Machine learning based control of data communications in wireless communication system

Abstract

The present subject matter relates to an apparatus for a wireless communication system. The apparatus includes a machine learning model configured to receive an input and provide an output that enables control of data communication in a wireless communication system by a node. The apparatus comprises means configured to: receive, from a node, a request to send data, the data comprising at least one of: an output of a trained machine learning model or an indication of a detected event that may affect control of data communication; inferring the machine learning model using the current value of the input to obtain a current output; determining whether an event that affects the output of the machine learning model occurs; data is transmitted to the node based on the occurrence of the event.

Description

Technical Field

[0001] Various example embodiments relate to telecommunication systems, and more particularly to machine learning based control of data communications in wireless communication systems. Background Art

[0002] The fifth generation of wireless networks (5G) refers to a new generation of radio systems and network architectures. It is expected that 5G will provide higher bit rates and coverage than the current long term evolution (LTE) system. It is expected that 5G will also be able to increase network scalability to hundreds of thousands of connections. However, there is a need to improve the communication services provided at these systems, and in particular to improve the resource status process. For example, some aspects of the resource status process with periodic reporting of load information between adjacent nodes are described in standards such as TS 38.423, TS 38.473, and TS 36.423. Summary of the invention

[0003] Example embodiments provide an apparatus for a wireless communication system. The apparatus includes a machine learning model. The machine learning model is configured to receive input and provide output. The output enables control of data communication in the wireless communication system by a node. The apparatus includes a component configured for the following items: receiving a request to send data from a node, the data including at least one of the following: an output of the machine learning model or an indication of an event that may affect control of data communication; using the current value of the input to reason about the machine learning model to obtain the current output; determining whether an event that affects the output of the machine learning model occurs; sending data to the node based on the occurrence of the event.

[0004] Example embodiments provide a method for using a machine learning model. The machine learning model is configured to receive an input and provide an output. The output enables control of data communications in a wireless communication system by a node. The method includes: receiving a request to send data from a node, the data including at least one of the following: an output of a trained machine learning model or an indication of an event that may affect control of data communications; using the current value of the input to reason about the machine learning model to obtain the current output; determining whether an event that affects the output of the machine learning model occurs; and sending data to the node based on the occurrence of the event.

[0005] An example embodiment provides a computer program comprising instructions for causing a device to perform at least the following: receiving a request to send data from a node, the data comprising at least one of: an output of a trained machine learning model or an indication of an event that may affect control of data communications; reasoning about the machine learning model using current values ​​of inputs to the machine learning model to obtain a current output; determining whether an event that affects the output of the machine learning model has occurred; and sending data to the node based on a detected event.

[0006] Example embodiments provide a node for controlling data communications in a wireless communication system. The node includes components configured to: configure a device of the wireless communication system to send data, the data including at least one of: an output of a trained machine learning model or an indication of a detected event that may affect control of data communications; receive requested data from the device; and perform control of data communications in the wireless communication system based on the received data.

[0007] Example embodiments provide a method for controlling data communications in a wireless communication system. The method includes: configuring a device of the wireless communication system to send data, the data including at least one of: an output of a trained machine learning model or an indication of a detected event that may affect control of data communications; receiving requested data from the device; and performing control of data communications in the wireless communication system based on the received data.

[0008] An example embodiment provides a computer program comprising instructions for causing a node to perform at least the following: configuring a device of a wireless communication system to send data, the data comprising at least one of the following: an output of a trained machine learning model or an indication of a detected event that may affect control of data communications; receiving requested data from the device; and performing control of data communications in the wireless communication system based on the received data. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The accompanying drawings are included to provide a further understanding of the examples, and are incorporated in and constitute a part of this specification. In the drawings:

[0010] Figure 1 illustrates a portion of an exemplary radio access network;

[0011] Figure 2 is a schematic diagram of a wireless communication system;

[0012] Figure 3 is a flow chart of a method used in an apparatus according to an example of the present subject matter;

[0013] Figure 4 is a flow chart of a method used in an apparatus according to an example of the present subject matter;

[0014] Figure 5 is a flowchart of a method used in a node according to an example of the present subject matter;

[0015] Figure 6 is a flowchart of a method used in a node according to an example of the present subject matter;

[0016] Figure 7is a flow chart of a signaling method for controlling data communication according to an example of the present subject matter;

[0017] Figure 8 is a flow chart of a signaling method for controlling data communication according to an example of the present subject matter;

[0018] Fig. 9 is a flow chart of a signaling method for controlling data communication according to an example of the present subject matter;

[0019] Fig.10 is a flow chart of a signaling method for controlling data communication according to an example of the present subject matter;

[0020] Fig.11 is a functional framework for model training and inference according to the examples of this topic;

[0021] Fig.12 is a block diagram illustrating an example of an apparatus according to examples of the present subject matter. DETAILED DESCRIPTION

[0022] In the following description, for the purpose of explanation rather than limitation, specific details are set forth, such as specific architectures, interfaces, techniques, etc., in order to provide a thorough understanding of the examples. However, those skilled in the art will appreciate that the disclosed subject matter can be practiced in other illustrative examples that deviate from these specific details. In some instances, detailed descriptions of well-known devices and / or methods are omitted to avoid obscuring the description with unnecessary details.

[0023] A communication system may be provided. The communication system includes nodes such as base stations, each of which may serve user equipment (UE) located within a service geographic area or cell of the node. The communication system may support one or more radio access technologies (RATs). For example, the radio access technology in the radio access technology may be Evolved Universal Terrestrial Radio Access (E-UTRA) or 5G New Radio (NR), but is not limited thereto, as those skilled in the art may apply the subject matter to other communication systems having the necessary attributes.

[0024] The communication system enables data communication between elements of the communication system. The data may include text data, image data, audio data, video data, and other types of data. Data communication may be random, time-varying, and unbalanced. This may make uneven load scenarios for adjacent cells. For example, one cell may be overloaded, while another cell may have far fewer users and its resources may not be fully utilized. This can be solved by controlling data communication. Different types of control of data communication can be provided. For example, data communication in a communication system can be controlled using means such as load balancing, mobility management, and energy saving. The present subject may use a machine learning (ML) model (referred to herein as ML1) to achieve improved and accurate control of data communication in a communication system. The present subject may enable event-based reporting between two nodes. For example, this may provide improvements in resource status processes defined on a network interface between gNBs or between a gNB-CU and a gNB-DU.

[0025] The machine learning model ML1 may be a trained machine learning model. For example, the machine learning model ML1 may be obtained by training a machine learning algorithm. The term "machine learning algorithm" may refer to a computer algorithm that is used to extract useful information from a training data set by automatically constructing a probabilistic model (called a machine learning model or "prediction model"). The machine learning algorithm constructs a mathematical model based on sample data (called "training data") so as to make predictions or decisions without being explicitly programmed to perform a task. Machine learning may be performed using learning algorithms such as supervised or unsupervised learning, [clustering, classification, linear regression] reinforcement algorithms, self-learning, etc. Machine learning may be based on various techniques such as clustering, classification, linear regression, support vector machines, neural networks, etc. For example, a "model" or "prediction model" may be a data structure or program such as a neural network, a support vector machine, a decision tree, a Bayesian network, etc. The model is suitable for predicting unmeasured values ​​(e.g., which label corresponds to a given token) and / or predicting or selecting an action to maximize future rewards based on other known values. Note that in some examples, maximizing rewards may be equivalent to minimizing future costs.

[0026] The machine learning model ML1 may be configured to receive input and provide output in response to receiving the input. The output of the machine learning model may also be referred to as artificial intelligence (AI) / ML model inference output.

[0027] The input of the machine learning model can be provided by one or more elements of the communication system. For example, the input can be provided by the UE and / or the node and / or the neighboring node running the AI / ML model. The input can be provided according to the control type of the data communication. The input can enable the execution of at least one of the following: performing mobility management, control switching, load balancing or network energy saving in the communication system, but is not limited thereto, because those skilled in the art can use this subject for other types of control. In one example, the input can describe the location of the user equipment. The input can indicate the location coordinates, the serving cell identifier (ID) of the cell serving the UE, and the moving speed of the UE. The input can also indicate measurements, such as the cell level of the cell serving the UE, the beam level in the cell, the reference signal received power (RSRP) measurement, the reference signal received quality (RSRQ) measurement, and the signal to interference and noise ratio (SINR) measurement. The input can also indicate the resources of the node adjacent to the first node, such as the current or predicted resource state corresponding to the load information at the neighboring node of the first node, and the current or expected energy efficiency at the node. In one example, the input can be evaluated by the corresponding element of the communication system through prediction, for example, using another ML model that can predict the input. Depending on the type of control of the data communication, the machine learning model ML1 can receive corresponding inputs to provide outputs that can be used to perform this type of control of the data communication in the communication system.

[0028] The output of the machine learning model ML1 may, for example, include predictions of trajectories of one or more user devices, predictions of cell loads at one or more cells served by a node of the communication system, predictions of target cells for switching, predictions of traffic of user equipment, predictions of target cells for load balancing, predictions of resource states corresponding to predicted load information of nodes of the communication system, predictions of user equipment to be switched to a target node of the communication system, predictions of energy saving strategies (such as predictions of cell activation or deactivation recommendations), predicted energy efficiency, predicted energy states, and predicted model output validity time.

[0029] The values ​​of the inputs used to train the machine learning model ML1 may have a defined pattern or distribution, such as along time or location. The pattern of the input may refer to a geometric property of the distribution of the values ​​of the input. For example, the values ​​of the inputs used during training may have a corresponding distribution or pattern. For example, the pattern of the input may indicate a range of allowed values ​​of the input and / or a maximum / minimum value. In order to provide reliable reasoning of the machine learning model ML1, the values ​​of the inputs used during reasoning may have to match or be consistent with the corresponding pattern. For example, during training, the load may vary between a given minimum load value and a maximum load value. Therefore, reliable reasoning of the machine learning model ML1 may require similar load conditions. However, due to the occurrence of an event, the pattern of the input may be affected or changed. For example, an event may be the occurrence of a failure. For example, an event may be that a node of a communication system has unloaded a set of user devices so that they can be served by the device. An event may be that network conditions have changed, making the machine learning model ML1 unsuitable for the environment in which it is running. If the event occurs before reasoning the machine learning model ML1, the reasoning of the machine learning model ML1 may be considered invalid (e.g., no longer accurate enough), and the machine learning model ML1 may require retraining. The present subject matter can solve this problem as described herein.

[0030] A given node of a communication system may be provided, referred to herein as a first node. The first node may use the output of the machine learning model ML1 to control data communication in the communication system. To this end, the machine learning model ML1 may be deployed in a device of the communication system. The machine learning model ML1 may be trained in the device or in an external system. For example, the external system may be an element of the communication system. The external system may be, for example, a next generation Node B (gNB), a gNB centralized unit (gNB-CU), a gNB distributed unit (gNB-DU), a UE, or an operation administration and maintenance (OAM) unit. For example, a UE may have sufficient processing resources to perform training.

[0031] For example, the device may be a second node of a radio access network of a communication system, a core network node, or a user equipment. The device may be configured to perform reasoning of a machine learning model ML1 and / or retraining of a machine learning model ML1. For example, the device may use a model reasoning function to perform reasoning of a machine learning model ML1. To this end, the device may require input of the machine learning model ML1. The input may be provided by one or more elements of the communication system (referred to herein as training elements). For example, the one or more training elements providing the input may include one or more nodes of a communication system and / or device, and / or one or more user equipment. For example, each of the one or more training elements may be configured to provide a value for a corresponding portion of the input. The value of the input may be determined by measurements and / or calculations and / or predictions performed by the ML model at the training element. Having several training elements that can provide input may be advantageous because the prediction of the machine learning model ML1 can take into account several contributors to the communication system, and thus reliable predictions can be provided.

[0032] The first node may configure the device to send data including model reasoning output information. The first node may configure the device so that the device provides model reasoning output information with specific content. For example, the first node may configure the device so that the device provides model reasoning output information, which includes the current output of the machine learning model ML1 and / or an indication of an event that may affect the control of the first node over data communication. The indication of the event may, for example, include a description of the event and / or a description of the output, such as a description of how the output is affected. For example, the model reasoning output information may be a tuple or a data structure including one or two entries. The first of the two entries may include the current output of the machine learning model ML1, and the other second entry may include a variable, the value of which provides a prompt or indication of the detected event. For example, a variable may include a description of an event that may affect the control of the first node over data communication, for example, if there is an event of type "Y" in the data structure, it indicates that the event is affecting the output of the machine learning model ML1. The data structure may include a first entry and / or a second entry. For example, if the detected event has an impact on the control of data communication, the data structure may only include the second entry. If the detected event has no effect on the control of data communications, the data structure may include only the first entry. If the device cannot determine whether the detected event has an effect on the control of data communications, the data structure may include the first entry and the second entry. If the event is not detected, the data structure may include only the first entry.

[0033] In response to being configured by the first node, the device may determine the current value of the input of the machine learning model ML1. For example, the device may request one or more training elements to provide the current value of its corresponding input. The device may receive all values ​​of the input from one or more training elements. Alternatively, the device may be configured to automatically receive the current value of the input (i.e., no request for the input may be required). The device may provide the current value of the input to the machine learning model ML1 in order to perform reasoning of the machine learning model ML1. In response, the machine learning model ML1 may provide the current output of the machine learning model ML1. The device may be configured to check whether an event affecting the output has occurred. The device may detect an event, i.e., determine that the event has occurred. The current output of the machine learning model ML1 may depend on the detected event. For example, the detected event may cause a change in the output quality of the machine learning model ML1. The device may be configured to detect a change in the output quality after the event is detected. The output quality may include the validity time of the output and / or the accuracy of the output and / or the maximum difference between the input and the threshold. If the event detected by the device causes a change in the output quality of the machine learning model ML1, it may indicate that the event may also affect the control of the data communication by the first node. If the event does not cause a change in the output quality of the machine learning model ML1, it can be indicated that the event will not affect the first node's control of data communication. Therefore, based on the occurrence of the event, the device can prepare model reasoning output information, which has an indication of the current output and / or detected event as content.

[0034] For example, if the detected event has no effect on the output quality of the machine learning model ML1, that is, the output is not affected after the event is detected, the model reasoning output information may include only the current output. If the detected event has an effect on the output quality of the machine learning model ML1, that is, the output is affected after the event is detected, the model reasoning output information may include the current output and an indication of the above event. If the event has not been detected (and the reporting cycle is reached), the model reasoning output information may include only the current output. If the output is affected, it may indicate that the control of data communication may be affected. For example, if the device determines that the output is affected according to the defined conditions or rules provided by the first node after the event is detected, the control of data exchange will also be affected. In another example, if the device determines that the output is affected according to its own inspection after the event is detected, the device may assume that the first node's control of data communication will also be affected, or it may send both the output and the indication of the output quality change / event to the first node so that the first node can decide whether the control of data communication will be affected. Therefore, if the event is not detected, the device may send the model reasoning output information to the first node according to the reporting period, and if the event is detected, the device may send the model reasoning output information to the first node immediately.

[0035] In one example, the device detects an event but is unable to determine whether the output of the machine learning model ML1 affects the control of data communications by the first node, for example, this may be the case if the device does not receive a list of influencing conditions from the first node and / or the device is unable to determine whether the output is affected. In this case, the device may determine that the detected event affects the output and data control at the first node, and send model reasoning output information including both the current output and the event indication to the first node. This may provide a secure method for the first node to control data communications.

[0036] In response to receiving the model reasoning output information, the first node may use the current output to perform control of data communication in the communication system. Alternatively, if the affected output quality may have an impact on data communication control, the first node may request the device to stop sending the model reasoning output information.

[0037] For example, the output of the machine learning model ML1 may be a predicted load at one or more network nodes. The first node may use the received predicted load and its own load information to predict the load in its cell. Based on this, the first node may predict that it will be overloaded in the future. Therefore, the first node may perform control of data communication in the communication system by actively requesting neighboring nodes of the communication system to activate capacity cells.

[0038] In another example, the output of the machine learning model ML1 may be resource status information from one or more network nodes. The first node may use information from neighboring nodes and local information to calculate the predicted load in its cell and neighboring cells. In particular, the first node may identify the node with the smallest expected load as the node for offloading the traffic. Thus, the first node may perform control of data communications in the communication system by performing a load balancing decision, according to which the traffic is offloaded to the identified node.

[0039] In another example, the output of the machine learning model ML1 may be a predicted UE trajectory. The first node may use the output to determine that the UE will follow one or more paths with a high probability. Thus, the first node may perform control of data communications in the communication system by configuring a conditional handover configuration on one or more most likely paths of the expected handover.

[0040] In one example, the device may provide the requested model reasoning output information as follows. Upon detecting an event, the device may check whether the output quality of the current output is affected. In response to determining that the output quality of the current output is affected, the device may send model reasoning output information, the content of which indicates that the event may affect the control of data communication, and further indicates the current output. The device may send the model reasoning output information together with the current output and the above indication, because if the first node has no influence on its control decision, the first node can still use the affected output to perform control of data communication in the communication system.

[0041] The device can check or determine whether the output quality of the current output is affected or changed or affected by, for example, the execution of the first check. The first check includes comparing the current output with one or more reference output values ​​of the machine learning model ML1, or comparing the output quality with the reference output quality. If the current output differs from the reference value by at least one offset value, it can be indicated that the current output is affected. The first check can provide a simple and accurate assessment of the output of the machine learning model ML1. Alternatively, the device can perform a second check. The second check includes determining whether the current value of the input matches the (multiple) patterns defined during training. In the case where the current value of the input does not match the pattern, the device can determine that the output quality of the machine learning model is affected, otherwise it can be determined that the output quality is not affected. For example, for each input "X" in the input, the device can identify a pattern related to the input "X", and if it finds a pattern, it can compare the value of the above input "X" with the above pattern to determine whether the value of the input "X" matches the relevant pattern. If the value of at least a portion of the input does not match the corresponding pattern, it can be indicated that the output quality of the machine learning model is affected. In one example, if at least one of the first check and the second check indicates that the quality of the current output is affected, the apparatus may determine that the quality of the current output is affected.

[0042] In one example, the first node may configure the device to repeatedly send model reasoning output information, for example according to the reporting cycle time described herein, until an event is detected. For each repetition, the device re-evaluates the input to obtain the current value of the input, and uses the updated input value to infer the machine learning model ML1. This can achieve reliable and accurate use of the machine learning model ML1. This can improve the control of data communications in the communication system. However, after the event is detected, the first node may check the event indicated in the received model reasoning output information, and if the event may affect its decision to control data communications in the communication system, the first node may send a stop request to the device to stop the repeated submission of the model reasoning output information. Therefore, the device can stop sending the model reasoning output information. If the detected event does not trigger a stop request from the first node, the device can also be configured to repeatedly send the model reasoning output information according to the reporting cycle until the event is detected again, and so on.

[0043] In order to configure the device to send model reasoning output information, the first node may, for example, send a request named measurement request to the device. The measurement request may indicate an element of the communication system whose input can be evaluated. The elements of the communication system may, for example, include one or more cells and one or more user equipments. This may be advantageous because the first node may control the inputs used to reason about the machine learning model ML1. This may further optimize the control of data communication in the communication system.

[0044] In one example, the measurement request may also indicate a condition, referred to herein as an impact condition. Alternatively, the device may access the impact condition in a shared database between nodes of the communication system so that the first node does not have to include the impact condition in the request. This may save resources in the case where the first node has to configure multiple devices of the communication system.

[0045] After receiving the measurement request and detecting the event, the device may determine whether the current output of the machine learning model ML1 satisfies the impact condition. If at least one of the impact conditions is satisfied, the current output of the machine learning model ML1 satisfies the impact condition. In response to determining that the current output satisfies the impact condition, the device may send model reasoning output information, the content of which only includes an indication that the impact condition is satisfied. This can save resources required to further send the current output, although it may not be used by the first node.

[0046] For example, the influencing condition may require that the validity time of the output is reduced by a specific value compared to the reference validity time, and / or the accuracy of the output is reduced by a specific value compared to the reference accuracy, and / or the input exceeds a certain threshold. The validity time may be the time during which the output can be reliably used to control data communications in the communication system. For example, the load forecast may be valid only for a specific time period and may need to be updated so that it can reliably reflect the status of the communication system.

[0047] In one example, the measurement request may also indicate a report configuration, according to which the device may perform the submission of model reasoning output information. For example, the report configuration may include report configuration parameters, such as a report cycle time. The report may be periodic, with a period higher than a minimum threshold, wherein when the output of the machine learning model ML1 satisfies the impact condition, an event may be reported in addition. The present subject matter may advantageously use the report configuration to improve the efficiency of the use of the machine learning model ML1. In particular, the report configuration parameters may be associated with one or more impact conditions. This correlation may enable the device to adapt the report configuration parameters to offset the impact of the output of the machine learning model ML1 on the decision of the first node. For example, if the impact condition requires that the validity time of the output is reduced by a specific value compared to the reference validity time and the validity time of the current output is reduced by this value (ie, the impact condition is met), the report cycle time may be changed so that it is less than or equal to the validity time of the current output. The device may continue to check for the occurrence of an event. If an event occurs, the device may immediately send the model reasoning output information. If the event is not detected and the event that triggers the stop request has not been previously detected, the device may periodically send the model reasoning output information according to the current / updated report cycle time.

[0048] In response to receiving the model reasoning output information, the first node may use the model reasoning output information to determine or decide whether to use the current output to control data communication or to request the device to stop sending the model reasoning output information. For example, if the model reasoning output information includes an indication of an event that may affect the control of data communication and does not include the output of the machine learning model ML1, the first node may send a stop request to the device to stop sending the model reasoning output information. The device may retrain the machine learning model ML1 in response to receiving the stop request. Therefore, based on the decision, the device may retrain the machine learning model ML1 or continue to reason about the machine learning model ML1. The output may be sufficiently accurate and may therefore be advantageously used for the first node's control of data communication in the communication system.

[0049] In one example, the first node may include another machine learning model named ML2, which is configured to receive as input the output of the machine learning model ML1 and provide as output an estimate or prediction of a control decision or action that the first node can take to control data communication in the communication system.

[0050] In one example, the first node may configure the device, but may also configure one or more other devices. The one or more other devices may each have their own trained ML model that provides corresponding outputs. The first node may use all outputs from the device to control data communications of the communication system. The first node may configure each device as described above to provide corresponding model reasoning output information, and use the model reasoning output information to control data communications of the communication system.

[0051] Figure 1 An example of a simplified system architecture is depicted, which shows only some elements and functional entities, which are all logical units, the implementation of which may differ from what is shown. Figure 1 The connections shown are logical connections; the actual physical connections may be different. It is clear to those skilled in the art that the system may generally include in addition to Figure 1 Other functions and structures than those shown.

[0052] However, the embodiments are not limited to the systems given as examples, but a person skilled in the art may apply the solution to other communication systems having the necessary characteristics.

[0053] Figure 1 The example of FIG. 1 shows a portion of an exemplary radio access network.

[0054] Figure 1 Devices 110 and 112 are shown. For example, devices 110 and 112 can be user equipment. Devices 110 and 112 are configured to be wirelessly connected to node 114 on one or more communication channels. Node 114 is also connected to core network 120. In one example, node 114 can be an access node (such as (e / g)NodeB) 114 that provides or serves devices in a cell. In one example, node 114 can be a non-3GPP access node. The physical link from the device to the (e / g)NodeB is called an uplink or reverse link, and the physical link from the (e / g)NodeB to the device is called a downlink or forward link. It should be understood that the (e / g)NodeB or its functions can be implemented by using any node, host, server or access point entity suitable for such a purpose.

[0055] The communication system usually includes more than one (e / g)NodeB, in which case the (e / g)NodeB may also be configured to communicate with each other via wired or wireless links designed for this purpose. These links may be used for signaling purposes. (e / g)NodeB is a computing device configured to control the radio resources of the communication system to which it is coupled. NodeB may also be referred to as a base station, an access point, or any other type of interface device, including a relay station capable of operating in a wireless environment. (e / g)NodeB includes or is coupled to a transceiver. From the transceiver of the (e / g)NodeB, a connection is provided to an antenna unit, which establishes a bidirectional radio link to the device. The antenna unit may include multiple antennas or antenna elements. (e / g)NodeB is further connected to a core network 20 (CN or next generation core NGC). For example, (e / g)NodeB may be connected to an access and mobility management function (AMF) and a user plane function (UPF) in the control plane and the user plane, respectively. Depending on the system, the counterpart on the CN side may be a serving gateway (S-GW, routing and forwarding user data packets), a packet data network gateway (P-GW, used to provide a connection between the device (UE) and an external packet data network), a mobility management entity (MME), etc.

[0056] The device (also referred to as user equipment, UE, user equipment, user terminal, terminal device, etc.) illustrates a type of device to which resources on the air interface are allocated and assigned, and therefore any features described herein with respect to the device can be implemented with a corresponding device (such as a relay node). An example of such a relay node is a layer 3 relay (self-backhaul relay) toward a base station.

[0057] The device generally refers to a device (e.g., a portable or non-portable computing device) including a wireless mobile communication device with or without a subscriber identity module (SIM), including but not limited to the following types of devices: mobile stations (mobile phones), smart phones, personal digital assistants (PDAs), handsets, devices using wireless modems (alarm or measurement devices, etc.), portable computers and / or touch screen computers, tablet computers, game consoles, laptops, and multimedia devices. It should be understood that the device can also be an almost exclusive uplink-only device, an example of which is a camera or video camera that loads images or video clips to a network. The device can also be a device with the ability to operate in an Internet of Things (IoT) network, in which an object can have the ability to transmit data over a network without requiring human-to-human or human-to-computer interaction, such as for smart grids and connected vehicles. The device can also use the cloud. In some applications, the device can include a user-portable device (such as a watch, headset, or glasses) with a radio component, and the calculation is performed in the cloud. The user device (or in some embodiments, a layer 3 relay node) is configured to perform one or more of the user device functions. The apparatus may also be called a subscriber unit, a mobile station, a remote terminal, an access terminal, a user terminal, or user equipment (UE), to name just a few names or devices.

[0058] The various techniques described in this article can also be applied to cyber-physical systems (CPS) (systems of collaborative computing elements that control physical entities). CPS can enable the implementation and development of a large number of interconnected ICT devices (sensors, actuators, processor microcontrollers, etc.) embedded in different locations in physical objects. Mobile cyber-physical systems, in which the physical systems in question have inherent mobility, are a subcategory of cyber-physical systems. Examples of mobile physical systems include mobile robots and electronic devices transported by humans or animals.

[0059] Furthermore, although the apparatus is depicted as a single entity, different units, processors and / or memory units may be implemented. Figure 1 Not all are shown).

[0060] 5G is able to use Multiple-Input Multiple-Output (MIMO) antennas, many more base stations or nodes than existing LTE systems (the so-called small cell concept), including macro sites cooperating with smaller base stations and employing multiple radio technologies, depending on service requirements, use cases and / or available spectrum. 5G mobile communications supports a variety of use cases and related applications (including video streaming, augmented reality, different ways of sharing data and various forms of machine type applications, such as (massive) machine type communications (mMTC), including vehicle safety, different sensors and real-time control. 5G is expected to have multiple radio interfaces, namely, sub-6GHz, cmWave and mmWave, and is also integrable with existing conventional radio access technologies such as LTE. The integration with LTE can be implemented, at least in the early stages, as a system in which macro coverage is provided by LTE and 5G radio interface access comes from small cells by aggregation to LTE. In other words, 5G plans to support both inter-RAT operability (such as LTE-5G) and inter-RI operability (inter-radio interface operability, such as sub-6GHz cmWave, sub-6GHz cmWave, mmWave). One of the concepts considered for use in 5G networks is network slicing, in which multiple independent and dedicated virtual subnets (network instances) can be created in the same infrastructure to run services with different requirements for latency, reliability, throughput and mobility.

[0061] The current architecture in LTE networks is completely distributed in the radio and completely centralized in the core network. Low-latency applications and services in 5G require content to be close to the radio, which causes local bursts and multi-access edge computing (MEC). 5G enables analysis and knowledge generation to be performed at the data source. This approach requires full use of resources that may not be continuously connected to the network, such as laptops, smart phones, tablets, and sensors. MEC provides a distributed computing environment for application and service hosting. It also has the ability to store and process content near cellular subscribers to speed up response time. Edge computing covers a wide range of technologies, such as wireless sensor networks, mobile data collection, mobile signature analysis, collaborative distributed peer-to-peer self-organizing networks and processing (also classified as local cloud / fog computing and grid / mesh computing), dew computing, mobile edge computing, cloudlet, distributed data storage and retrieval, autonomous self-healing networks, remote cloud services, augmented and virtual reality, data caching, Internet of Things (large-scale connectivity and / or latency critical), critical communications (autonomous driving cars, traffic safety, real-time analysis, time-critical control, healthcare applications).

[0062] The communication system can also communicate with other networks, such as the public switched telephone network or the Internet, as shown by the component 122, or utilize services provided by them. The communication network can also be capable of supporting the use of cloud services, for example, at least a portion of the core network operations can be provided as a cloud service (this is in the Figure 1 The communication system may also include a central control entity that provides facilities for networks of different operators to cooperate, for example, in spectrum sharing.

[0063] Edge cloud technology can be introduced into the radio access network (RAN) using network function virtualization (NVF) and software defined networking (SDN). The use of edge cloud technology can mean that access node operations will be performed at least partially in a server, host or node that is operatively coupled to a remote radio head or base station including the radio part. Node operations can also be distributed among multiple servers, nodes or hosts. The application of cloudRAN architecture enables RAN real-time functions to be performed on the RAN side (in the distributed unit DU 114), and non-real-time functions can be performed in a centralized manner (in the centralized unit CU 118).

[0064] It should also be understood that the workload distribution between core network operations and base station operations may be different from LTE, or even non-existent. Some other technological advancements that may be used are big data and all-IP, which may change the way networks are built and managed. 5G is designed to support multiple hierarchical structures, where MEC servers can be placed between the core and the base station or nodeB (gNB). It should be understood that MEC can also be applied to 4G networks.

[0065] 5G can also make use of satellite communications to enhance or supplement the coverage of 5G services, for example by providing backhaul. Possible use cases are to provide service continuity for machine-to-machine (M2M) or Internet of Things (IoT) devices or for passengers on board, or to ensure service availability for critical communications and future rail / maritime / aeronautical communications. Satellite communications can make use of geostationary orbit (GEO) satellite systems as well as low Earth orbit (LEO) satellite systems, in particular mega-constellations (systems in which hundreds of (nano) satellites are deployed). Each satellite 116 in a mega-constellation can cover the network entities of several satellite-enabled cells that create a ground cell. The ground cell can be created via a ground relay node 114 or by a gNB located on the ground or in a satellite.

[0066] Those skilled in the art will appreciate that the depicted system is only an example of a portion of a radio access system, and in practice, the system may include multiple (e / g)NodeBs, which may access multiple radio cells, and the system may also include other devices, such as physical layer relay nodes or other network elements. One of the (e / g)NodeBs may be a home (e / g)NodeB. In addition, in a geographical area of ​​a radio communication system, multiple different types of radio cells and multiple radio cells may be provided. Radio cells may be macro cells (or umbrella cells), which are large cells with a diameter of typically tens of kilometers, or smaller cells such as micro, femto or pico cells. Figure 1 The (e / g)NodeB can provide any kind of these cells. A cellular radio system can be implemented as a multi-layer network comprising several kinds of cells. Typically, in a multi-layer network, one access node provides one or more cells of one kind, and thus providing such a network structure requires multiple (e / g)NodeBs.

[0067] In order to meet the needs of improving the deployment and execution of communication systems, the concept of "plug and play" (e / g) NodeB is introduced. Usually, in addition to the home (e / g) NodeB (H(e / g) NodeB), the network that can use "plug and play" (e / g) NodeB also includes the home nodeB gateway or HNB-GW ( Figure 1 The HNB gateway (HNB-GW) usually installed in the operator's network can aggregate services from a large number of HNBs back to the core network.

[0068] Figure 2 is a schematic diagram of a wireless communication system 200. The communication system 200 may be configured to use a time division duplex (TDD) technique for data transmission.

[0069] For simplicity, the communication system 200 is shown as including four base stations BS1, BS2, BS3 and BS4. Each of the base stations BS1, BS2, BS3 and BS4 may be, for example, an eNodeB or a gNB, as shown in FIG. Figure 1 That is, the communication system 200 may support the same RAT or different RATs.

[0070] Each of the base stations BS1 to BS4 may serve UEs within a respective service geographical coverage area or cell 203.1-4. For simplicity, only one user equipment 201 is illustrated. In this particular example, the UE 201 is served by the base station BS1.

[0071] In this particular example, base station BS2 may include a trained machine learning model 205. Machine learning model 205 may be configured to receive inputs and provide outputs in response to receiving the inputs. The output of machine learning model 205 may be a prediction of any of the following: the trajectory of user equipment 201; the cell loads at cells 203.2, 203.3, and 203.4, the target cell for handover of user equipment 201, and the like. For example, the values ​​of the inputs may be provided by user equipment 201 and / or base stations BS3 to BS4 and / or base station BS2. Base station BS4 may include a trained machine learning model 207. For example, machine learning model 207 may be configured to receive its inputs so as to predict the traffic load in communication system 200. User equipment 201 may include a trained machine learning model 209. Machine learning model 209 may be configured to receive its inputs so as to predict the target cell for load balancing in communication system 200.

[0072] Base station BS1 may be configured to use the output of machine learning model 205 and / or the output of machine learning model 207 and / or the output of machine learning model 209 in order to control data communication in communication system 200 .

[0073] Figure 3 is a flow chart of a method according to an example of the present subject matter. For the purpose of explanation, Figure 3 The method described in Figure 2 For example, the method may be performed by a device such as a base station BS2, but is not limited thereto, because those skilled in the art may apply the present invention to another element of a communication system having the necessary attributes.

[0074] The method starts at step 301, where the base station BS2 may receive a request to send data from a node such as the base station BS1, the data including at least one of the following: the output of the trained machine learning model 205 or an indication of an event that may affect the control of the data communication by the base station BS1. For example, the base station BS1 may be one of the one or more nodes requesting the base station BS2 to send data. In response to receiving the request, the method proceeds to step 303, where the base station BS2 may determine the current value of the input of the machine learning model 205. To this end, the base station BS2 may request the user equipment 201 and / or the base station BS3 and / or the base station BS4 to provide the current value of the input, for example, each of them may provide the value of its corresponding part of the input. The method proceeds to step 305, where the base station BS2 may perform reasoning of the machine learning model 205 using the determined current value of the input. The machine learning model 205 may provide a current output in response to receiving the current input. The base station BS2 may determine in step 306 whether an event that affects the output of the machine learning model 205 has occurred. To this end, the base station BS2 can determine whether the event occurs in step 306, and can also determine whether the current output is affected after the occurrence of the event. If the event occurs and the output is affected, it indicates that the event affecting the output has occurred; otherwise, it indicates that the event affecting the output has not occurred. Alternatively, the base station BS2 can determine whether the current output is affected in step 306. If the output is affected, it can be determined that the event affecting the output has occurred; otherwise, it can be determined that the event affecting the output has not occurred. The output of the machine learning model 205 is affected, which can mean that the output quality of the machine learning model 205 is affected. For example, the base station BS2 can use the output to detect inference impairments and their types. The type of inference impairment can be, for example, an invalid impairment type or an inaccurate impairment type. An event that has occurred and has an impact on the output can have an event type of an impairment type, for example, it can be determined that the event type is an inaccurate impact event type. Figure 4 An alternative implementation is provided. The event that occurs may also be based on a condition indicated by the first BS1 requesting BS2 to send data, for example, one or more of the requested outputs satisfy a threshold condition. The base station BS2 may determine that the detected event may affect the control of the data communication by the base station BS1 because the event has an impact on the current output. Alternatively, if Figure 4As shown, if the current output is affected and satisfies the impact condition provided by the base station BS1, the base station BS2 can determine that the detected event affects the control of the data communication. Alternatively, the base station BS2 can determine whether the control of the data communication by the base station BS1 will be affected based on the determined event type. For example, using a predefined event type list that affects the control of the communication at BS1, the base station BS2 can determine whether the detected event type belongs to the list in order to determine whether the control of the data communication will be affected by the inference result. In step 307, the base station BS2 can send data to the base station BS1 based on the occurrence of an event that affects the output of the machine learning model 205. The base station BS2 can also send data to other nodes that have requested data. For example, if the event that affects the output of the machine learning model 205 is not detected (does not occur) in step 306, the data can be sent to the base station BS1 according to a predefined period in step 307. However, if the event that affects the output of the machine learning model 205 is detected in step 306, the data can be sent to the base station BS1 immediately in step 307. For example, if an event that affects the output of the machine learning model 205 is detected in step 306, data may be sent to the base station BS1 in step 307 T time units after the time point at which the event is detected, where the time unit may be microseconds, seconds, minutes, etc. For example, the data may be sent 10 microseconds after the detection of the event. In one example, if no periodic configuration is provided, data may be reported only when an event that affects the output of the machine learning model is detected.

[0075] For example, the base station BS2 may create a tuple or data structure with the following entries {OUT="", EVT="}. The first entry "OUT" may be set to the output of the machine learning model 205, and the second entry "EVT" may have a value indicating an event that may affect the control of data communication. Figure 3 For example, the base station BS2 may send one of the following tuples: {OUT="val", EVT="-"}, {OUT="-", EVT="val"} or {OUT="val", EVT="val"}, where "val" indicates that there is a value in the entry and "-" indicates that the entry is empty. Figure 4 An example implementation of step 307 is provided.

[0076] For example, if the event does not occur, the base station BS2 may send the current output to the base station BS1 in step 307, for example, it may send {OUT="val", EVT="-"}. For example, if the event is detected and the output quality of the current output is not affected by the detected event, the base station BS2 may send the current output to the base station BS1 in step 307, for example, it may send {OUT="val", EVT="-"}. However, if the detected event has an impact on the output quality, which means that the output quality is affected after the event is detected, the base station BS2 may send both the current output and an indication of the event to the base station BS1 in step 307, for example, it may send {OUT="val", EVT="val"}, so that the base station BS1 can decide whether the control of the communication will be affected by the output. Alternatively, if the detected event has an impact on the output quality, the base station BS2 may decide (itself) that the current output may not be available for controlling data communication at the base station BS1 (e.g., based on conditions or rules defined by the base station BS1), and may therefore send the output quality to the base station BS1 in step 307, e.g., it may send {OUT="-", EVT="val"}.

[0077] Figure 4 is a flow chart of a method according to an example of the present subject matter. For the purpose of explanation, Figure 4 The method described in Figure 2 For example, the method may be performed by a device such as a base station BS2, but is not limited thereto, because those skilled in the art may apply the present invention to another element of a communication system having the necessary attributes.

[0078] The method starts at step 401, where base station BS2 may receive a request to send data from a node such as base station BS1, the data including at least one of the following: the output of the trained machine learning model 205 or an indication of an event that may affect the control of data communication by base station BS1. The request may also include an impact condition. For example, the impact condition may indicate whether the output of the trained machine learning model 205 may affect the decision of base station BS1 to control data communication. For example, the impact condition may require that the validity time is reduced by a specific value compared to a reference validity time, and / or the accuracy of the output is reduced by a specific value compared to a reference accuracy, and / or the output exceeds a specific threshold.

[0079] In response to receiving the request, the method proceeds to step 403, where the base station BS2 may determine the current value of the input of the machine learning model 205. To this end, the base station BS2 may request the user equipment 201 and / or the base station BS3 and / or the base station BS4 to provide the current value of the input, for example, each of them may provide the value of its respective part of the input.

[0080] The method proceeds to step 405, where the base station BS2 may use the determined current value of the input to perform inference of the machine learning model 205. The machine learning model 205 may provide a current output in response to receiving the current input.

[0081] Base station BS2 may detect an event that may affect the output of machine learning model 205 in step 406. For example, base station BS2 may detect the event by receiving one or more signals indicating the event, for example, an event from a base station such as BS3 may be received at base station BS2. The event may be related to a failure, an unexpected event (e.g., an unexpected handover from BS3 or a series of handovers to offload a group of UEs to BS2, etc.).

[0082] The base station BS2 may check in step 407 whether the current output satisfies the imparting condition. For example, it may check whether the validity time of the current output is reduced by a specific value compared to the reference validity time and / or whether the accuracy of the current output is reduced by a specific value. Therefore, determining whether an event affecting the output of the machine learning model has occurred includes steps 406 and 407. Specifically, determining that the event affects the output of the machine learning model includes steps 406 and 407, wherein the current output satisfies the condition in step 407. Therefore, steps 406 and 407 may trigger event-based reporting.

[0083] If the current output satisfies the imparting condition, the method proceeds to step 409, where the base station BS2 sends data indicating that the imparting condition is satisfied to the base station BS1, for example, it may send {OUT="-", EVT="val"}. After determining that the output satisfies the condition, the data may be sent immediately. If the current output does not satisfy the imparting condition, the method proceeds to step 411, where the base station BS2 sends data including the current output to the base station BS1, for example, it may send: {OUT="val", EVT="-"}.

[0084] Figure 5 is a flow chart of a method according to an example of the present subject matter. For the purpose of explanation, Figure 5 The method described in Figure 2 For example, the method may be executed by the base station BS1, but is not limited thereto, because those skilled in the art may apply the present invention to another element of the communication system having the necessary properties.

[0085] The method starts at step 501, where a node such as a base station BS1 can configure a device such as a base station BS2 to send data, the data including at least one of the following: the output of a trained machine learning model or an indication of an event that may affect the control of data communication by the base station BS1. After configuring the base station BS2, the base station BS1 can receive the requested data from the base station BS2 in step 503. Based on the received data, the base station BS1 can perform control of the data communication. For example, the received data can include a tuple {OUT, EVT}. The base station BS1 can check the value stored in the entry EVT to make a decision. For example, if the entry EVT has no value, the base station BS1 can use the value of OUT to control the data communication in the communication system. For example, if the value of OUT is not set and the value stored in EVT (e.g., 2) indicates that the output of ML1 is affected (e.g., invalid), the base station BS1 can send a request to BS2 to stop the submission of data (e.g., tuple). If the value stored in the EVT (e.g., 1) indicates that the quality of the output of ML1 (which may also be part of the EVT) is affected, and OUT has a value, the base station BS1 can decide on its own whether the quality is acceptable; and this can be advantageous in the case where the base station BS2 cannot decide whether the output affects the control of data communications. If the value of the output with this quality does not affect the control of data communications, the quality is acceptable. Affecting the control of data communications means that it cannot provide the expected control results. If the quality is acceptable, the base station BS1 can use the value of OUT to control the data communication in the communication system. If the quality is unacceptable, the base station BS1 can send a request to BS2 to stop the submission of data. Alternatively, if the value stored in the EVT (e.g., 3) indicates that the quality of the output of ML1 is affected, the base station BS1 can decide to use the value of OUT to control the data communication in the communication system.

[0086] In an alternative implementation, base station BS1 may configure base station BS2 to periodically send data including at least one of the following: an output of a trained machine learning model or an indication of an event that may affect base station BS1's control of data communications. In this case, steps 503 to 505 may be repeated by base station BS1 for each data received from base station BS2.

[0087] Figure 6 is a flow chart of a method according to an example of the present subject matter. For the purpose of explanation, Figure 6 The method described in Figure 2For example, the method may be executed by the base station BS1, but is not limited thereto, because those skilled in the art may apply the present invention to another element of the communication system having the necessary properties.

[0088] The method starts from step 601, where a node such as base station BS1 can configure multiple elements of the communication system, such as configuring base station BS2, base station BS4 and user equipment 201, so that the data sent by each of them includes at least one of the following: the output of the associated trained machine learning model 205 to 209 or an indication of an event that may affect the control of data communication by base station BS1. After configuring base station BS2, base station BS4 and user equipment 201, in step 503, base station BS1 can receive requested data from base station BS2, receive corresponding requested data from base station BS4, and receive corresponding requested data from user equipment 201. For example, it can receive the tuple {OUT1, EVT1} from base station BS2, the tuple {OUT2, EVT2} from base station BS4, and the tuple {OUT3, EVT3} from the user equipment. Based on all three received data, base station BS1 can Figure 5 The control of data communication is performed in a similar manner as described above. For example, if any one of EVT1, EVT2, and EVT3 (e.g., EVT2) has a value of "2", the base station BS1 may not perform control over data communication, and may send a request to the element associated with the one of EVT1, EVT2, and EVT3 (e.g., BS4) to stop transmission and retrain its corresponding machine learning model. If EVT1, EVT2, and EVT3 (e.g., EVT2) all have no value, the base station BS1 may use the values ​​of OUT1, OUT2, and OUT3 to control data communication in the communication system.

[0089] For example, if any one of EVT1, EVT2, and EVT3 (e.g., EVT1) has a value of "1", the base station BS1 can decide on its own whether the quality is acceptable. If the quality is acceptable, the base station BS1 can use the values ​​of OUT1, OUT2, and OUT3 to control data communication in the communication system. If the quality is unacceptable, the base station BS1 can send a request to BS2 to stop the submission of data.

[0090] Figure 7 An example of a signaling diagram between a first node (such as NG-RAN node 1, referred to as NG-RAN1 and 701), a second node (such as NG-RAN node 2, referred to as NG-RAN2 and 703), and a third node (such as NG-RAN node 3, referred to as NG-RAN3 and 705) according to an example of the present subject matter is depicted. Figure 7In the example of , the NG-RAN 2703 may have a trained ML model.

[0091] At 1, NG-RAN1 701 configures inference measurements (e.g., predictions / estimations) to NG-RAN2 703, which includes a set of conditions that, when changed at NG-RAN2 703, may also affect decisions at NG-RAN1 701, for example by invalidating ML models available at NG-RAN 701. At 2, NG-RAN2 703 accepts the measurement configuration. At 3, NG-RAN2 703 runs the inference of the ML. At 4, NG-RAN2 703 sends the inference measurements (prediction / estimation) to NG-RAN1 701. At 5, a failure or unexpected event (e.g., an unexpected handover from NG-RAN3 705 offloads a set of UEs to NG-RAN2 703) occurs. At 6, NG-RAN2 703 detects that the ML inference has become invalid. At 7, NG-RAN2 703 checks the conditions in the configuration. If the inferred invalidity is of a type that affects decisions at NG-RAN1 701, for example by causing the ML model at NG-RAN1 701 (in training or execution) to require retraining, then NG-RAN2 703 may need to update this information to the neighbors. At 8, NG-RAN2 703 sends an indication to NG-RAN1 701, which is used to inform its neighbors that the conditions affecting the validity of the ML model at NG-RAN1 701 have changed. At 9, NG-RAN1 701 may instruct NG-RAN2 703 to stop measuring.

[0092] exist Figure 7 In the example, the reporting events are based on inferred impairments detected by the reporting node, here NG-RAN2 703. Specifically, NG-RAN2 703 may detect and report those pre-configured impairments or changes in its cells that are expected to affect the reliability of the ML model trained at the first NG-RAN1 701. In the presence of inferred impairments, the initiating node may cancel the measurement configuration.

[0093] Figure 8 An example of a signaling diagram between a first node (such as NG-RAN node 1, referred to as NG-RAN1 and 801), a second node (such as NG-RAN node 2, referred to as NG-RAN2 and 803), and a third node (such as NG-RAN node 3, referred to as NG-RAN3 and 805) according to an example of the present subject matter is depicted. Figure 8 In the example of , NG-RAN2803 can have a trained ML model.

[0094] This can be Figure 7An additional option is provided where the first node (here NG-RAN1 801) can confirm that it can continue to receive measurements with updated entries and conditions. In this case, the configuration from NG-RAN1 801 can allow NG-RAN2 803 to continue reporting measurements even in the presence of inferred impairments, thereby enabling NG-RAN3 803 to perform "best effort" reporting operations by re-adjusting the reporting configuration. For example, if NG-RAN1 801 has configured NG-RAN2 803 to provide predicted load information with a validity period of 4 hours in the absence of any inferred impairments, if the configuration allows, then NG-RAN2 can still report these measurements by adjusting the reporting to once every 2 hours when the validity period is shortened by 50%.

[0095] Thus, at 1, NG-RAN1 801 configures an inference measurement (prediction / estimate) to NG-RAN2 803, which includes a set of conditions that, when changed at NG-RAN2 803, may also affect the decision at NG-RAN1 801, for example by invalidating the ML model available at NG-RAN 801. NG-RAN1 801 may indicate in the configuration that it is willing to accept inferences even in the presence of certain inference impairments, thereby enabling NG-RAN2 803 to perform "best effort" reporting operations by re-adjusting the reporting configuration. At 2, NG-RAN2 803 accepts the measurement configuration. At 3, NG-RAN2 803 runs the inference. At 4 and 5, NG-RAN2 803 sends the inference measurement (prediction / estimate) to NG-RAN1 801 within a reporting interval T with a prediction validity time V. At 6, events are received at NG-RAN2 803 from neighboring NG-RAN nodes. The event may be related to a failure, an unexpected event (e.g., an unexpected handover from NG-RAN3 805 offloading a group of UEs to NG-RAN2 803, etc.). At 7, NG-RAN2 803 detects that the ML reasoning has become invalid because the validity time of the reasoning has been halved. At 8, NG-RAN2 803 checks the conditions in the configuration and determines that its own reasoning invalidity may affect the decision at NG-RAN1 801 (e.g., invalidate another ML model in training or execution). If "best effort" measurement reporting is indicated with impairment of validity time in the configuration from NG-RAN1 801, NG-RAN2 803 decides to continue reporting to NG-RAN1 801. In addition, it decides to update the configuration to compensate for the impairment by halving the reporting cycle time. At 9, NG-RAN2 803 continues reasoning measurement reporting under the new configuration.

[0096] Therefore, if the configuration allows, the reporting interval can be dynamically updated to allow the reporting node to dynamically adapt the measurement reporting interval. In some other cases, the reporting node can only indicate (multiple) impairments to the first node without updating the reporting configuration accordingly. In this case, the first node can initiate another process to the other node to reconfigure the measurement reporting.

[0097] Fig. 9 An example of a signaling diagram between a first node (such as NG-RAN node 1, referred to as NG-RAN1 and 901), a second node (such as NG-RAN node 2, referred to as NG-RAN2 and 903), and a third node (such as NG-RAN node 3, referred to as NG-RAN3 and 905) according to an example of the present subject matter is depicted. Fig. 9 In the example of , NG-RAN 2903 may have a trained ML model.

[0098] At 1, NG-RAN1 901 configures an inference measurement (prediction / estimation) to NG-RAN2 903, which includes a set of conditions that, when changed at NG-RAN2 903, may also affect the decision at NG-RAN1 901, for example by invalidating the ML model available at NG-RAN 901. At 2, NG-RAN2 903 accepts the measurement configuration. At 3, NG-RAN2 903 runs the inference. At 4, NG-RAN2 903 sends the inference measurement (prediction / estimation) to NG-RAN1 901. At 5, a failure or unexpected event (e.g., an unexpected handover from NG-RAN3 905 offloads a group of UEs to NG-RAN2 903. At 6, NG-RAN2 903 detects that the ML reasoning has become invalid (e.g., reliability has been affected). At 7, NG-RAN2 903 checks the conditions in the configuration. If the reliability is not of the type that affects the decision at NG-RAN1 901, for example, causing the ML model (in training or execution) at NG-RAN1 901 to require retraining, then NG-RAN2 903 will continue to report normally. At 8, NG-RAN2 903 sends the reasoning measurements (predictions / estimations) to NG-RAN1 901 without making any changes to the reporting.

[0099] Therefore, according to Fig. 9 In the method, impairments may affect information such as the reliability of the measurement and also need to be communicated. If the reliability of the measurement is not considered an impairment of the original configuration, the reliability may be reduced, even significantly reduced, without any impact on the report (for example, the same report may be considered despite the measurement reliability being reduced from 90% to 70%).

[0100] Fig.10An example of a signaling diagram between a first node (such as NG-RAN node 1, referred to as NG-RAN1 and 1001), a second node (such as NG-RAN node 2, referred to as NG-RAN2 and 1003), and a third node (such as NG-RAN node 3, referred to as NG-RAN3 and 1005) according to an example of the present subject matter is depicted.

[0101] In addition to predictions, other measurements from NG-RAN2 1003 may also invalidate the trained ML model or the ML model in training at NG-RAN1 1001. This may occur, for example, when certain events (e.g., offloading operations, errors, etc.) significantly change the statistical distribution of the reported information. To this end, NG-RAN2 1003 may be configured to report load information to NG-RAN1 1001 when certain thresholds are exceeded: a measurement on one cell or a measurement on a group of cells exceeds a threshold (e.g., a load measurement), a measurement on one cell or a measurement on a group of cells is below a threshold (e.g., a load measurement). This is done in Fig.10 Shown in.

[0102] At 1, NG-RAN1 1001 configures measurements to NG-RAN2 1003, e.g., the load measurements include a set of conditions that, when changed at NG-RAN2 1003, may also affect decisions at NG-RAN1 1001, e.g., by invalidating ML models available at NG-RAN1 1001. In this example, these conditions include the load exceeding a threshold. At 2, NG-RAN2 1003 accepts the measurement configuration. At 3, NG-RAN2 1003 calculates the load. At 4, NG-RAN2 1003 checks whether the load exceeds a threshold. At 5, when the load does not exceed the threshold, NG-RAN2 1003 reports the load to NG-RAN1 1001. At 6, a failure or unexpected event (e.g., an unexpected handover from NG-RAN3 1005 offloads a set of UEs to NG-RAN2 1003, which changes the load statistics at NG-RAN2 1003). At 7, NG-RAN2 1003 checks whether the load exceeds a threshold. At 8, NG-RAN2 1003 sends an indication to NG-RAN1 1001 notifying that a change in the impact on NG-RAN2 1003 has been detected. It may also convey the experienced load. Fig.10 In the example, no periodic reporting is configured and measurements are reported only if the corresponding event is met. Event-based reporting can be achieved by introducing a process responsible for reporting events that may affect the training of artificial intelligence (AI) / ML models at the NG-RAN node and updating the existing resource status process to include the events in the report.

[0103] Fig.11 is a functional framework for model training and inference according to the examples in this topic. The functional framework can be found at Figure 2 The system is implemented in the illustrated embodiment, but is not limited to such implementation.

[0104] Two nodes may be provided. The first node 1031 may be configured to receive data from the second node 1032. The second node 1032 may be configured to perform training of the model training function 1034 and reasoning of the model reasoning function 1035. Figure 2 In the example of FIG. 1 , the first node 1031 may be a base station BS1, and the second node 1032 may be a base station BS2.

[0105] The data collection function 1030 may be a function that provides input data to the model training function 1034 at the second node and the model reasoning function 1035 at the second node 1032. AI / ML algorithm specific data preparation (e.g., data preprocessing and cleaning, formatting, and transformation) may not be performed in the data collection function 1030. Examples of input data may include measurements from a UE or a different network entity, feedback from the first node 1031, and output from an AI / ML model.

[0106] The data collection function 1030 may send training data to the model training function 1034 and send inference data to the model inference function 1035. The training data refers to the data required as input to the AI / ML model training function 1034. The inference data refers to the data required as input to the AI / ML model inference function 1035.

[0107] Model training function 1034 can be a function that performs AI / ML model training, validation, and testing, which can be used to generate model execution metrics as part of the model testing process. Figure 2 The machine learning model 205. If required, the model training function 1034 can also be responsible for data preparation (e.g., data preprocessing and cleaning, formatting and transformation) based on the training data delivered by the data collection function 1030.

[0108] The model training function 1034 can be configured to perform model deployment / update. Model deployment / update can be used to initially deploy the trained, validated, and tested AI / ML model to the model reasoning function 1035, or to deliver an updated model to the model reasoning function 1035. The AI / ML model can be, for example, the machine learning model 205. That is, the model deployment / update can initially deploy the trained, validated, and tested machine learning model 205 to the model reasoning function 1035.

[0109] The model reasoning function 1035 may be a function that provides an AI / ML model reasoning output (e.g., a prediction or decision). Where applicable, the model reasoning function 1035 may provide model execution feedback to the model training function 1034. If required, the model reasoning function 1035 may also be responsible for data preparation (e.g., data preprocessing and cleaning, formatting, and transformation) based on the reasoning data delivered by the data collection function 1030. The output refers to the reasoning output of the AI / ML model generated by the model reasoning function 1035. Model execution feedback can be used to monitor the execution of the AI / ML model (if available).

[0110] The first node 1031 may include a participant function that receives output from the model reasoning function 1035 and triggers or performs corresponding actions. The participant function may trigger actions for other entities or itself. The participant function may provide feedback. The feedback may be information needed to obtain training data, reasoning data, or monitor the execution of the AI / ML model and its impact on the network by updating KPIs and execution counters.

[0111] exist Fig.11 In the text, the terms “output”, “inference output”, “AI / ML model inference output”, “output from the AI / ML model”, and “output from the model inference function” refer to the output of the AI / ML model.

[0112] exist Fig.12 , a circuit block diagram of a configuration of an illustrated device 1070 is shown, which is configured to implement at least a portion of the present subject matter. It should be noted that Fig.12The device 1070 shown may include several other elements or functions in addition to those described below, which are omitted here for simplicity because they are not necessary for understanding. In addition, the device may also be another device with similar functions, such as a chipset, a chip, a module, etc., which may also be a part of the device or attached to the device 1070, etc. as a separate element. The device 1070 may include a processing function or a processor 1071, such as a central processing unit (CPU), etc., which executes instructions given by a program, etc., associated with a flow control mechanism. The processor 1071 may include one or more processing parts dedicated to a specific process as described below, or the process may be run in a single processor. The part for performing such a specific process may also be provided as a discrete element, or provided in one or more other processors or processing parts, such as in a physical processor (such as a CPU) or in several physical entities. Reference numeral 1072 represents a transceiver or input / output (I / O) unit (interface) connected to the processor 1071. The I / O unit 1072 may be used to communicate with one or more other network elements, entities, terminals, etc. The I / O unit 1072 may be a combined unit including communication devices toward several network elements, or may include a distributed structure with a plurality of different interfaces for different network elements. Reference numeral 1073 denotes a memory that may be used, for example, to store data and programs to be executed by the processor 1071 and / or used as a working memory for the processor 1071.

[0113] The processor 1071 is configured to perform processing related to the above-mentioned subject matter. Specifically, the device 1070 may be configured to perform a combination of Figure 3 , Figure 4 , Figure 5 or Figure 6 The method described.

[0114] For example, processor 1071 is configured to: receive a request to send data from a node, the data including at least one of: an output of a trained machine learning model or an indication of an event that may affect control of data communications; infer the machine learning model using current values ​​of inputs to the machine learning model to obtain current outputs; detect events that may affect control of data communications; and send data to the node based on the detected events.

[0115] Alternatively, processor 1071 is configured to: configure a device of the wireless communication system to send data, the data comprising at least one of: an output of a trained machine learning model or an indication of an event that may affect control of data communications; receive requested data from the device; and perform control of data communications based on the received data.

[0116] Those skilled in the art will appreciate that aspects of the present invention may be embodied as devices, methods, computer programs or computer program products. Therefore, aspects of the present invention may take the form of a complete hardware embodiment, a complete software embodiment (including firmware, resident software, microcode, etc.), or an embodiment combining software and hardware aspects, which are generally referred to herein as "circuits", "modules" or "systems". In addition, aspects of the present invention may take the form of a computer program product embodied in one or more computer-readable media, which includes computer executable code. A computer program includes computer executable code or "program instructions".

[0117] Any combination of one or more computer readable media may be utilized. A computer readable medium may be a computer readable storage medium. As used herein, "computer readable storage medium" includes any tangible storage medium that can store instructions that can be executed by a processor of a computing device. A computer readable storage medium may be referred to as a computer readable non-transitory storage medium. A computer readable storage medium may also be referred to as a tangible computer readable medium. In some embodiments, a computer readable storage medium may also be capable of storing data that can be accessed by a processor of a computing device.

[0118] "Computer memory" or "memory" is an example of a computer-readable storage medium. Computer memory is any memory that is directly accessible to a processor. "Computer storage" or "storage" is another example of a computer-readable storage medium. Computer storage is any non-volatile computer-readable storage medium. In some embodiments, computer storage may also be computer memory and vice versa.

[0119] As used herein, 'processor' includes an electronic component capable of executing a program or machine executable instructions or computer executable code. References to computing devices that include a "processor" should be interpreted as possibly containing more than one processor or processing core. A processor may, for example, be a multi-core processor. A processor may also refer to a collection of processors within a single computer system or distributed between multiple computer systems. The term computing device should also be interpreted as possibly referring to a collection or network of computing devices, each of which includes one or more processors. Computer executable code may be executed by multiple processors, which may be within the same computing device, or may even be distributed across multiple computing devices.

[0120] Computer executable code may include machine executable instructions or programs that cause a processor to perform an aspect of the present invention. Computer executable code for performing the operations of aspects of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​(such as Java, Smalltalk, C++, etc.) and traditional procedural programming languages ​​(such as "C" programming language or similar programming languages), and compiled into machine executable instructions. In some instances, the computer executable code may be in the form of a high-level language or in a precompiled form and used in conjunction with an interpreter that generates machine executable instructions on the fly.

[0121] Typically, program instructions can be executed on one processor or several processors. In the case of multiple processors, the processors may be distributed across several different entities. Each processor may execute a portion of the instructions for that entity. Thus, when referring to a system or process involving multiple entities, a computer program or program instructions is understood to be suitable for execution by a processor associated or associated with the respective entity.

Claims

1. An apparatus (201, BS2) for a wireless communication system (200), the apparatus comprising a machine learning model (209, 205), the machine learning model (207) being configured to receive an input and provide an output, the output enabling control of data communication in the wireless communication system by a node; the apparatus comprising components configured to: receiving (301) a request from the node to send data, the data comprising at least one of: an output of the machine learning model, or an indication of an event that may affect the control of the data communication; inferring (305) the machine learning model using the current value of the input to obtain a current output; determining (306) whether an event that affects the output of the machine learning model occurs; The data is sent (307) to the node based on the occurrence of the event.

2. The apparatus according to claim 1, wherein the component is configured to: In response to determining that the event occurs, immediately sending to the node the data including at least one of: the current output, or the indication of the event; In response to determining that the event has not occurred, sending the data including the current output to the node according to a reporting period.

3. An apparatus according to any one of the preceding claims, wherein the machine learning model is trained with values ​​of the input having a defined pattern, and the component is configured to determine whether the current value of the input matches the pattern; in response to determining that the current value of the input does not match the pattern, determining that the event has occurred.

4. The apparatus according to any one of the preceding claims 1 to 2, wherein the determining comprises: Determine whether the event occurs, and determine whether the event that occurs causes a change in the output quality of the machine learning model. 5 . The apparatus according to claim 4 , wherein the output quality comprises at least one of the following: validity time of the output, or accuracy of the output.

6. The apparatus according to any one of the preceding claims 2 to 5, the request indicating a reporting configuration, wherein the reporting configuration indicates the reporting period.

7. The apparatus according to claim 6, wherein the component is configured to: in response to the occurrence of the event and the sending of the data, receive a request to stop submitting the data and retrain the machine learning model, and stop submitting the data.

8. The apparatus according to any one of the preceding claims, wherein the request further indicates an influencing condition, the component being configured to: In response to determining that the event has occurred, determining whether the current output of the machine learning model satisfies the impact condition; In response to determining that the current output satisfies the condition, sending the data including the indication that the condition is satisfied.

9. The apparatus according to claim 8, wherein the component is configured to: in response to sending the data indicating that the condition is met, receive a request to stop submitting the data and retraining the machine learning model, and stop submitting the data.

10. The device according to any one of the preceding claims, wherein the influencing condition requires at least one of the following: The validity time of the output is reduced by a certain value, or the accuracy of the output is reduced by a certain value.

11. An apparatus according to any one of the preceding claims 8 to 10, wherein the component is configured to: in response to determining that the current output does not satisfy the impact condition, modify a reporting configuration to prevent the current output from affecting the node control of data communication, and send the data according to the modified reporting configuration.

12. The apparatus of claim 11, wherein the influencing condition requiring the validity time of the output to be reduced by a specific value, wherein modifying the report configuration comprises: The reporting cycle time is changed to a time that is less than or equal to the validity time of the current output.

13. The device according to any of the preceding claims, comprising any of: a node of a radio access network of said communication system, Core network nodes, and User equipment.

14. The apparatus of any preceding claim, wherein the output of the machine learning model is a prediction of any of: Trajectory of one or more user devices; cell load at one or more cells served by a node of the communication system; A target cell for handover; Services of user equipment; target cell for load balancing; resource status of the node or other nodes of the communication system; a user equipment to be handed over to a target node of the communication system; Cell activation or deactivation recommendation.

15. The apparatus of any preceding claim, wherein the input comprises at least one of: User device location; Cell-level measurements; Beam-level measurements; The load of the cell; UE Reference Signal Received Power (RSRP) measurement, UE Reference Signal Received Quality (RSRQ) measurement, UE signal-to-interference-and-noise ratio (SINR) measurement; current or predicted resource status at one or more nodes; Current or predicted energy efficiency at one or more nodes.

16. An apparatus according to any one of the preceding claims, wherein the component comprises at least one processor; and at least one memory comprising computer program code, wherein the at least one memory and the computer program code are configured to, together with the at least one processor, cause execution of the apparatus.

17. A method for using a machine learning model (209, 205), the machine learning model being configured to receive an input and provide an output, the output enabling control of data communications in the wireless communication system (200) by a node; the method comprising: receiving (301) a request from the node to send data, the data comprising at least one of: an output of the trained machine learning model, or an indication of an event that may affect the control of the data communication; inferring (305) the machine learning model using the current value of the input to obtain a current output; determining (306) whether an event that affects the output of the machine learning model occurs; The data is sent (307) to the node based on the occurrence of the event.

18. The method according to claim 17, further comprising: The inference, determination and sending steps are repeated.

19. The method according to claim 17 or 18, wherein the sending comprises: In response to determining that the event occurs, immediately sending the data including at least one of: the current output, or the indication of the event; or In response to determining that the event has not occurred, the data including the current output is sent according to a reporting period.

20. The method according to claim 19, further comprising: A stop request is received from the node in response to immediately sending the data, and submission of the data is stopped in response to receiving the stop request.

21. A computer program comprising instructions for causing an apparatus to perform at least the following: Receiving (301) a request from a node to send data, the data comprising at least one of: an output of a trained machine learning model, or an indication of an event that may affect the control of the data communication; inferring (305) the machine learning model using current values ​​of inputs to the machine learning model to obtain current outputs; determining (306) whether an event that affects the output of the machine learning model occurs; The data is sent (307) to the node based on the detected event.

22. A node (BS1) for controlling data communication in a wireless communication system (200); the node (BS1) comprising components configured to: configuring (501) an apparatus (201, BS2) of the wireless communication system to transmit data, the data comprising at least one of: an output of a trained machine learning model or an indication of a detected event that may affect the control of the data communication; receiving (503) the requested data from the device; The controlling of the data communication is performed (505) based on the received data.

23. The node of claim 22, wherein the components include at least one processor; and at least one memory including computer program code, the at least one memory and the computer program code being configured to, together with the at least one processor, cause execution of the node.

24. A method for controlling data communication in a wireless communication system; the method comprising: configuring (501) an apparatus of the wireless communication system to transmit data, the data comprising at least one of: an output of a trained machine learning model, or an indication of a detected event that may affect the control of the data communication; The requested data is received (503) from the device; and the control of the data communication is performed (505) in the wireless communication system based on the received data.

25. A computer program comprising instructions for causing a node to perform at least the following: configuring a device of the wireless communication system to send data, the data comprising at least one of the following: an output of a trained machine learning model, or an indication of a detected event that may affect the control of the data communication; receiving the requested data from the device; and performing the control of the data communication in the wireless communication system based on the received data.

Images