, the limitation and sharing property of the wireless medium are simulated by limiting the number of clients participating in the federal learning and adding a packet error in the local federal learning model updating uploading stage; and then, selecting a client for learning by adopting a scheduling strategy, discarding the parameter model with errors at the access point through a checking mechanism, and the like to obtain a Federal learning framework with robustness to wireless environment changes in the power Internet of things. The invention solves the problem that the Federal learning performance is rapidly deteriorated due to the resource limitation of the communication medium and the sharing property of the wireless medium in the wireless edge network, including the loss of the parameter update from the user equipment caused by the resource limitation of the communication medium; due to the sharing property of a wireless medium, communication receiving interval interference, average path loss of a communication link, large-scale fading and the like can cause the problems that parameter updating encounters inevitable faults and the like.
It should be noted that: the sequence of the embodiments in this specification is merely for description, and does not represent the advantages or disadvantages of the embodiments. And specific embodiments thereof have been described above. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing may also be possible or may be advantageous.
The embodiments in the present specification are described in a progressive manner, and the same and similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, for the embodiments of the apparatus, the device and the storage medium, since they are substantially similar to the method embodiments, the description is relatively simple, and the relevant points can be referred to the partial description of the method embodiments.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, but rather as the subject matter of the invention is to be construed in all aspects and as broadly as possible, and all changes, equivalents and modifications that fall within the true spirit and scope of the invention are therefore intended to be embraced therein.

Claims (9)

1. The federal learning method for robustness to wireless environment changes in the power internet of things is characterized by comprising the following steps:
the method comprises the following steps: building a neural network model for learning, initializing weight parameters of the neural network model, and distributing a data set to multiple User Equipment (UEs);
step two: carrying out federal learning through the built neural network model to obtain a trained neural network model;
step three: uploading the parameters of the trained neural network model to a wireless Access Point (AP), and simulating wireless environment factors to analyze the convergence performance of the global FL model at the moment;
step four: analyzing the convergence performance condition of the global FL model at the moment through simulating the wireless environment factors;
step five: and selecting a client executing the federated learning at the wireless access point AP by adopting a check mechanism and a scheduling strategy according to the acquired convergence performance condition to weaken the influence of the corresponding wireless environment factors on the convergence of the global model, and finally obtaining the federated learning method having robustness on the wireless environment change in the power Internet of things.
2. The federal learning method for robustness to wireless environment changes in the power internet of things as claimed in claim 1, wherein the building of the neural network model for learning in the first step comprises:
determining a type of neural network used;
determining the number of layers of the used neural network and the number of ganglion points of each layer;
the weights and bias parameters of all neural networks are initialized.
3. The Federal learning method for robustness against wireless environment changes in the electric Internet of things as claimed in claim 1,
the second step comprises the following steps:
determining parameters of a neural network model, training by using a data set based on the neural network model, and determining the parameters of the neural network model to comprise communication round times, client number K, client score C, local training generation E, local small batch B, test data small batch, learning rate lr and an optimization method;
distributing the original data set to K clients in a uniform distribution or non-uniform distribution mode;
the K clients learn by using local training data sets thereof until the total loss function reaches a minimum value, and upload the local parameter model of the neural network to the wireless access point AP through the wireless link;
after the wireless access point AP receives local model parameters from all the multi-user equipment ends UEs, the wireless access point AP starts to aggregate K local parameter models to obtain a global average model and transmits the model to all the multi-user equipment ends UEs, the multi-user equipment ends UEs start a new round of learning according to the updated local FL model, and the trained model is obtained through the learning of a plurality of communication rounds.
4. The Federal learning method for robustness against wireless environment changes in the Internet of things for electric power of claim 3, wherein in the step of distributing the original data set to K clients in a uniform or non-uniform manner,
the uniform and uniform distribution is to directly divide the original data set into K equal parts, and the non-uniform and uniform distribution requires that the original data set is sorted according to the size of the tag data and then distributed to K clients, namely the data set of each client is in non-uniform distribution.
5. The data synchronization method of claim 4, wherein the uploading the parameters of the trained neural network model to the wireless access point AP in the third step comprises:
in the training process of the communication turn, each user equipment firstly trains the local FL model w by using the training data of each user equipmentk；
Each user equipment transmits w through a wireless cellular linkkSending the data to the AP;
the AP receives the local FL models of all the participating user equipment ends, updates the global FL model and updates the global FL model wt+1Sending the data to all user equipment terminals to optimize a local FL model; the global FL model is obtained from all local FL models of all ue participating in learning, and can be represented by the following equation:
Figure FDA0002825675780000021
wherein, wt+1Is a global FL model generated by the wireless access point AP, nkRepresenting the size of a local data set of user equipment K, K representing a Kth client, t representing a ttth communication turn, wkRepresenting the local FL model.
6. The Federal learning method for robustness against wireless environment changes in the Internet of things for electric power of claim 1, wherein the convergence performance of the global FL model at this time is analyzed by simulating wireless environment factors in step four, wherein the wireless environment factors include bandwidth limitation, inter-cell interference and path loss,
by limiting the number of clients participating in federal learning to simulate bandwidth limitations in wireless environmental factors,
inter-zone interference and path loss in wireless environmental factors are simulated by adding errors to local FL model parameters.
7. The federal learning method for robustness against wireless environment changes in the power internet of things as claimed in claim 6, wherein the step of simulating inter-section interference and path loss in wireless environment factors by adding errors to local FL model parameters comprises:
in each communication round, selecting K.C (C belongs to (0, 1)) clients from K clients to participate in training;
in K x C clients, respectively testing the influence of the error of the model parameters of a x K x C clients, b x K x C clients and C x K x C clients on the identification precision of the final polymerization model; wherein a, b, c is belonged to (0, 1) and different from each other;
the error mode of the client model parameter comprises the following steps:
setting L layers of the model neural network layer, selecting M layers to make errors, traversing all parameters of the M layers, and enabling each parameter to have the probability of making errors of Q, wherein M is less than L, and Q is the error of (0, 1) and is set as:
W＝W+randomuniform(-1,1)
wherein randomniform (-1, 1) randomly generates a floating point number from-1 to 1, and w represents the local FL model.
8. The federal learning method for robustness against wireless environment changes in the power internet of things as claimed in claim 1, wherein the scheduling policy in the fifth step comprises:
allocating resource-limited radio channels to appropriate multi-user equipment UEs includes, in each communication round, randomly selecting multiple relevant multi-user equipment UEs uniformly by the wireless access point AP for parameter updating, and allocating a dedicated sub-channel to transmit training parameters for each selected UE.
9. The federal learning method for robustness against wireless environment changes in the power internet of things as claimed in claim 1, wherein the checking mechanism is used to check the data errors in the local FL model received at the wireless access point AP;
the wireless access point AP does not use the received data error local FL model to update the global FL model;
the wireless access point AP will update the global FL model directly using the remaining correct local FL model.
CN202011432782.8A 2020-12-09 2020-12-09 Federal learning method with robustness to wireless environment change in electric power Internet of things Active CN112637883B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011432782.8A CN112637883B (en) 2020-12-09 2020-12-09 Federal learning method with robustness to wireless environment change in electric power Internet of things

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011432782.8A CN112637883B (en) 2020-12-09 2020-12-09 Federal learning method with robustness to wireless environment change in electric power Internet of things

Publications (2)

Publication Number Publication Date
CN112637883A true CN112637883A (en) 2021-04-09
CN112637883B CN112637883B (en) 2023-04-28

Family

ID=75309394

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011432782.8A Active CN112637883B (en) 2020-12-09 2020-12-09 Federal learning method with robustness to wireless environment change in electric power Internet of things

Country Status (1)

Country Link
CN (1) CN112637883B (en)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111654027A (en) * 2020-06-15 2020-09-11 江苏方天电力技术有限公司 Power distribution Internet of things intelligent decision method based on reinforcement learning
CN113762533A (en) * 2021-09-07 2021-12-07 北京航空航天大学 Federal learning method with robustness to data pollution in wireless edge network
CN114554459A (en) * 2022-01-19 2022-05-27 苏州大学 Internet of vehicles federal learning client selection method assisted by near-end strategy optimization
WO2023096214A1 (en) * 2021-11-25 2023-06-01 엘지전자 주식회사 Method for performing federated learning in wireless communication system, and apparatus therefor
CN117135693A (en) * 2023-10-27 2023-11-28 四川长虹新网科技有限责任公司 Real-time service distribution method based on federal learning under multi-AP environment
CN117692940A (en) * 2024-02-04 2024-03-12 北京大业桥科技股份有限公司 Microwave system performance detection method based on microwave link
CN117692940B (en) * 2024-02-04 2024-04-26 北京大业桥科技股份有限公司 Microwave system performance detection method based on microwave link

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110493826A (en) * 2019-08-28 2019-11-22 重庆邮电大学 A kind of isomery cloud radio access network resources distribution method based on deeply study
CN111190487A (en) * 2019-12-30 2020-05-22 中国科学院计算技术研究所 Method for establishing data analysis model
US20200234110A1 (en) * 2019-01-22 2020-07-23 Adobe Inc. Generating trained neural networks with increased robustness against adversarial attacks
CN111754000A (en) * 2020-06-24 2020-10-09 清华大学 Quality-aware edge intelligent federal learning method and system
CN111860581A (en) * 2020-06-11 2020-10-30 南京邮电大学 Federal learning training method based on model dispersion

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200234110A1 (en) * 2019-01-22 2020-07-23 Adobe Inc. Generating trained neural networks with increased robustness against adversarial attacks
CN110493826A (en) * 2019-08-28 2019-11-22 重庆邮电大学 A kind of isomery cloud radio access network resources distribution method based on deeply study
CN111190487A (en) * 2019-12-30 2020-05-22 中国科学院计算技术研究所 Method for establishing data analysis model
CN111860581A (en) * 2020-06-11 2020-10-30 南京邮电大学 Federal learning training method based on model dispersion
CN111754000A (en) * 2020-06-24 2020-10-09 清华大学 Quality-aware edge intelligent federal learning method and system

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111654027A (en) * 2020-06-15 2020-09-11 江苏方天电力技术有限公司 Power distribution Internet of things intelligent decision method based on reinforcement learning
CN111654027B (en) * 2020-06-15 2022-04-29 江苏方天电力技术有限公司 Power distribution Internet of things intelligent decision method based on reinforcement learning
CN113762533A (en) * 2021-09-07 2021-12-07 北京航空航天大学 Federal learning method with robustness to data pollution in wireless edge network
CN113762533B (en) * 2021-09-07 2023-06-06 北京航空航天大学 Federal learning method with robustness to data pollution in wireless edge network
WO2023096214A1 (en) * 2021-11-25 2023-06-01 엘지전자 주식회사 Method for performing federated learning in wireless communication system, and apparatus therefor
CN114554459A (en) * 2022-01-19 2022-05-27 苏州大学 Internet of vehicles federal learning client selection method assisted by near-end strategy optimization
CN117135693A (en) * 2023-10-27 2023-11-28 四川长虹新网科技有限责任公司 Real-time service distribution method based on federal learning under multi-AP environment
CN117135693B (en) * 2023-10-27 2024-01-23 四川长虹新网科技有限责任公司 Real-time service distribution method based on federal learning under multi-AP environment
CN117692940A (en) * 2024-02-04 2024-03-12 北京大业桥科技股份有限公司 Microwave system performance detection method based on microwave link
CN117692940B (en) * 2024-02-04 2024-04-26 北京大业桥科技股份有限公司 Microwave system performance detection method based on microwave link

Also Published As

Publication number Publication date
CN112637883B (en) 2023-04-28

Similar Documents

Publication Publication Date Title
CN112637883B (en) Federal learning method with robustness to wireless environment change in electric power Internet of things
Lu et al. Differentially private asynchronous federated learning for mobile edge computing in urban informatics
Tang et al. Computational intelligence and deep learning for next-generation edge-enabled industrial IoT
Dhakal et al. Coded federated learning
Lin et al. Task offloading for wireless VR-enabled medical treatment with blockchain security using collective reinforcement learning
Drainakis et al. Federated vs. centralized machine learning under privacy-elastic users: A comparative analysis
CN111629380B (en) Dynamic resource allocation method for high concurrency multi-service industrial 5G network
Xie et al. Adaptive online decision method for initial congestion window in 5G mobile edge computing using deep reinforcement learning
Yoshida et al. MAB-based client selection for federated learning with uncertain resources in mobile networks
Wang et al. Content-based vehicle selection and resource allocation for federated learning in IoV
Cui et al. Optimal rate adaption in federated learning with compressed communications
Shi et al. Privacy-aware edge computing based on adaptive DNN partitioning
Yang et al. Learning the interference graph of a wireless network
Taïk et al. Data-quality based scheduling for federated edge learning
CN116471286A (en) Internet of things data sharing method based on block chain and federal learning
CN114301935A (en) Reputation-based method for selecting edge cloud collaborative federated learning nodes of Internet of things
Liu et al. Time minimization in hierarchical federated learning
US20240097993A1 (en) Machine Learning Model Distribution
Imani et al. iSample: Intelligent client sampling in federated learning
Guo et al. Feat: A federated approach for privacy-preserving network traffic classification in heterogeneous environments
Salama et al. Flcc: Efficient distributed federated learning on iomt over csma/ca
CN113691594A (en) Method for solving data imbalance problem in federal learning based on second derivative
Liu et al. Federated Learning and Meta Learning: Approaches, Applications, and Directions
Lin et al. Heuristic-learning-based network architecture for device-to-device user access control
Ma et al. Paving the way for distributed artificial intelligence over the air

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
GR01 Patent grant
GR01 Patent grant