US-12628224-B2 - Methods for improved federated machine learning in wireless networks

US12628224B2US 12628224 B2US12628224 B2US 12628224B2US-12628224-B2

Abstract

A computer implemented method for federated machine learning, FL, in a wireless communication system, the method comprising establishing a first wireless access radio link between a first access node and a wireless device, initiating an FL process involving the first access node and the wireless device, transmitting FL information from the first access node to the wireless device, where the FL information comprises data indicative of the FL process, establishing a second wireless access radio link between a second access node and the wireless device, where the second access node is communicatively coupled to the first access node, exchanging at least part of the FL information between the wireless device and the second access node, and resuming the FL process involving the first access node and the wireless device by communication via the second access node over the second wireless access radio link.

Inventors

Henrik Rydén

Assignees

TELEFONAKTIEBOLAGET LM ERICSSON (PUBL)

Dates

Publication Date: 20260512
Application Date: 20210325

Claims (20)

1 . A computer implemented method for federated machine learning, FL, in a wireless communication system, the method comprising: establishing a first wireless access radio link between a first access node and a wireless device, initiating an FL process involving the first access node and the wireless device, transmitting FL information from the first access node to the wireless device, where the FL information comprises data indicative of the FL process, establishing a second wireless access radio link between a second access node and the wireless device, where the second access node is communicatively coupled to the first access node, exchanging at least part of the FL information between the wireless device and the second access node, and resuming the FL process involving the first access node and the wireless device by communication via the second access node over the second wireless access radio link.
2 . The method according to claim 1 , where the wireless communication system is a system operating according to a standard defined by the third generation partnership program, 3GPP.
3 . The method according to claim 1 , where the initiated FL process comprises training a machine learning function based on channel state information, CSI, data obtained locally by the wireless device.
4 . The method according to claim 1 , where the initiated FL process comprises training a machine learning function based on radiolocation(s) and/or geolocation data obtained locally by the wireless device.
5 . The method according to claim 1 , performed in response to an anticipated interruption of the initiated FL process.
6 . The method according to claim 5 , wherein the anticipated interruption of the FL process comprises the wireless device entering into an idle mode of operation.
7 . The method according to claim 5 , wherein the anticipated interruption of the FL process comprises the wireless device entering into a handover operation.
8 . The method according to claim 5 , wherein the anticipated interruption of the FL process comprises the wireless device entering into a recovery mode of operation due to declaring radio-link failure.
9 . The method according to claim 1 , comprising transmitting FL information comprising any of: an FL process identifier, a priority metric indicating a priority of the FL process relative to one or more other FL processes in the wireless communication system, an FL process training round indicator value, a use-case description associated with the FL process, a set of global FL model weights associated with the FL process, an instruction to store a dataset of the FL process, or a part of a dataset of the FL process, an instruction indicating when stored FL process data can be deleted by the wireless device, an instruction indicating a set of access nodes in the wireless communication system with which FL information can be shared, and/or an instruction associated with encryption of the model parameters when communicating with the second node.
10 . The method according to claim 1 , comprising connecting the second access node to the first access node via a core network.
11 . The method according to claim 1 , comprising the second access node requesting the wireless device to provide the FL information.
12 . The method according to claim 1 , comprising transmitting information from the wireless device to the second access node indicative of a set of access nodes where the wireless device has participated in an FL process.
13 . The method according to claim 1 , comprising determining a throughput metric for communication between the wireless device and the first access node via the second access node prior to resuming the FL process between the first access node and the wireless device.
14 . The method according to claim 1 , comprising determining a relevance metric associated with a local data set of the wireless device for the FL process prior to resuming the FL process between the first access node and the wireless device.
15 . The method according to claim 1 , comprising determining an energy storage state of the wireless device prior to resuming the FL process between the first access node and the wireless device.
16 . The method according to claim 1 , comprising transmitting an instruction to the wireless device to delete a local data set related to the FL process.
17 . A computer implemented method for federated machine learning, FL, performed in a first access node, the method comprising: establishing a first wireless access radio link between the first access node and a wireless device, initiating an FL process involving the first access node and the wireless device, and transmitting FL information from the first access node to the wireless device, where the FL information comprises data indicative of the FL process.
18 . The method according to claim 17 , where the initiated FL process comprises training a machine learning function based on CSI data obtained locally by the wireless device.
19 . A computer implemented method for federated machine learning, FL, performed in a second access node, where the second access node is communicatively coupled to a first access node, the method comprising: establishing a second wireless access radio link between the second access node and a wireless device, exchanging FL information related to an initiated FL process between the wireless device and the second access node, where the FL information has previously been communicated between the first access node and the wireless device, and facilitating resumption of an FL process involving the first access node and the wireless device by forwarding communication between the wireless device and the first access node via the second access node over the second wireless access radio link.
20 . A computer implemented method for federated machine learning, FL, performed in a wireless device, the method comprising: establishing a first wireless access radio link between a first access node and the wireless device, initiating an FL process involving the first access node and the wireless device, receiving FL information from the first access node, where the FL information comprises data indicative of the FL process, establishing a second wireless access radio link between a second access node and the wireless device, where the second access node is communicatively coupled to the first access node, exchanging at least part of the FL information between the wireless device and the second access node, and resuming the FL process involving the first access node and the wireless device by communication via the second access node over the second wireless access radio link.

Description

TECHNICAL FIELD The present disclosure relates to machine learning in wireless communication networks comprising a plurality of access points and one or more wireless devices which are serviced from the access points over wireless radio access links. There are disclosed methods, network nodes, wireless devices, computer programs and computer program products for, e.g., data collection and data analysis based on federated machine learning methods. BACKGROUND Machine learning relates generally to techniques where a model having a certain pre-determined structure is adapted to provide a desired function by means of some form of training mechanism. Machine learning techniques have gained tremendous interest over recent years due to the versatility and adaptability of the methods in different types of applications. Some example machine learning structures comprise neural networks, autoencoder networks, and random forest structures. For instance, WO2017/162262 A1 discusses the use of machine learning methods for predicting radio performance on different radio carriers. Federated learning (FL), also known as collaborative learning, is a machine learning technique in which a machine learning function is trained across multiple decentralized edge devices which have access to respective local data samples, without the edge devices exchanging those data samples with a central server or with other edge devices. This approach stands in contrast to traditional centralized machine learning techniques where all the local datasets are uploaded to one server for central processing. The approach is also different from decentralized machine learning techniques where data samples are exchanged among edge devices. Since FL enables multiple actors to build a common and robust machine learning model without sharing data, critical issues such as data privacy, data security, data access rights and access to heterogeneous data are inherently addressed, which is an advantage. The training of FL models often requires multiple training rounds, where each training round involves signaling to and from the edge devices. When wireless devices in a wireless access network are participating in the FL process, there are several issues that can occur which have an adverse effect on performance. For example, a wireless device may go into idle mode, and thus not be available for all rounds in the FL process, which can cause undesired interruptions in the FL process and thus introduce latencies in the model training. There is a need for FL techniques specifically adapted for use in wireless communication networks comprising mobile wireless devices. There is also a need for improving robustness and data integrity associated with FL processes in wireless access network. SUMMARY It is an object of the present disclosure to provide methods for federated learning which alleviates or even resolves some of the above-mentioned issues. This object is obtained by a computer implemented method for federated machine learning (FL) in a wireless communication system. The method comprises establishing a first wireless access radio link between a first access node and a wireless device, initiating an FL process involving the first access node and the wireless device and transmitting FL information from the first access node to the wireless device, where the FL information comprises data indicative of the FL process. The method also comprises establishing a second wireless access radio link between a second access node and the wireless device, where the second access node is communicatively coupled to the first access node, exchanging at least part of the FL information between the wireless device and the second access node, followed by resuming the FL process involving the first access node and the wireless device by communication via the second access node over the second wireless access radio link. This way an FL process may be resumed via relay over a second access node. This is an advantage, since the first access node can update its global FL model using the data available locally at the wireless device, which otherwise would have been out of reach from the first access node. The accuracy of the FL model is thus improved. The wireless communication system is optionally a system operating according to a standard defined by the third generation partnership program (3GPP). According to aspects, the initiated FL process comprises training a machine learning function based on channel state information (CSI) data obtained locally by the wireless device 150. Machine learning processes involving CSI data are advantageously implemented using FL techniques, since there is no need to communicate the large size CSI data sets. Also, privacy issues are more easily handles in FL processes since the CSI information is not communicated from the wireless device, which is an advantage. According to aspects, the initiated FL process comprises training a machine learning functio