Method for operating ai / ML model in wireless communication system, and apparatus therefor

Abstract

This method performed by a first wireless device in a wireless communication system supporting an AI / ML model comprises the steps of: training a surrogate model and training a secondary model on the basis that a value related to loss in a primary model is greater than a first threshold value; and, on the basis of a value indicating the difference between a parameter of the surrogate model and a parameter of the secondary model, changing, from the primary model to the secondary model, a model activated for performing a task related to wireless communication.

Description

Method and device for operating an AI / ML model in a wireless communication system

[0001] This relates to a method for operating an AI / ML model in a wireless communication system and a device therefor.

[0002] Mobile communication systems were developed to provide voice services while ensuring user activity. However, they have expanded beyond voice to include data services. Currently, explosive growth in traffic is leading to resource shortages and users are demanding faster services, necessitating a more advanced mobile communication system.

[0003] Next-generation mobile communication systems must support explosive data traffic growth, dramatically increasing data rates per user, a vastly increased number of connected devices, ultra-low end-to-end latency, and high energy efficiency. To achieve these goals, various technologies are being studied, including dual connectivity, massive multiple input multiple output (MIMO), in-band full duplex, non-orthogonal multiple access (NOMA), super wideband support, and device networking.

[0004] Meanwhile, Deep Neural Networks (DNNs) can be utilized in wireless communication systems that support AI / ML (Artificial Intelligence / Machine Learning). While DNNs offer high performance, they have low explainability. Therefore, it is difficult to define the correlation between inputs and outputs in DNNs. This problem applies not only to DNNs but also to other AI / algorithms / learning methods with low explainability.

[0005] As described above, AI / ML models with high performance but low explainability can be utilized for wireless communications. If the performance of these AI / ML models degrades, it can be difficult to clearly identify the cause of the performance degradation, and it can take a long time for wireless communication quality to return to normal.

[0006] The purpose of this specification is to minimize the extent to which wireless communication quality deteriorates when AI / ML model performance degrades.

[0007] In a wireless communication system according to one embodiment of the present specification supporting an AI / ML (Artificial Intelligence / Machine Learning) model, a method performed by a first wireless device includes the steps of: performing training of a surrogate model and training of a secondary model based on a value related to a loss of a primary model being greater than a first threshold; and changing a model activated for performing a task related to the wireless communication from the primary model to the secondary model based on a value representing a difference between a parameter of the surrogate model and a parameter of the secondary model. The primary model is a model activated for performing a task related to the wireless communication, and the surrogate model and the secondary model have the same model type and structure and different parameters based on learning. The training of the surrogate model is based on transfer learning of the primary model. The training of the secondary model is online learning based on a wireless signal currently being transmitted and received.

[0008] The above first model may include a deep neural network (DNN) model.

[0009] The above secondary model may include a linear model, a kernel model, a graph model, and / or a probabilistic model, excluding a deep neural network.

[0010] The explainability of the above first model may be lower than that of the above second model.

[0011] The value representing the difference between the parameters of the above proxy model and the parameters of the second model may be an average value calculated based on the Euclidean distance between the parameters of the above proxy model and the parameters of the second model.

[0012] i) a value related to the loss of the first model is greater than a value obtained by adding an offset value to a value related to the loss of the second model, and ii) a value representing a difference between a parameter of the surrogate model and a parameter of the second model is greater than a second threshold: a model activated for performing a task related to the wireless communication may be changed from the first model to the second model.

[0013] The method may further include the step of transmitting a message related to a change in the activated model to the second wireless device. The message may include parameters of the surrogate model and parameters of the secondary model.

[0014] The state related to the operation of the above first model can transition from the ACTIVE state to the PREPARE state. The state related to the operation of the above second model can transition from the PREPARE state to the ACTIVE state.

[0015] Based on the value related to the performance of the first model being greater than the first threshold value, the state related to the operation of the surrogate model and the second model can be transitioned from the IDLE state to the PREPARE state.

[0016] The learning of the above proxy model and the learning of the second model can be performed in the PREPARE state.

[0017] The above transfer learning may be related to knowledge distillation.

[0018] The task related to the above wireless communication may be related to at least one of reporting of channel state information, measurement of reference signal, signal modulation, beam tracking, or beam search.

[0019] The size of the training data for learning the above second model may be smaller than the size of the training data for learning the above first model.

[0020] Training data for learning the above secondary model may include i) input and ii) output corresponding to the input.

[0021] The input may include a vector relating to at least one of a position, a velocity, or an attitude of the first wireless device. The output may include information determined based on the vector.

[0022] Information determined based on the above vector may include an index of a beam associated with transmission or reception of the first wireless device.

[0023] i) a value related to the loss of the secondary model is greater than a value obtained by adding an offset value to a value related to the loss of the primary model, and ii) a value representing a difference between a parameter of the surrogate model and a parameter of the secondary model is less than a second threshold value: a model activated for performing a task related to the wireless communication may be changed from the secondary model to the primary model.

[0024] The state related to the operation of the above first model can transition from the PREPARE state to the ACTIVE state. The state related to the operation of the above second model can transition from the ACTIVE state to the IDLE state.

[0025] The above first wireless device may be a sender, a receiver, a terminal, or a base station.

[0026] A first wireless device operating in a wireless communication system according to another embodiment of the present disclosure supporting an AI / ML (Artificial Intelligence / Machine Learning) model includes one or more transceivers, one or more processors controlling the one or more transceivers, and one or more memories operably connectable to the one or more processors and storing instructions that, when executed by the one or more processors, cause the one or more processors to perform operations.

[0027] The above operations include a step of performing training of a surrogate model and training of a secondary model based on a value related to a loss of the primary model being greater than a first threshold value, and a step of changing a model activated for performing a task related to wireless communication from the primary model to the secondary model based on a value representing a difference between parameters of the surrogate model and parameters of the secondary model. The primary model is a model activated for performing a task related to wireless communication, and the surrogate model and the secondary model have the same model type and structure and different parameters based on learning. The training of the surrogate model is based on transfer learning of the primary model. The training of the secondary model is online learning based on a wireless signal currently being transmitted and received.

[0028] A device according to another embodiment of the present disclosure comprises one or more memories and one or more processors functionally connected to the one or more memories.

[0029] The one or more memories store instructions that, when executed by the one or more processors, cause the one or more processors to perform operations.

[0030] The above operations include a step of performing training of a surrogate model and training of a secondary model based on a value related to a loss of the primary model being greater than a first threshold value, and a step of changing a model activated for performing a task related to wireless communication from the primary model to the secondary model based on a value representing a difference between parameters of the surrogate model and parameters of the secondary model. The primary model is a model activated for performing a task related to wireless communication, and the surrogate model and the secondary model have the same model type and structure and different parameters based on learning. The training of the surrogate model is based on transfer learning of the primary model. The training of the secondary model is online learning based on a wireless signal currently being transmitted and received.

[0031] In another embodiment of the present disclosure, one or more non-transitory computer-readable media store one or more instructions.

[0032] The one or more instructions, when executed by one or more processors, cause the one or more processors to perform operations.

[0033] The above operations include a step of performing training of a surrogate model and training of a secondary model based on a value related to a loss of the primary model being greater than a first threshold value, and a step of changing a model activated for performing a task related to wireless communication from the primary model to the secondary model based on a value representing a difference between parameters of the surrogate model and parameters of the secondary model. The primary model is a model activated for performing a task related to wireless communication, and the surrogate model and the secondary model have the same model type and structure and different parameters based on learning. The training of the surrogate model is based on transfer learning of the primary model. The training of the secondary model is online learning based on a wireless signal currently being transmitted and received.

[0034] When the performance of a model with low explainability degrades, the wireless communication quality may be severely degraded or communication may be interrupted until the performance of the model is restored. On the other hand, according to an embodiment of the present disclosure, when the performance of the primary model degrades, a wireless communication-related task may be performed based on a secondary model. Specifically, training of a surrogate model and a secondary model is performed based on values ​​related to the loss of the primary model, and the model activated for the wireless communication-related task is changed from the primary model to the secondary model based on the difference value between the parameters of the surrogate model and the secondary model.

[0035] Therefore, compared to using only AI / ML models with low explainability, the frequency and duration of deterioration in wireless communication quality can be reduced. Specifically, the overall reliability of AI / ML model-based wireless communication can be improved. Furthermore, AI / ML models with high explainability can more clearly identify causal relationships between system problems between the transmitter and receiver than models with low explainability, as the causal relationship between inputs and outputs can be explained by the model's trained parameters.

[0036] According to the existing method, when a terminal or base station operates a transmitter / receiver in conjunction with side information data transmitted from various additional sensors, the fallback operation to a transmission, reception, and measurement device using a mathematically based model, which is the existing communication method, cannot be performed. As described above, even when the transmitter / receiver is in conjunction with side information data and mathematically based modeling is limited or impossible, the fallback operation can be supported according to the present embodiment.

[0037] The effects that can be obtained from this specification are not limited to the effects mentioned above, and other effects that are not mentioned can be clearly understood by a person having ordinary skill in the technical field to which this specification belongs from the description below.

[0038] The accompanying drawings are intended to aid understanding of the present specification and may provide embodiments of the present specification along with detailed descriptions. However, the technical features of the present specification are not limited to specific drawings, and the features disclosed in each drawing may be combined with each other to form new embodiments. Reference numerals in each drawing may indicate structural elements.

[0039] Figure 1 is a drawing showing an example of a communication system applicable to this specification.

[0040] Figure 2 is a drawing showing an example of a wireless device applicable to this specification.

[0041] FIG. 3 is a diagram illustrating a method for processing a transmission signal applicable to the present specification.

[0042] FIG. 4 is a drawing showing another example of a wireless device applicable to this specification.

[0043] FIG. 5 is a drawing showing an example of a mobile device applicable to this specification.

[0044] Figure 6 is a diagram showing physical channels applicable to this specification and a signal transmission method using them.

[0045] Figure 7 is a diagram showing an example of a perceptron structure.

[0046] Figure 8 is a diagram showing an example of a multilayer perceptron structure.

[0047] Figure 9 is a diagram showing an example of a deep neural network.

[0048] Figure 10 is a diagram showing an example of a convolutional neural network.

[0049] Figure 11 is a diagram showing an example of a filter operation in a convolutional neural network.

[0050] Figure 12 shows an example of a neural network structure in which a recurrent loop exists.

[0051] Figure 13 shows an example of the operating structure of a recurrent neural network.

[0052] Figure 14 illustrates the functional framework of an AI / ML model.

[0053] Figure 15 illustrates a channel decoder including a neural network.

[0054] Figure 16 illustrates an autoencoder method in which neural networks are configured at both the transmitter and receiver.

[0055] Figure 17 illustrates deep learning-based channel coding and decoding operations.

[0056] Figure 18 illustrates learning performed in a wireless channel environment.

[0057] Figure 19 illustrates the trade-off between explainability and performance of AI / ML.

[0058] Figure 20 shows the structure of a sender and a receiver according to an embodiment of the present specification.

[0059] FIG. 21 is a flowchart illustrating a procedure for transmitting performance information related to whether or not support is provided for a primary-secondary AI / ML model according to an embodiment of the present specification.

[0060] Figure 22 is a flowchart illustrating a procedure for setting up a primary-secondary AI / ML model according to an embodiment of the present specification.

[0061] Figure 23 is a flowchart illustrating a procedure for releasing a primary-secondary AI / ML model according to an embodiment of the present specification.

[0062] Figure 24 shows the state transition of a secondary model according to an embodiment of the present specification.

[0063] Figure 25 illustrates state transitions of a surrogate model according to an embodiment of the present specification.

[0064] Figure 26 shows the state transition of the primary model according to an embodiment of the present specification.

[0065] Figure 27 illustrates a procedure for a secondary model of a transmitter to move from an IDLE state to a PREPARE state according to an embodiment of the present specification.

[0066] Figure 28 illustrates a procedure in which a secondary model of a transmitter moves from a PREPARE state to an ACTIVE state according to an embodiment of the present specification.

[0067] Figure 29 illustrates a procedure for a secondary model of a transmitter to move from a PREPARE state to an IDLE state according to an embodiment of the present specification.

[0068] Figure 30 illustrates a procedure for a secondary model of a transmitter to move from an ACTIVE state to an IDLE state according to an embodiment of the present specification.

[0069] Figure 31 illustrates a procedure for a secondary model of a receiver to move from an IDLE state to a PREPARE state according to an embodiment of the present specification.

[0070] Figure 32 illustrates a procedure in which a secondary model of a receiver according to an embodiment of the present specification moves from a PREPARE state to an ACTIVE state.

[0071] Figure 33 illustrates a procedure in which a secondary model of a receiver according to an embodiment of the present specification moves from a PREPARE state to an IDLE state.

[0072] Figure 34 illustrates a procedure for a secondary model of a receiver to move from an ACTIVE state to an IDLE state according to an embodiment of the present specification.

[0073] FIG. 35 illustrates a beam tracking task performed according to an embodiment of the present specification.

[0074] FIG. 36 is a flowchart illustrating a method performed by a first wireless device according to an embodiment of the present specification.

[0075] The following embodiments combine the components and features of this specification in a predetermined form. Each component or feature may be considered optional unless explicitly stated otherwise. Each component or feature may be implemented without being combined with other components or features. Furthermore, some components and / or features may be combined to form embodiments of this specification. The order of operations described in the embodiments of this specification may be changed. Some components or features of one embodiment may be included in another embodiment or may be replaced with corresponding components or features of another embodiment.

[0076] In the description of the drawings, procedures or steps that may obscure the gist of the present specification are not described, and procedures or steps that can be understood by a person skilled in the art are also not described.

[0077] Throughout the specification, when a part is said to "comprising" (or including) a certain component, this does not mean that other components are excluded, but rather that other components can be included, unless specifically stated otherwise. In addition, terms such as "...part," "...unit," and "module" described in the specification mean a unit that processes at least one function or operation, which may be implemented by hardware, software, or a combination of hardware and software. In addition, the words "a" or "an," "one," "the," and similar related words may be used in the context of describing this specification (especially in the context of the claims below) to include both the singular and the plural, unless otherwise indicated herein or clearly contradicted by context.

[0078] The embodiments of this specification have been described with a focus on the data transmission and reception relationship between a base station and a mobile station. Here, the base station is understood as a terminal node of a network that directly communicates with the mobile station. Certain operations described herein as being performed by the base station may, in some cases, be performed by an upper node of the base station.

[0079] That is, in a network consisting of multiple network nodes including a base station, various operations performed for communication with a mobile station may be performed by the base station or other network nodes other than the base station. In this case, the term 'base station' may be replaced by terms such as fixed station, Node B, eNB (eNode B), gNB (gNode B), ng-eNB, advanced base station (ABS), or access point.

[0080] Additionally, in the embodiments of the present specification, the term terminal may be replaced with terms such as user equipment (UE), mobile station (MS), subscriber station (SS), mobile subscriber station (MSS), mobile terminal, or advanced mobile station (AMS).

[0081] Additionally, a transmitter refers to a fixed and / or mobile node that provides data or voice services, and a receiver refers to a fixed and / or mobile node that receives data or voice services. Therefore, for uplink, a mobile station can be the transmitter, and a base station can be the receiver. Similarly, for downlink, a mobile station can be the receiver, and a base station can be the transmitter.

[0082] Embodiments of the present specification may be supported by standard documents disclosed in at least one of wireless access systems, such as IEEE 802.xx system, 3rd Generation Partnership Project (3GPP) system, 3GPP Long Term Evolution (LTE) system, 3GPP 5G (5th generation) NR (New Radio) system and 3GPP2 system, and in particular, embodiments of the present specification may be supported by 3GPP TS (technical specification) 38.211, 3GPP TS 38.212, 3GPP TS 38.213, 3GPP TS 38.321 and 3GPP TS 38.331 documents.

[0083] Furthermore, the embodiments of this specification may be applied to other wireless access systems and are not limited to the aforementioned systems. For example, they may also be applicable to systems implemented after the 3GPP 5G NR system, and are not limited to a specific system.

[0084] That is, obvious steps or parts not described in the embodiments of this specification may be explained by reference to the above documents. In addition, all terms disclosed in this specification may be explained by the above standard documents.

[0085] Hereinafter, preferred embodiments according to the present specification will be described in detail with reference to the accompanying drawings. The detailed description set forth below, together with the accompanying drawings, is intended to illustrate exemplary embodiments of the present specification and is not intended to represent the only embodiments in which the technical components of the present specification may be implemented.

[0086] Additionally, specific terms used in the embodiments of this specification are provided to aid in understanding of this specification, and the use of these specific terms may be changed to other forms without departing from the technical spirit of this specification.

[0087] The following technology can be applied to various wireless access systems such as CDMA (code division multiple access), FDMA (frequency division multiple access), TDMA (time division multiple access), OFDMA (orthogonal frequency division multiple access), and SC-FDMA (single carrier frequency division multiple access).

[0088] In order to make the following description clear, the following description is based on a 3GPP communication system (e.g., LTE, NR, etc.), but the technical idea of ​​the present invention is not limited thereto. LTE may refer to technology after 3GPP TS 36.xxx Release 8. Specifically, LTE technology after 3GPP TS 36.xxx Release 10 may be referred to as LTE-A, and LTE technology after 3GPP TS 36.xxx Release 13 may be referred to as LTE-A pro. 3GPP NR may refer to technology after TS 38.xxx Release 15. 3GPP 6G may refer to technology after TS Release 17 and / or Release 18. "xxx" refers to a standard document detail number. LTE / NR / 6G may be collectively referred to as a 3GPP system.

[0089] For background information, terms, abbreviations, etc. used in this specification, reference may be made to standard documents published prior to the invention of the present invention. For example, reference may be made to the 36.xxx and 38.xxx standard documents.

[0090] Communication systems applicable to this specification

[0091] Although not limited thereto, the various descriptions, functions, procedures, proposals, methods and / or operational flowcharts disclosed herein may be applied to various fields requiring wireless communication / connectivity (e.g., 5G) between devices.

[0092] Hereinafter, more specific examples will be provided with reference to the drawings. In the drawings / descriptions below, the same drawing reference numerals may represent identical or corresponding hardware blocks, software blocks, or functional blocks, unless otherwise described.

[0093] FIG. 1 is a diagram illustrating an example of a communication system applicable to the present specification. Referring to FIG. 1, a communication system (100) applicable to the present specification includes a wireless device, a base station, and a network. Here, a wireless device refers to a device that performs communication using a wireless access technology (e.g., 5G NR, LTE) and may be referred to as a communication / wireless / 5G device. Although not limited thereto, the wireless device may include a robot (100a), a vehicle (100b-1, 100b-2), an XR (extended reality) device (100c), a hand-held device (100d), a home appliance (100e), an IoT (Internet of Things) device (100f), and an AI (artificial intelligence) device / server (100g). For example, the vehicle may include a vehicle equipped with a wireless communication function, an autonomous vehicle, a vehicle capable of performing vehicle-to-vehicle communication, etc. Here, the vehicles (100b-1, 100b-2) may include unmanned aerial vehicles (UAVs) (e.g., drones). The XR devices (100c) include augmented reality (AR) / virtual reality (VR) / mixed reality (MR) devices, and may be implemented in the form of head-mounted devices (HMDs), head-up displays (HUDs) installed in vehicles, televisions, smartphones, computers, wearable devices, home appliances, digital signage, vehicles, robots, etc. The portable devices (100d) may include smartphones, smart pads, wearable devices (e.g., smartwatches, smart glasses), computers (e.g., laptops, etc.), etc. The home appliances (100e) may include TVs, refrigerators, washing machines, etc. The IoT devices (100f) may include sensors, smart meters, etc.For example, the base station (120) and the network (130) may also be implemented as wireless devices, and a specific wireless device (120a) may act as a base station / network node to other wireless devices.

[0094] Wireless devices (100a to 100f) can be connected to a network (130) via a base station (120). AI technology can be applied to the wireless devices (100a to 100f), and the wireless devices (100a to 100f) can be connected to an AI server (100g) via a network (130). The network (130) can be configured using a 3G network, a 4G (e.g., LTE) network, a 5G (e.g., NR) network, etc. The wireless devices (100a to 100f) can communicate with each other via the base station (120) / network (130), but can also communicate directly (e.g., sidelink communication) without going through the base station (120) / network (130). For example, vehicles (100b-1, 100b-2) can communicate directly (e.g., V2V (vehicle to vehicle) / V2X (vehicle to everything) communication). In addition, IoT devices (100f) (e.g., sensors) can communicate directly with other IoT devices (e.g., sensors) or other wireless devices (100a to 100f).

[0095] Wireless communication / connection (150a, 150b, 150c) can be established between wireless devices (100a to 100f) / base stations (120), and base stations (120) / base stations (120). Here, the wireless communication / connection can be established through various wireless access technologies (e.g., 5G NR) such as uplink / downlink communication (150a), sidelink communication (150b) (or D2D communication), and base station-to-base station communication (150c) (e.g., relay, IAB (integrated access backhaul)). Through the wireless communication / connection (150a, 150b, 150c), the wireless device and base station / wireless device, and base stations and base stations can transmit / receive wireless signals to / from each other. For example, the wireless communication / connection (150a, 150b, 150c) can transmit / receive signals through various physical channels. To this end, at least some of the various configuration information setting processes for transmitting / receiving wireless signals, various signal processing processes (e.g., channel encoding / decoding, modulation / demodulation, resource mapping / demapping, etc.), and resource allocation processes may be performed based on various proposals of this specification.

[0096] Communication systems applicable to this specification

[0097] FIG. 2 is a diagram illustrating an example of a wireless device applicable to this specification.

[0098] Referring to FIG. 2, the first wireless device (200a) and the second wireless device (200b) can transmit and receive wireless signals via various wireless access technologies (e.g., LTE, NR). Here, {the first wireless device (200a), the second wireless device (200b)} can correspond to {the wireless device (100x), the base station (120)} and / or {the wireless device (100x), the wireless device (100x)} of FIG. 1.

[0099] A first wireless device (200a) includes one or more processors (202a) and one or more memories (204a), and may further include one or more transceivers (206a) and / or one or more antennas (208a). The processor (202a) controls the memories (204a) and / or the transceivers (206a), and may be configured to implement the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed herein. For example, the processor (202a) may process information in the memory (204a) to generate first information / signals, and then transmit a wireless signal including the first information / signals via the transceivers (206a). In addition, the processor (202a) may receive a wireless signal including second information / signals via the transceivers (206a), and then store information obtained from signal processing of the second information / signals in the memory (204a). The memory (204a) may be connected to the processor (202a) and may store various information related to the operation of the processor (202a). For example, the memory (204a) may perform some or all of the processes controlled by the processor (202a), or may store software code including instructions for performing the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed herein. Here, the processor (202a) and the memory (204a) may be part of a communication modem / circuit / chip designed to implement wireless communication technology (e.g., LTE, NR). The transceiver (206a) may be connected to the processor (202a) and may transmit and / or receive wireless signals via one or more antennas (208a). The transceiver (206a) may include a transmitter and / or a receiver. The transceiver (206a) may be used interchangeably with an RF (radio frequency) unit. In this specification, wireless device may also mean a communication modem / circuit / chip.

[0100] The second wireless device (200b) includes one or more processors (202b), one or more memories (204b), and may further include one or more transceivers (206b) and / or one or more antennas (208b). The processor (202b) controls the memories (204b) and / or the transceivers (206b), and may be configured to implement the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed herein. For example, the processor (202b) may process information in the memory (204b) to generate third information / signals, and then transmit a wireless signal including the third information / signals via the transceivers (206b). In addition, the processor (202b) may receive a wireless signal including fourth information / signals via the transceivers (206b), and then store information obtained from signal processing of the fourth information / signals in the memory (204b). The memory (204b) may be connected to the processor (202b) and may store various information related to the operation of the processor (202b). For example, the memory (204b) may perform some or all of the processes controlled by the processor (202b), or may store software code including instructions for performing the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed herein. Here, the processor (202b) and the memory (204b) may be part of a communication modem / circuit / chip designed to implement wireless communication technology (e.g., LTE, NR). The transceiver (206b) may be connected to the processor (202b) and may transmit and / or receive wireless signals via one or more antennas (208b). The transceiver (206b) may include a transmitter and / or a receiver. The transceiver (206b) may be used interchangeably with an RF unit. In this specification, wireless device may also mean a communication modem / circuit / chip.

[0101] Hereinafter, hardware elements of the wireless device (200a, 200b) will be described in more detail. Although not limited thereto, one or more protocol layers may be implemented by one or more processors (202a, 202b). For example, one or more processors (202a, 202b) may implement one or more layers (e.g., functional layers such as physical (PHY), media access control (MAC), radio link control (RLC), packet data convergence protocol (PDCP), radio resource control (RRC), and service data adaptation protocol (SDAP)). One or more processors (202a, 202b) may generate one or more Protocol Data Units (PDUs) and / or one or more Service Data Units (SDUs) according to the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed herein. One or more processors (202a, 202b) may generate messages, control information, data or information according to the descriptions, functions, procedures, proposals, methods and / or operational flowcharts disclosed herein. One or more processors (202a, 202b) may generate signals (e.g., baseband signals) including PDUs, SDUs, messages, control information, data or information according to the functions, procedures, proposals and / or methods disclosed herein and provide the signals to one or more transceivers (206a, 206b). One or more processors (202a, 202b) may receive signals (e.g., baseband signals) from one or more transceivers (206a, 206b) and obtain PDUs, SDUs, messages, control information, data or information according to the descriptions, functions, procedures, proposals, methods and / or operational flowcharts disclosed herein.

[0102] One or more processors (202a, 202b) may be referred to as a controller, a microcontroller, a microprocessor, or a microcomputer. One or more processors (202a, 202b) may be implemented by hardware, firmware, software, or a combination thereof. For example, one or more application specific integrated circuits (ASICs), one or more digital signal processors (DSPs), one or more digital signal processing devices (DSPDs), one or more programmable logic devices (PLDs), or one or more field programmable gate arrays (FPGAs) may be included in one or more processors (202a, 202b). The descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed herein may be implemented using firmware or software, and the firmware or software may be implemented to include modules, procedures, functions, etc. The descriptions, functions, procedures, suggestions, methods and / or operation flowcharts disclosed in this specification may be implemented using firmware or software configured to perform one or more processors (202a, 202b) or stored in one or more memories (204a, 204b) and executed by one or more processors (202a, 202b). The descriptions, functions, procedures, suggestions, methods and / or operation flowcharts disclosed in this specification may be implemented using firmware or software in the form of codes, instructions and / or sets of instructions.

[0103] One or more memories (204a, 204b) may be coupled to one or more processors (202a, 202b) and may store various forms of data, signals, messages, information, programs, codes, instructions, and / or commands. The one or more memories (204a, 204b) may be configured as read only memory (ROM), random access memory (RAM), erasable programmable read only memory (EPROM), flash memory, hard drives, registers, cache memory, computer readable storage media, and / or combinations thereof. The one or more memories (204a, 204b) may be located internally and / or externally to the one or more processors (202a, 202b). Additionally, the one or more memories (204a, 204b) may be coupled to the one or more processors (202a, 202b) via various technologies, such as wired or wireless connections.

[0104] One or more transceivers (206a, 206b) can transmit user data, control information, wireless signals / channels, etc., as mentioned in the methods and / or flowcharts of this specification, to one or more other devices. One or more transceivers (206a, 206b) can receive user data, control information, wireless signals / channels, etc., as mentioned in the descriptions, functions, procedures, proposals, methods and / or flowcharts of this specification, from one or more other devices. For example, one or more transceivers (206a, 206b) can be coupled to one or more processors (202a, 202b) and can transmit and receive wireless signals. For example, one or more processors (202a, 202b) can control one or more transceivers (206a, 206b) to transmit user data, control information, or wireless signals to one or more other devices. Additionally, one or more processors (202a, 202b) may control one or more transceivers (206a, 206b) to receive user data, control information, or wireless signals from one or more other devices. Additionally, one or more transceivers (206a, 206b) may be coupled to one or more antennas (208a, 208b), and one or more transceivers (206a, 206b) may be configured to transmit and receive user data, control information, wireless signals / channels, or the like, as referred to in the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed herein, via one or more antennas (208a, 208b). In the present specification, one or more antennas may be multiple physical antennas or multiple logical antennas (e.g., antenna ports). One or more transceivers (206a, 206b) can convert received user data, control information, wireless signals / channels, etc. from RF band signals to baseband signals in order to process the received user data, control information, wireless signals / channels, etc. using one or more processors (202a, 202b).One or more transceivers (206a, 206b) may convert user data, control information, wireless signals / channels, etc. processed by one or more processors (202a, 202b) from baseband signals to RF band signals. For this purpose, one or more transceivers (206a, 206b) may include an (analog) oscillator and / or filter.

[0105] FIG. 3 is a diagram illustrating a method for processing a transmission signal applied to the present specification. For example, the transmission signal may be processed by a signal processing circuit. At this time, the signal processing circuit (300) may include a scrambler (310), a modulator (320), a layer mapper (330), a precoder (340), a resource mapper (350), and a signal generator (360). At this time, as an example, the operations / functions of FIG. 3 may be performed in the processors (202a, 202b) and / or the transceivers (206a, 206b) of FIG. 2. Furthermore, as an example, the hardware elements of FIG. 3 may be implemented in the processors (202a, 202b) and / or the transceivers (206a, 206b) of FIG. 2. For example, blocks 310 to 350 may be implemented in the processor (202a, 202b) of FIG. 2, and block 360 may be implemented in the transceiver (206a, 206b) of FIG. 2, and are not limited to the above-described embodiments.

[0106] The codeword can be converted into a wireless signal through the signal processing circuit (300) of FIG. 3. Here, the codeword is an encoded bit sequence of an information block. The information block may include a transport block (e.g., a UL-SCH transport block, a DL-SCH transport block). The wireless signal may be transmitted through various physical channels (e.g., a PUSCH, a PDSCH) of FIG. 6. Specifically, the codeword can be converted into a bit sequence scrambled by a scrambler (310). The scramble sequence used for scrambling is generated based on an initialization value, and the initialization value may include ID information of the wireless device, etc. The scrambled bit sequence can be modulated into a modulation symbol sequence by a modulator (320). The modulation scheme may include pi / 2-binary phase shift keying (pi / 2-BPSK), m-phase shift keying (m-PSK), m-quadrature amplitude modulation (m-QAM), etc.

[0107] A complex modulation symbol sequence can be mapped to one or more transmission layers by a layer mapper (330). The modulation symbols of each transmission layer can be mapped to the corresponding antenna port(s) by a precoder (340) (precoding). The output z of the precoder (340) can be obtained by multiplying the output y of the layer mapper (330) by an N*M precoding matrix W. Here, N is the number of antenna ports, and M is the number of transmission layers. Here, the precoder (340) can perform precoding after performing transform precoding (e.g., discrete Fourier transform (DFT) transform) on the complex modulation symbols. In addition, the precoder (340) can perform precoding without performing transform precoding.

[0108] The resource mapper (350) can map modulation symbols of each antenna port to time-frequency resources. The time-frequency resources can include multiple symbols (e.g., CP-OFDMA symbols, DFT-s-OFDMA symbols) in the time domain and multiple subcarriers in the frequency domain. The signal generator (360) generates a wireless signal from the mapped modulation symbols, and the generated wireless signal can be transmitted to another device through each antenna. To this end, the signal generator (360) can include an inverse fast Fourier transform (IFFT) module, a cyclic prefix (CP) inserter, a digital-to-analog converter (DAC), a frequency uplink converter, and the like.

[0109] The signal processing process for receiving signals in a wireless device can be configured in reverse order of the signal processing process (310-360) of FIG. 3. For example, a wireless device (e.g., 200a, 200b of FIG. 2) can receive wireless signals from the outside through an antenna port / transceiver. The received wireless signals can be converted into baseband signals through a signal restorer. For this purpose, the signal restorer can include a frequency downlink converter, an analog-to-digital converter (ADC), a CP remover, and a fast Fourier transform (FFT) module. Thereafter, the baseband signal can be restored to a codeword through a resource demapper process, a postcoding process, a demodulation process, and a descrambling process. The codewords can be restored to the original information blocks through decoding. Accordingly, a signal processing circuit (not shown) for a received signal may include a signal restorer, a resource de-mapper, a postcoder, a demodulator, a de-scrambler, and a decoder.

[0110] Wireless device structure applicable to this specification

[0111] FIG. 4 is a diagram illustrating another example of a wireless device to which the present specification applies.

[0112] Referring to FIG. 4, the wireless device (400) corresponds to the wireless devices (200a, 200b) of FIG. 2 and may be composed of various elements, components, units, and / or modules. For example, the wireless device (400) may include a communication unit (410), a control unit (420), a memory unit (430), and additional elements (440). The communication unit may include a communication circuit (412) and a transceiver(s) (414). For example, the communication circuit (412) may include one or more processors (202a, 202b) and / or one or more memories (204a, 204b) of FIG. 2. For example, the transceiver(s) (414) may include one or more transceivers (206a, 206b) and / or one or more antennas (208a, 208b) of FIG. 2. The control unit (420) is electrically connected to the communication unit (410), the memory unit (430), and the additional elements (440) and controls the overall operation of the wireless device. For example, the control unit (420) may control the electrical / mechanical operation of the wireless device based on the program / code / command / information stored in the memory unit (430). In addition, the control unit (420) may transmit information stored in the memory unit (430) to an external device (e.g., another communication device) via a wireless / wired interface through the communication unit (410), or store information received from an external device (e.g., another communication device) via a wireless / wired interface in the memory unit (430).

[0113] The additional element (440) may be configured in various ways depending on the type of the wireless device. For example, the additional element (440) may include at least one of a power unit / battery, an input / output unit, a driving unit, and a computing unit. Although not limited thereto, the wireless device (400) may be implemented in the form of a robot (Fig. 1, 100a), a vehicle (Fig. 1, 100b-1, 100b-2), an XR device (Fig. 1, 100c), a portable device (Fig. 1, 100d), a home appliance (Fig. 1, 100e), an IoT device (Fig. 1, 100f), a digital broadcasting terminal, a hologram device, a public safety device, an MTC device, a medical device, a fintech device (or a financial device), a security device, a climate / environmental device, an AI server / device (Fig. 1, 140), a base station (Fig. 1, 120), a network node, etc. Wireless devices may be mobile or stationary depending on the use / service.

[0114] In FIG. 4, various elements, components, units / parts, and / or modules within the wireless device (400) may be entirely interconnected via a wired interface, or at least some may be wirelessly connected via a communication unit (410). For example, within the wireless device (400), the control unit (420) and the communication unit (410) may be wired, and the control unit (420) and a first unit (e.g., 430, 440) may be wirelessly connected via the communication unit (410). In addition, each element, component, unit / part, and / or module within the wireless device (400) may further include one or more elements. For example, the control unit (420) may be composed of a set of one or more processors. For example, the control unit (420) may be composed of a set of a communication control processor, an application processor, an electronic control unit (ECU), a graphics processing processor, a memory control processor, etc. As another example, the memory unit (430) may be composed of RAM, DRAM (dynamic RAM), ROM, flash memory, volatile memory, non-volatile memory, and / or a combination thereof.

[0115] Mobile devices to which this specification applies

[0116] FIG. 5 is a drawing illustrating an example of a mobile device to which the present specification applies.

[0117] Figure 5 illustrates an example of a mobile device applicable to the present specification. The mobile device may include a smartphone, a smart pad, a wearable device (e.g., a smart watch, smart glasses), or a portable computer (e.g., a laptop, etc.). The mobile device may be referred to as a mobile station (MS), a user terminal (UT), a mobile subscriber station (MSS), a subscriber station (SS), an advanced mobile station (AMS), or a wireless terminal (WT).

[0118] Referring to FIG. 5, the portable device (500) may include an antenna unit (508), a communication unit (510), a control unit (520), a memory unit (530), a power supply unit (540a), an interface unit (540b), and an input / output unit (540c). The antenna unit (508) may be configured as a part of the communication unit (510). Blocks 510 to 530 / 540a to 540c correspond to blocks 410 to 430 / 440 of FIG. 4, respectively.

[0119] The communication unit (510) can transmit and receive signals (e.g., data, control signals, etc.) with other wireless devices and base stations. The control unit (520) can control components of the portable device (500) to perform various operations. The control unit (520) can include an AP (application processor). The memory unit (530) can store data / parameters / programs / codes / commands required for operating the portable device (500). In addition, the memory unit (530) can store input / output data / information, etc. The power supply unit (540a) supplies power to the portable device (500) and can include a wired / wireless charging circuit, a battery, etc. The interface unit (540b) can support connection between the portable device (500) and other external devices. The interface unit (540b) can include various ports (e.g., audio input / output ports, video input / output ports) for connection with external devices. The input / output unit (540c) can input or output video information / signals, audio information / signals, data, and / or information input from a user. The input / output unit (540c) may include a camera, a microphone, a user input unit, a display unit (540d), a speaker, and / or a haptic module.

[0120] For example, in the case of data communication, the input / output unit (540c) obtains information / signals (e.g., touch, text, voice, image, video) input by the user, and the obtained information / signals can be stored in the memory unit (530). The communication unit (510) can convert the information / signals stored in the memory into wireless signals, and transmit the converted wireless signals directly to other wireless devices or to a base station. In addition, the communication unit (510) can receive wireless signals from other wireless devices or base stations, and then restore the received wireless signals to the original information / signals. The restored information / signals can be stored in the memory unit (530) and then output in various forms (e.g., text, voice, image, video, haptic) through the input / output unit (540c).

[0121] Physical channels and general signal transmission

[0122] FIG. 6 is a diagram illustrating physical channels applicable to this specification and a signal transmission method using them.

[0123] When a terminal is powered on again from a powered-off state or newly enters a cell, it performs an initial cell search operation, such as synchronizing with the base station, in step S611. To this end, the terminal receives a primary synchronization channel (P-SCH) and a secondary synchronization channel (S-SCH) from the base station to synchronize with the base station and obtain information such as the cell ID.

[0124] After that, the terminal can obtain broadcast information within the cell by receiving a physical broadcast channel (PBCH) signal from the base station. Meanwhile, the terminal can check the downlink channel status by receiving a downlink reference signal (DL RS) in the initial cell search phase. After completing the initial cell search, the terminal can obtain more specific system information by receiving a physical downlink control channel (PDCCH) and a physical downlink shared channel (PDSCH) based on the physical downlink control channel information in step S612.

[0125] Thereafter, the terminal may perform a random access procedure such as steps S613 to S616 to complete connection to the base station. To this end, the terminal may transmit a preamble through a physical random access channel (PRACH) (S613) and receive a random access response (RAR) for the preamble through a physical downlink control channel and a physical downlink shared channel corresponding thereto (S614). The terminal may transmit a physical uplink shared channel (PUSCH) using scheduling information in the RAR (S615) and perform a contention resolution procedure such as receiving a physical downlink control channel signal and a physical downlink shared channel signal corresponding thereto (S616).

[0126] A terminal that has performed the procedure described above can then perform reception of a physical downlink control channel signal and / or a physical downlink shared channel signal (S617) and transmission of a physical uplink shared channel (PUSCH) signal and / or a physical uplink control channel (PUCCH) signal (S618) as a general uplink / downlink signal transmission procedure.

[0127] Control information transmitted from a terminal to a base station is collectively referred to as uplink control information (UCI). UCI includes hybrid automatic repeat and request acknowledgment / negative ACK (HARQ-ACK / NACK), scheduling request (SR), channel quality indication (CQI), precoding matrix indication (PMI), rank indication (RI), and beam indication (BI) information. UCI is generally transmitted periodically through PUCCH, but depending on the embodiment (e.g., when control information and traffic data must be transmitted simultaneously), it may be transmitted through PUSCH. In addition, the terminal may transmit UCI aperiodically through PUSCH upon request / instruction from the network.

[0128] 6G communication system

[0129] The 6G (wireless communication) system aims to achieve (i) very high data rates per device, (ii) a very large number of connected devices, (iii) global connectivity, (iv) very low latency, (v) low energy consumption for battery-free IoT devices, (vi) ultra-reliable connectivity, and (vii) connected intelligence with machine learning capabilities. The vision of the 6G system can be divided into four aspects: "intelligent connectivity," "deep connectivity," "holographic connectivity," and "ubiquitous connectivity," and the 6G system can satisfy the requirements as shown in Table 1 below. In other words, Table 1 is a table showing the requirements of the 6G system.

[0130]

[0131] At this time, the 6G system may have key factors such as enhanced mobile broadband (eMBB), ultra-reliable low latency communications (URLLC), massive machine type communications (mMTC), AI integrated communication, tactile internet, high throughput, high network capacity, high energy efficiency, low backhaul and access network congestion, and enhanced data security.

[0132] Core implementation technology of 6G systems

[0133] Artificial Intelligence

[0134] The most crucial and newly introduced technology for 6G systems is AI. 4G systems did not involve AI. 5G systems will support partial or very limited AI. However, 6G systems will fully support AI for automation. Advances in machine learning will create more intelligent networks for real-time communications in 6G. Incorporating AI into communications can streamline and improve real-time data transmission. AI can use numerous analyses to determine how complex target tasks should be performed. In other words, AI can increase efficiency and reduce processing delays.

[0135] Time-consuming tasks such as handover, network selection, and resource scheduling can be performed instantly using AI. AI can also play a crucial role in machine-to-machine (M2M), machine-to-human, and human-to-machine communications. Furthermore, AI can facilitate rapid communication in brain-computer interfaces (BCIs). AI-based communication systems can be supported by metamaterials, intelligent structures, intelligent networks, intelligent devices, intelligent cognitive radios, self-sustaining wireless networks, and machine learning.

[0136] Recent attempts to integrate AI into wireless communication systems have focused on the application layer, network layer, and especially deep learning in wireless resource management and allocation. However, this research is increasingly evolving to the MAC layer and physical layer, with attempts to combine deep learning with wireless transmission, particularly at the physical layer. AI-based physical layer transmission refers to the application of AI-based signal processing and communication mechanisms, rather than traditional communication frameworks, in the fundamental signal processing and communication mechanisms. For example, this may include deep learning-based channel coding and decoding, deep learning-based signal estimation and detection, deep learning-based MIMO mechanisms, and AI-based resource scheduling and allocation.

[0137] Machine learning can be used for channel estimation and channel tracking, as well as for power allocation and interference cancellation in the physical layer of the downlink (DL). Furthermore, machine learning can be used for antenna selection, power control, and symbol detection in MIMO systems.

[0138] Below, we will look at machine learning in more detail.

[0139] Machine learning refers to a series of operations that train machines to perform tasks that humans can or cannot perform. Machine learning requires data and a learning model. Data learning methods in machine learning can be broadly categorized into three types: supervised learning, unsupervised learning, and reinforcement learning.

[0140] Neural network training aims to minimize output errors. It involves repeatedly inputting training data into a neural network, calculating the neural network output and target error for the training data, and backpropagating the neural network error from the output layer to the input layer to update the weights of each node in the neural network to reduce the error.

[0141] Supervised learning uses labeled training data, while unsupervised learning may not have labeled training data. For example, in the case of supervised learning for data classification, the training data may be data in which each training data category is labeled. Labeled training data is input to a neural network, and the error can be calculated by comparing the output (categories) of the neural network with the training data labels. The calculated error is backpropagated through the neural network in the backward direction (i.e., from the output layer to the input layer), and the connection weights of each node in each layer of the neural network can be updated through backpropagation. The amount of change in the connection weights of each updated node can be determined by the learning rate. The neural network's calculation of the input data and the backpropagation of the error can constitute a learning cycle (epoch). The learning rate can be applied differently depending on the number of iterations of the neural network's learning cycle. For example, in the early stages of training a neural network, a high learning rate can be used to quickly allow the network to achieve a certain level of performance, thereby increasing efficiency. In the later stages of training, a low learning rate can be used to increase accuracy.

[0142] Learning methods may vary depending on the characteristics of the data. For example, if the goal is to accurately predict data transmitted by a transmitter in a communication system, supervised learning is preferable to unsupervised learning or reinforcement learning.

[0143] The learning model corresponds to the human brain, and the most basic linear model can be thought of, but the machine learning paradigm that uses highly complex neural network structures, such as artificial neural networks, as learning models is called deep learning.

[0144] The neural network cores used in learning methods are mainly divided into deep neural networks (DNN), convolutional deep neural networks (CNN), and recurrent boltzmann machines (RNN).

[0145] An artificial neural network is an example of a network of multiple perceptrons.

[0146] Referring to Fig. 7, when an input vector x=(x1,x2,...,xd) is input, the entire process of multiplying each component by a weight (W1,W2,...,Wd), adding up all the results, and then applying the activation function σ() is called a perceptron. A large-scale artificial neural network structure can extend the simplified perceptron structure illustrated in Fig. 7 to apply the input vector to perceptrons of different dimensions. For convenience of explanation, input values ​​or output values ​​are called nodes.

[0147] Meanwhile, the perceptron structure illustrated in Fig. 7 can be explained as consisting of a total of three layers based on input and output values. An artificial neural network in which there are H perceptrons of dimension (d+1) between the 1st layer and the 2nd layer, and K perceptrons of dimension (H+1) between the 2nd layer and the 3rd layer can be expressed as in Fig. 8.

[0148] The layer where the input vector is located is called the input layer, the layer where the final output value is located is called the output layer, and all layers located between the input layer and the output layer are called hidden layers. The example in Fig. 8 shows three layers, but when counting the number of layers in an actual artificial neural network, the input layer is excluded, so it can be viewed as a total of two layers. An artificial neural network is composed of perceptrons, which are basic blocks, connected in two dimensions.

[0149] The aforementioned input, hidden, and output layers can be applied jointly not only to multilayer perceptrons but also to various artificial neural network structures, such as CNNs and RNNs, which will be described later. The greater the number of hidden layers, the deeper the artificial neural network. The machine learning paradigm that uses sufficiently deep artificial neural networks as learning models is called deep learning. Furthermore, the artificial neural network used for deep learning is called a deep neural network (DNN).

[0150] The deep neural network illustrated in Figure 9 is a multilayer perceptron consisting of eight hidden layers and eight output layers. The multilayer perceptron structure is referred to as a fully-connected neural network. In a fully-connected neural network, there is no connection between nodes located in the same layer, and there is a connection only between nodes located in adjacent layers. DNN has a fully-connected neural network structure and is composed of a combination of multiple hidden layers and activation functions, and can be usefully applied to identify correlation characteristics between inputs and outputs. Here, the correlation characteristic can mean the joint probability of inputs and outputs. Figure 9 is a diagram illustrating an example of a deep neural network.

[0151] Meanwhile, depending on how multiple perceptrons are connected to each other, various artificial neural network structures different from the aforementioned DNN can be formed.

[0152] In DNN, nodes within a single layer are arranged vertically in a one-dimensional manner. However, Fig. 10 can assume a case where nodes are arranged two-dimensionally, with w nodes in width and h nodes in height (the convolutional neural network structure of Fig. 10). In this case, since a weight is added to each connection in the connection process from one input node to the hidden layer, a total of h×w weights must be considered. Since there are h×w nodes in the input layer, a total of h2w2 weights are required between two adjacent layers.

[0153] Figure 10 is a diagram showing an example of a convolutional neural network.

[0154] The convolutional neural network of Fig. 10 has a problem in that the number of weights increases exponentially according to the number of connections. Therefore, instead of considering the connections of all modes between adjacent layers, it assumes that there are small filters, and performs weighted sum and activation function operations on the overlapping portions of the filters, as in Fig. 10.

[0155] Each filter has a weight corresponding to its size, and weight learning can be performed to extract and output a specific feature on the image as a factor. In Fig. 10, a 3×3 filter is applied to the upper left 3×3 region of the input layer, and the output value resulting from performing weighted sum and activation function operations on the corresponding node is stored in z22.

[0156] The above filter performs weighted sum and activation function operations while moving horizontally and vertically at a certain interval while scanning the input layer, and places the output value at the current filter position. This operation method is similar to the convolution operation for images in the field of computer vision, so a deep neural network with this structure is called a convolutional neural network (CNN), and the hidden layer generated as a result of the convolution operation is called a convolutional layer. In addition, a neural network with multiple convolutional layers is called a deep convolutional neural network (DCNN).

[0157] Figure 11 is a diagram showing an example of a filter operation in a convolutional neural network.

[0158] In the convolutional layer, the number of weights can be reduced by calculating a weighted sum that includes only the nodes located in the area covered by the filter, starting from the node where the current filter is located. This allows a single filter to focus on features within a local area. Accordingly, CNNs can be effectively applied to image data processing where physical distance in a two-dimensional area is an important criterion for judgment. Meanwhile, CNNs can apply multiple filters immediately before the convolutional layer, and can generate multiple output results through the convolution operation of each filter.

[0159] Meanwhile, depending on the data properties, there may be data for which sequence characteristics are important. Considering the length variability and chronological relationship of such sequence data, a structure that applies a method of inputting one element of the data sequence at each timestep and inputting the output vector (hidden vector) of the hidden layer output at a specific timestep together with the immediately following element in the sequence is called a recurrent neural network structure.

[0160] Referring to Figure 12, a recurrent neural network (RNN) is a structure that inputs elements (x1(t), x2(t), ,..., xd(t)) of a data sequence at a time point t into a fully connected neural network, and then inputs the hidden vectors (z1(t-1), z2(t-1),..., zH(t-1)) of the immediately preceding time point t-1 together and applies a weighted sum and activation function. The reason for transmitting the hidden vector to the next time point in this way is because the information in the input vectors of the preceding time points is considered to be accumulated in the hidden vector of the current time point.

[0161] Figure 12 shows an example of a neural network structure in which a recurrent loop exists.

[0162] Referring to Figure 12, the recurrent neural network operates in a predetermined order of time for the input data sequence.

[0163] When the input vector (x1(t), x2(t), ,..., xd(t)) at time point 1 is input to the recurrent neural network, the hidden vector (z1(1), z2(1),..., zH(1)) is input together with the input vector (x1(2), x2(2),..., xd(2)) at time point 2, and the vector (z1(2), z2(2),..., zH(2)) of the hidden layer is determined through a weighted sum and an activation function. This process is repeatedly performed until time points 2, 3, ,,, T.

[0164] Figure 13 shows an example of the operating structure of a recurrent neural network.

[0165] Meanwhile, when multiple hidden layers are placed within a recurrent neural network, it is called a deep recurrent neural network (DRNN). Recurrent neural networks are designed to be useful for processing sequence data (e.g., natural language processing).

[0166] It is a neural network core used in a learning manner, and includes various deep learning techniques such as DNN, CNN, RNN, Restricted Boltzmann Machine (RBM), Deep Belief Network (DBN), and Deep Q-Network, and can be applied to fields such as computer vision, speech recognition, natural language processing, and speech / signal processing.

[0167] Figure 14 illustrates the functional framework of an AI / ML model.

[0168] In the framework according to Fig. 14, the definition of each term and the operation of each function can be based on Table 2 below.

[0169]

[0170]

[0171] Data Set

[0172] Data sets used in AI / ML are divided into training data, validation data, and test data, and their definitions are as follows.

[0173] - Training data

[0174] Data set for training the model

[0175] - Validation data

[0176] Data set for verifying a model whose learning has already been completed

[0177] Validation data is a data set usually used to prevent over-fitting of the training data set.

[0178] Validation data is a data set used to select the best model among the various models learned during the learning process. Therefore, validation data can be viewed as a data set related to learning.

[0179] - Test data

[0180] Data set for final evaluation. Test data is unrelated to learning.

[0181] For the above Data Set, the entire training set containing the above-described data in a certain proportion can be used.

[0182] For example, a training set containing training data and validation data in a ratio of 8:2 or 7:3 can be used.

[0183] For example, a training set containing training data, validation data, and test data in a ratio of 6:2:2 can be used.

[0184] Collaboration level

[0185] Depending on the capability of AI / ML functions between the base station and the terminal, the level of cooperation can be defined as shown in Table 3 below.

[0186]

[0187] The collaboration levels listed in Table 3 are merely examples and may be modified and utilized differently depending on the implementation method. For example, a collaboration level combining two or more of the illustrated collaboration levels may be defined and utilized. For example, a collaboration level(s) excluding one or more of the illustrated collaboration levels may be utilized.

[0188] AI-based communication (AI4C, AI for communication)

[0189] AI-based communications represent a shift from traditional mathematical model-based physical layer communications to data-driven AI / ML technologies. This shift is driven by advances in deep learning technology, which effectively approximates the actual data distributions being handled. By partially replacing or expanding the functions of physical layer communications, such as channel encoders, decoders, channel estimation and correction, synchronization, and waveform transmitters and receivers, AI / ML technologies can ultimately replace all existing communications.

[0190] Figure 15 illustrates a channel decoder including a neural network.

[0191] Referring to Figure 15, attempts to apply machine learning to the physical layer may primarily focus on optimizing specific functions of the receiver. For example, a neural network can be configured for decoding in a channel decoder. When the channel decoder is configured with a neural network, its performance can be improved. Typically, a signal transmitted from a transmitter passes through a wireless channel (H) and is delivered to the receiver as a received signal with added noise. Therefore, the receiver can decode the received signal with added noise to derive the actual signal. For example, existing channel decoders perform decoding according to a preset method, which can result in limited performance improvements. Conversely, configuring a neural network in the channel decoder can improve decoding performance through learning, thereby reducing delay and potentially improving performance.

[0192] As another example, if a MIMO detector is implemented as a neural network in a MIMO system with multiple transmit and receive antennas, the performance of the MIMO system can be improved. As another example, an autoencoder method can be applied. An autoencoder can be a method that performs end-to-end optimization by configuring both the transmitter and receiver as neural networks to improve performance. When communication is performed based on an autoencoder, the transmitter and receiver can each include a neural network. In this case, the transmitter and receiver can learn optimal communication settings and coding techniques, including the channel environment. The autoencoder can perform encoding and decoding using the information acquired through learning. Specifically, both the transmitter and receiver can include neural networks, and encoding and decoding can be considered together as a pair.

[0193] For example, when using an autoencoder, a communication system can form an optimized communication chain by considering the connection environment or situation. In other words, the communication system can create an optimal communication channel through the autoencoder by considering UE capabilities or channel characteristics. Based on the above, when the autoencoder operates, the communication system can perform communication in a shorter time with fewer resources. For example, in the case of high-capacity transmission (e.g., Tera-bps communication), the transmitter and receiver can reach an optimal communication environment even if the initial connection takes time, thereby reducing the probability of retransmission. Figure 16 illustrates an autoencoder method in which neural networks are configured at both the transmitter and receiver.

[0194] Figure 17 illustrates deep learning-based channel coding and decoding operations.

[0195] Referring to FIG. 17, channel coding and decoding can be performed based on deep learning. The transmitter can determine a channel to be applied based on deep learning and a coding scheme to be applied to the channel based on historical information about the channel environment. For example, the transmitter can transmit a reference signal and location information to the receiver. The receiver can generate an AI coding initial value using the reference signal and the location information of the transmitter and transmit it to the transmitter. The transmitter can determine a deep learning-based channel and a coding scheme to be applied to the channel based on the AI ​​coding initial value, perform encoding on the data, and transmit it to the receiver through the channel. The receiver can perform decoding to detect errors and generate reward information based on the detected error rate. For example, the receiver can generate a reward based on a single codeword. As another example, the receiver can generate a reward after receiving multiple blocks in the form of a batch and feed it back to the transmitter, but is not limited to the above-described embodiment. That is, after performing decoding, the receiver can feed back the decoding-based result information to the transmitter as reward information. Here, the transmitter can reflect the fed-back reward information in a deep learning-based algorithm and repeat the above-described process to select the channel and channel coding method most suitable for the channel environment.

[0196] Figure 18 illustrates learning performed in a wireless channel environment.

[0197] Referring to Figure 18, learning can be performed based on an artificial intelligence system in a wireless channel environment. For example, an artificial intelligence system agent can perform reinforcement learning based on an acquired reward value. Reinforcement learning can consist of two inputs and one output. More specifically, the two inputs can be state information and a reward value. In this case, the state information can be a factor acquired based on the wireless channel environment. For example, the state information can be information about an estimated channel. In addition, the reward value can be a value transmitted to the artificial intelligence system and updated and fed back based on the action of the artificial intelligence system agent. Thereafter, the artificial intelligence system agent can derive an action as an output based on the two inputs.

[0198] The descriptions / examples of FIGS. 1 to 18 described above can be applied to specify / implement the embodiments described below.

[0199] As an example, a device for performing a method according to an embodiment described below can be implemented by the devices of FIGS. 1 to 5.

[0200] For example, operations related to physical channel / signal transmission based on FIG. 6 may be performed in conjunction with operations of the embodiments described below.

[0201] For example, the AI / ML model according to the embodiment described below can be implemented based on the description of FIGS. 7 to 14.

[0202] As an example, a receiver / sender according to an embodiment described below may include a channel decoder and / or autoencoder based on FIGS. 15 and 16.

[0203] As an example, the channel coding and decoding operations according to FIG. 17 may be combined with the method according to the embodiment described below.

[0204] For example, learning according to the embodiment described below can be performed based on FIG. 18.

[0205] Below, we propose a device and operating method for robust online learning of deep neural network models that perform well in communications using AI / ML models. Before going into specifics, we briefly describe the operation and resulting effects of the embodiments described below.

[0206] AI / ML models that perform well but are difficult to interpret, such as deep neural network models, are defined as primary models. Secondary models, while not as powerful as neural networks, are trained with less data and are easily interpretable. If the primary model's performance deteriorates, a secondary model is trained using channel locality. Knowledge distillation from the primary model is performed on a surrogate model with the same architecture as the secondary model but different parameters. By comparing the surrogate model with the secondary model, the effectiveness of the deep neural network model is assessed and managed, thereby improving task continuity and resilience.

[0207] The problems of the existing method are specifically explained in relation to the embodiments described below.

[0208] 6G must support a wider variety of use cases than existing communications. For example, it will integrate communications between ground stations and satellites, between unmanned aerial vehicles (UAVs) and ground stations or UAVs, between vehicles, and support for IoT among small devices.

[0209] Wireless communications occur across diverse frequency ranges, including sub-6 GHz, millimeter-wave, and terahertz bands. As transmitter and receiver architectures diversify, wireless channel environments become increasingly complex. As the number of antennas increases, channel dimensionality increases. Channel probability distributions, due to their increased dimensionality and diversity, can exhibit long-tailed distributions.

[0210] As wireless communication technology embraces AI / ML technologies, it will be able to achieve even better communication performance in nonlinear channel environments. Among AI / ML models, deep neural networks demonstrate the highest performance. This is because they effectively learn and reflect nonlinear characteristics that conventional mathematical models fail to account for under the various scenarios described above.

[0211] AI / ML models are trained and operate based on data. However, as devices move, their performance will degrade when they encounter untrained communication channel environments. In particular, devices with limited memory and computational capacity have deep neural networks of limited complexity, increasing the likelihood of encountering new, untrained channel distributions as they move.

[0212] When deep neural networks encounter these new channels, they can cause communication quality degradation, such as data truncation. This degradation can be detected as a decrease in signal-to-noise ratio (SNR), bit error rate (BER), symbol error rate (SER), and block error rate (BER) when performing physical layer tasks. The aforementioned issues can arise when deep neural network models perform communication-related tasks. In particular, deep neural networks currently face the challenge of not fully understanding the causal relationships between their performance.

[0213] To address the above issues, we propose a method for operating a transmitter and receiver by introducing an explainable surrogate model that replaces the deep neural network model in the terminal. We propose a method for flexibly operating the surrogate model based on the deep neural network model's situation, thereby minimizing performance degradation.

[0214] The following describes the embodiments described above in the following order: 1) technical background, 2) concepts and structures related to the embodiments, 3) procedures related to the embodiments, and 4) specific operations when the embodiments are applied to tasks related to wireless communication.

[0215] The technical background of this embodiment is described in detail below.

[0216] Locality of communication channel environment

[0217] The regional nature of a communication channel environment refers to the distribution of communication propagation characteristics that are more distinctive across different regions. As frequencies increase, such as millimeter-wave and terahertz waves, propagation becomes more linear. These propagation characteristics, combined with diverse regional environments and diffraction, scattering, and reflection characteristics, create channel regionality, with unique distributions specific to each region. This characteristic results in long tails in the channel propagation probability distribution. For example, the TR 38.901 specification demonstrates the regional nature of a radio channel as shown below.

[0218] a) UMi (streets and building forests, open areas) with O2O (Outdoor-to-Outdoor) and O2I (Outdoor-to-Indoor)

[0219] b) UMa with O2O and O2I O2I

[0220] c) Indoor: Environment, shopping malls. Typical office environments include open cubicle areas and walled spaces.

[0221] d) Backhaul, including outdoor rooftop backhaul and street canyon scenarios with small cells in urban areas.

[0222] e) D2D / V2V. Device-to-device access in open spaces, canyons, and indoor scenarios. V2V is vehicle-to-vehicle communication.

[0223] f) Other scenarios such as stadiums (open roof) and gymnasiums (closed roof).

[0224] g) Indoor industry scenario

[0225] Locality can be defined as an area where the communication channel probability distribution is similar in the radio environment faced by a terminal. Locality can differentiate the propagation environment that affects the reference signal task in the millimeter-wave or terahertz bands. High-frequency propagation paths, unlike those in low-frequency bands, are more linear. These high-frequency propagation path distributions, which differ from low-frequency propagation in terms of reflection, diffraction, and refraction, have numerous variations, not only due to static buildings and terrain, but also due to moving objects, antenna placement, and terminal movement. The geographical distribution of a specific terminal and its dynamic state have the greatest impact on locality. This suggests that locality can be related to geographical distribution.

[0226] Figure 19 illustrates the tradeoff between explainability and performance in AI / ML. Referring to Figure 19, learning methods / AI / technologies with high performance have low explainability.

[0227] Explainable AI (XAI) technology enhances the understanding and interpretability of the behavior of AI / ML models. While deep neural network-based technologies have demonstrated outstanding performance, their internal behavior remains largely mathematically incomprehensible. Explainable AI combines mathematical and data-driven approaches to enhance this interpretability. Various explainable AI approaches exist. For example, one approach examines the output response to a combination of inputs to uncover correlations and relationships between inputs and outputs.

[0228] Complexity and explainability have a trade-off relationship. Deep neural networks are complex and perform well, but have poor explainability, while linear models are simple but struggle to explain nonlinear data, but are mathematically explainable.

[0229] Knowledge Distillation

[0230] In machine learning, knowledge distillation is the process of transferring knowledge from a large model to a smaller model. Large models (e.g., very deep neural networks or ensembles of many models) can be computationally expensive to evaluate or train than smaller models due to their larger capacity. Knowledge distillation transfers knowledge from a large model to a smaller model without significant loss of knowledge. The smaller model can be used to distill knowledge into an explainable model because it is much cheaper to train. Model cast Knowledge distillation can be performed using the loss function for knowledge distillation. is the knowledge distillation loss function (Knowledge Distillation loss) and cross entropy loss function (Cross Entropy loss) (Mathematical Formula 1 below). Knowledge distillation is a type of transfer learning, and other transfer learning methods are also possible.

[0231]

[0232] The proposed concept and structure related to this embodiment are described in detail below.

[0233] As described above, the present embodiment aims to ensure uninterrupted data communication when a deep neural network, a high-performance AI / ML model, experiences performance degradation. Specifically, the present invention proposes a method and device for resolving this issue by using a surrogate model and a secondary model that can learn and explain with a relatively small data set when the performance of a task between a transmitter and receiver operating as a primary model deteriorates.

[0234] Basic structure

[0235] The differences from the existing fallback operation are as follows:

[0236] The conventional fallback operation is performed as follows. A legacy model (i.e., a legacy communication model not based on an AI / ML model), a mathematically-based model specifically designed for a given task, is used to replace the deep neural network model. In contrast, the present embodiment does not utilize the legacy model. Specifically, the deep neural network model is replaced with an explainable model using a general-purpose, small training data set.

[0237] In this embodiment, the deep neural network model with the best performance but the most difficult to explain is defined as the primary model. If the performance of the primary model deteriorates, a data-driven, explainable, and less complex secondary model is utilized. This embodiment differs from the existing method of simply falling back to the legacy model in that it leverages the trade-off between performance and explainability.

[0238] The transmitter and receiver structures for implementing the proposed concept are described with reference to Fig. 20.

[0239] Figure 20 shows the structure of a sender and a receiver according to an embodiment of the present specification.

[0240] Referring to Figure 20, the sender (Sender, 2010) and the receiver (Receiver, 2020) may be two-sided models. In this case, only the sender (Sender, 2010) or the receiver (Receiver, 2020) may have an AI / ML model.

[0241] The primary model (2012, 2022) is primarily a deep neural network or an AI / ML model with high performance but low explainability. The surrogate model (2013, 2023) and secondary model (2014, 2024) are AI / ML models with average performance but high explainability, and are primarily composed of linear models. The surrogate model (2013, 2023) is the result of knowledge distillation or transfer learning of the primary model. The secondary model (2014, 2024) is a model that is trained online based on the current dataset, independently of the primary model (2012, 2022).

[0242] When the performance or loss falls below a specified level, the surrogate model that performed knowledge distillation or transfer learning of the primary model is compared with the secondary model that performed online learning based on the current data set. This allows for estimating and managing abnormalities in the current primary model's operation. The surrogate model and the secondary model are the same model, with only different parameter values. For example, the surrogate model can be expressed as y = ax + b, and the secondary model can be expressed as y = a'x + b'. Here, the parameters (a, b) of the surrogate model are different from the parameters (a', b') of the secondary model. As described below, the parameters of the secondary model and the surrogate model are mentioned in the SET_ACT condition (Equations 3 and 4) and the PRI_ACT condition (Equations 5 and 6). At this time, the comparison of the corresponding parameters (sender: , , recipient: , ) can correspond to a comparison of a and a' (and / or b and b').

[0243] The sender (2010) is the first model with the highest performance among AI / ML models for transmission. )(2012) and the surrogate model of the first model( )(2013) and the second model( )(2014) can be configured to include. The subscripts of each model are the parameters of that model.

[0244] The transmission data transmitted from the TX entity (2011) can be transmitted based on the first model (2012) and the second model (2014). At this time, the channel is transmitted based on one of the two models (2012, 2014). Transmission to can be performed.

[0245] The recipient (2020) is the first model, which is an AI / ML model. )(2022) and the proxy model of the first model( )(2023) and the second model( )(2024) can be configured to include. The subscripts of each model are the parameters of the model.

[0246] A signal arriving at the receiver (2020) can be input to both the primary model (2022) and the secondary model (2024), but only the task data of one of the primary model (2022) and the secondary model (2023) can be input to the RX entity (2021).

[0247] Data sets depending on the type of transmission / reception task And the role of each model may be different.

[0248] In case of data transmission and reception tasks, the primary models of the sender / receiver ( , )(2012, 2022) can also be the encoder and decoder of the autoencoder. In this case, and Since the transmitted and received data must be the same, they have the same value.

[0249] For channel estimation and data restoration tasks, the sender's first-order model ( )(2012) is a precoder on the transmitting side and the first-order model on the receiver ( )(2022) can be a channel estimator. At this time can be a vector consisting of the terminal state and the transmission reference signal, may be a channel estimation value.

[0250] For beam management tasks, the sender's primary model ( )(2012) can be a transmission beam selector. The first-order model of the receiver ( )(2022) may be a receiving beam selector. At this time, can be a vector consisting of the terminal state and the transmission beam signal, can be a receiving beam. The learning goal of the AI / ML model is a data set. Loss function for The goal is to find the parameters of the model. The learning objective of these AI / ML models can be expressed in the following mathematical expression 2.

[0251]

[0252] Equation 2 is the loss function for x and y Parameters of the first-order model of the sender (2012) with the minimum value and parameters of the receiver's first model (2022) It represents.

[0253] In this specification, the transmitter may refer to a first wireless device (200a or 200b of FIG. 2), a terminal, or a base station, and the receiver may refer to a second wireless device (200a or 200b of FIG. 2), a terminal, or a base station. For example, the transmitter may be a terminal and the receiver may be a base station. For example, the transmitter may be a base station and the receiver may be a terminal. For example, the transmitter may be a terminal and the receiver may be another terminal.

[0254] Channel locality

[0255] channel Training data from a set of terminal states in which the locality of the communication channel environment described above is maintained This locality is reflected in the state vector consisting of the terminal's position, posture, and motion vectors. and the status of the base station or receiver The channel is determined according to the combined channel distribution. can be explained using . Terminal status in case of a static receiving base station According to will be determined, so the channel distribution is can be expressed as . If , If given, depending on the region can have a log-normal distribution, a Rician distribution, or a special distribution depending on the terrain. Locality is the possible state of the terminal. set of can be defined as a set Any two belonging to , Is satisfies here is the Kullback-Leibler divergence, which is the distance between two distributions. can be determined by the capacity of the AI / ML model and the complexity of the channel distribution in the region.

[0256] When the performance of the deep neural network, which is the best performing primary model, falls below a specified monitoring threshold or the number of times it falls below the threshold within a window set to a specific timer exceeds a specified level, the sender and / or receiver can periodically perform knowledge distillation or transfer learning from the primary model (2012, 2022) to a surrogate model (2014, 2024). Learning of these surrogate models (2014, 2024) is related to the locality described above. It can be performed periodically to satisfy the requirement. The establishment of locality is a necessary condition for distilling the nonlinear part of the first-order model into a linear model of the second-order model. At the same time, the second-order model performs online learning based on the current data set.

[0257] The procedures and methods related to this embodiment are described in detail below.

[0258] Primary and secondary model capability procedure

[0259] Below, with reference to FIG. 21, a procedure for signaling capability related to whether the primary model and secondary model are supported is specifically described.

[0260] FIG. 21 is a flowchart illustrating a procedure for transmitting performance information related to whether or not support is provided for a primary-secondary AI / ML model according to an embodiment of the present specification.

[0261] Figure 21 is a process for exchanging information related to whether or not a primary-secondary AI / ML model that can support negotiation on a task between a sender and a receiver is supported.

[0262] In S2110, the sender requests performance information from the receiver regarding support for the first and second models. For example, the sender may send a primary-secondary model capability request message to the receiver.

[0263] In S2120, the receiver transmits performance information related to whether the first and second models are supported to the sender. For example, the receiver may transmit a primary-secondary model list to the sender as performance information related to whether the first and second models are supported.

[0264] As an example, the primary-secondary model list may include information based on Table 4 below.

[0265]

[0266] AI / ML models can be distinguished by identifiers. The identifiers can correspond to the type, architecture, various parameters, input features, and output feature structures of the AI / ML model. This information can be shared in advance between the receiver and the sender to conserve wireless resources. Subsequently, to convey information about supported AI / ML models, a list containing only the identifier(s) can be transmitted from the receiver to the sender. Although Figure 21 only illustrates the operation of the receiver transmitting the list information of supported models to the sender based on the sender's request, the operation of the sender transmitting the list information of supported models to the receiver based on the receiver's request can also be performed.

[0267] Setup procedure for primary and secondary model

[0268] Below, the setup procedure for the primary model and the secondary model is specifically described with reference to FIG. 22.

[0269] Figure 22 is a flowchart illustrating a procedure for setting up a primary-secondary AI / ML model according to an embodiment of the present specification.

[0270] Specifically, Figure 22 illustrates a procedure for setting a primary model and a secondary model between a sender and a receiver before a task related to wireless communication begins after the sender and receiver mutually confirm the AI / ML model support list in the primary secondary model capability procedure.

[0271] In S2210, the sender requests the receiver to set up the first / second model. For example, the sender may send a primary-secondary model setup request message to the receiver.

[0272] In S2220, the receiver transmits a message to the sender confirming the setup of the first / second model. For example, the receiver may transmit a primary-secondary model setup request confirm message to the sender.

[0273] As described in Fig. 21, the direction of signaling may be performed in the opposite direction to Fig. 22. For example, the sender may transmit a model setup confirmation message to the receiver based on the receiver's model setup request.

[0274] In the procedure of FIG. 22, the request message for setting up the first / second model or the message confirming the setting up of the first / second model may include information based on Table 5 below. In this case, some of the information based on Table 5 may be used in the first / second model online learning procedure (primary and secondary model online learning procedure).

[0275]

[0276] Release procedure for primary and secondary model

[0277] Below, the procedure for releasing the primary model and the secondary model is specifically described with reference to FIG. 23.

[0278] Figure 23 is a flowchart illustrating a procedure for releasing a primary-secondary AI / ML model according to an embodiment of the present specification. Specifically, Figure 23 illustrates a procedure for releasing the first / second models and all operations related to the first / second models.

[0279] In S2310, the Sender requests the Receiver to release the first / second model. For example, the Sender may send a primary-secondary model release request message to the Receiver.

[0280] In S2320, the receiver transmits a message to the sender confirming the release of the first / second model. For example, the receiver may transmit a primary-secondary model release confirm message to the sender.

[0281] As described in Figure 21, the direction of signaling may be performed in the opposite direction to Figure 23. For example, the sender may transmit a model release confirmation message to the receiver based on the receiver's model release request.

[0282] State of primary and secondary model

[0283] Below, each state and state transition of each model (primary model, surrogate model, secondary model) is described in detail.

[0284] The status of an AI / ML model can be IDLE, PREPARE, or ACTIVE.

[0285] Sender 1st model and receiver first-order model When a task operation starts, it moves to the ACTIVE state. At this time, the secondary model of the sender / receiver is in the IDLE state. An AI / ML model in the IDLE state does not perform any action.

[0286] Below, the state transition of each model is specifically described with reference to FIGS. 24 to 26.

[0287] Figure 24 illustrates state transitions of a secondary model according to an embodiment of the present specification. Figure 25 illustrates state transitions of a surrogate model according to an embodiment of the present specification. Figure 26 illustrates state transitions of a primary model according to an embodiment of the present specification.

[0288] State transition from IDLE to PREPARE for primary and secondary model

[0289] Referring to Figures 24 and 25, if the performance of the primary model is worse than PRETH (or if the loss is greater than PRETH, Loss > PRETH), the secondary model of the sender and the second model of the receiver and proxy model , moves from IDLE state to PREPARE state. This condition is SEC_PRE.

[0290] Conversely, if the performance of the primary model is better than PRETH (or the loss is equal to or less than PRETH, Loss ≥ PRETH), the sender's secondary model and the second model of the receiver and proxy model , moves from PREPARE state to IDLE state. This condition is SEC_IDL.

[0291] The secondary model performs online learning on the currently given task in the PREPARE state (Figure 24). Simultaneously, the surrogate model of the primary model performs knowledge distillation or transfer learning from the primary model (Figure 25).

[0292] State transition from PREPARE to ACTIVE for secondary model

[0293] Referring to Figure 24, the sender's secondary model and the receiver's secondary model can move to the ACTIVE state independently or together (in Figure 26, the primary model corresponding to each secondary model moves to the PREPARE state).

[0294] Referring to Figure 26, the primary model of the sender and the primary model of the receiver can move to the ACTIVE state independently or together (the secondary model corresponding to each primary model in Figure 24 moves to the IDLE state).

[0295] The secondary model performs actual task processing when it enters the ACTIVE state. The condition for the secondary model to move from the PREPARE state to the ACTIVE state (the condition for the primary model to move from the ACTIVE state to the PREPARE state) is SEC_ACT.

[0296] The SEC_ACT condition of the sender's second model can be expressed as in the following mathematical expression 3.

[0297]

[0298] In mathematical formula 3 is the offset required to determine whether the SEC_ACT condition is met. is training data The second model of the sender Through the channel The receiver's first model passes through is a composite function that processes the training data. The output of this function is the training data. and compare it with the loss function. Operator is the training data set Probability distribution for The average is calculated for the loss function over the training data set distribution. The final average loss is the average loss. As described above. is training data The sender's first model Through the channel The receiver's first model passes through It is a function that processes the .

[0299] In mathematical formula 3 The second model of the sender Wow, a proxy model It is the average value of the L2 Norm (Euclidean Distance) of the parameters. This value If this is the case (or If the second model and the surrogate model are greater than or equal to each other, the second model and the surrogate model can be judged as different models. It acts as a loss offset to reduce the sensitivity to fluctuations in loss values.

[0300] The SEC_ACT condition of the receiver's second model can be expressed as the following mathematical expression 4.

[0301]

[0302] In mathematical formula 4 is training data The sender's first model Through the channel The receiver's second model passes through is a composite function that processes the training data. The output of this function is the training data. and compare it with the loss function. Operator is the training data set Probability distribution for The average is calculated for the loss function over the training data set distribution. The final average loss is the average loss. As described above. is training data The sender's first model Through the channel The receiver's first model passes through It is a function that processes the .

[0303] In mathematical formula 4 Silver 2nd model Wow, a proxy model It is the average value of the L2 Norm (Euclidean Distance) of the parameters. This value In this case, the secondary model and the surrogate model can be judged as different models. As with the sender, It acts as a loss offset to reduce the sensitivity to fluctuations in loss values.

[0304] State transition from ACTIVE to PREPARE for primary model

[0305] Referring to Figure 26, based on the SEC_ACT condition described above, the primary model of the sender / receiver moves from the ACTIVE state to the PREPARE state.

[0306] State transition from PREPARE to ACTIVE for primary model

[0307] Referring to Figure 26, the primary models of the sender / receiver can independently or jointly transition from the PREPARE state to the ACTIVE state. At this time, the secondary model corresponding to each primary model transitions from the ACTIVE state to the IDLE state (Figure 24). The condition associated with this state transition is PRI_ACT.

[0308] The sender's PRI_ACT condition can be expressed as in the following mathematical expression 5.

[0309]

[0310] In mathematical equation 5 is an offset to prevent frequent state transitions due to large fluctuations in the measured value. The sender's PRI_ACT condition can be defined in contrast to the sender's SEC_ACT condition (Equation 3). Since the explanation of Equation 5 corresponds to Equation 3, redundant explanation is omitted.

[0311] The receiver's PRI_ACT condition can be expressed as in the following mathematical expression 6.

[0312]

[0313] The receiver's PRI_ACT condition can be defined in the opposite way to the receiver's SEC_ACT condition (Equation 4). The explanation of Equation 6 corresponds to Equation 4, so a duplicate explanation is omitted.

[0314] Hereinafter, the operations related to online learning and state transition of the primary model (and / or secondary model) of the sender / receiver are specifically described with reference to FIGS. 27 to 34.

[0315] Online learning procedure to PREPARE state of secondary model from IDLE for sender

[0316] Figure 27 illustrates a procedure for a secondary model of a transmitter to move from an IDLE state to a PREPARE state according to an embodiment of the present specification.

[0317] Referring to FIG. 27, in S2710, the primary model of the sender is in the ACTIVE state, and the secondary model of the sender is in the IDLE state.

[0318] When the performance of the primary model, which is a high-performance deep neural network, falls below the expected level (e.g., Loss < PRETH), a preparatory operation is performed to replace / change the model for performing the wireless communication-related task from the primary model to an explainable secondary model. That is, based on the performance of the primary model, the transmitter's secondary model can move from the IDLE state to the PREPARE state.

[0319] In S2720, the sender sends a SEC_PRE transition request message to the receiver. Since performance degradation is only known to the receiver through the loss function evaluation, the receiver monitors this condition and sends a SEC_PRE transition request message.

[0320] At S2730, the receiver sends a SEC_PRE transition confirm message to the sender.

[0321] In step S2740, the sender performs a state transition of the secondary model. Specifically, the sender's secondary model moves from the IDLE state to the PREPARE state. At this point, the receiver can freeze the parameters of its models before the sender initiates training of the secondary / surrogate models (Stop online learning).

[0322] In S2750, the sender performs online learning for the secondary model and knowledge distillation (or transfer learning) for the surrogate model, respectively. Specifically, the sender transmits training data for the learning to the receiver.

[0323] In S2760, the receiver feeds back the loss for the online learning to the sender.

[0324] Online learning procedure to ACTIVE state of secondary model from PREPARE for sender

[0325] Figure 28 illustrates a procedure in which a secondary model of a transmitter moves from a PREPARE state to an ACTIVE state according to an embodiment of the present specification.

[0326] Referring to Figure 28, at S2810, the sender's secondary model is in the PREPARE state. Although not shown in the drawing, the sender's primary model is in the ACTIVE state. At this time, the receiver suspends online learning of the primary model (and / or secondary model).

[0327] In step S2820, the sender determines whether the SEC_ACT condition is satisfied in the PREPARE state of the secondary model. Based on whether the SEC_ACT condition is satisfied, the sender transmits a request message for a state transition based on the SEC_ACT condition of the secondary model to the receiver. For example, the request message for a state transition based on the SEC_ACT condition of the secondary model may be a Request SEC_ACT transition for sender message that includes information that whether the model is changed is set to TRUE. The Request SEC_ACT transition for sender message may include parameters of the sender's proxy model and the secondary model.

[0328] In S2830, the receiver sends a confirmation message for the state transition based on the SEC_ACT condition of the secondary model (e.g., a Confirm SEC_ACT transition for sender message).

[0329] In S2840, the sender transitions the secondary model from the PREPARE state to the ACTIVE state. The sender transitions the primary model from the ACTIVE state to the PREPARE state. The receiver resumes online training of the primary model (and / or secondary model).

[0330] Online learning procedure to IDLE state of secondary model from PREPARE for sender

[0331] Figure 29 illustrates a procedure for a secondary model of a transmitter to move from a PREPARE state to an IDLE state according to an embodiment of the present specification.

[0332] Referring to Figure 29, at S2910, the sender's secondary model is in the PREPARE state. Although not shown in the drawing, the sender's primary model is in the ACTIVE state. At this time, the receiver suspends online learning of the primary model (and / or secondary model).

[0333] In step S2920, the sender determines whether the SEC_ACT condition is met in the PREPARE state of the secondary model. Based on the SEC_ACT condition not being met, the sender transmits a request message for a state transition based on the SEC_ACT condition of the secondary model to the receiver. For example, the request message for a state transition based on the SEC_ACT condition of the secondary model may be a Request SEC_ACT Transition for Sender message set to FALSE.

[0334] In S2930, the receiver sends a confirmation message for the state transition based on the SEC_ACT condition of the secondary model (e.g., a Confirm SEC_ACT transition for sender message).

[0335] At S2940, the sender transitions the secondary model from the PREPARE state to the IDLE state. The receiver resumes online training of the primary model (and / or secondary model).

[0336] Online learning procedure to ACTIVE state of primary model from PREPARE for sender (or to IDLE state of secondary model from ACTIVE for sender)

[0337] Figure 30 illustrates a procedure for a secondary model of a transmitter to move from an ACTIVE state to an IDLE state according to an embodiment of the present specification.

[0338] Referring to FIG. 30, in S3010, the primary model of the sender is in the PREPARE state, and the secondary model of the sender is in the ACTIVE state.

[0339] In S3020, the sender performs online training of the primary and secondary models. The sender transmits training data for the training to the receiver.

[0340] In S3030, the receiver feeds back the loss for the online learning to the sender.

[0341] In S3040, the sender evaluates whether the PRI_ACT condition is met (in the ACTIVE state of the secondary model). Based on whether the PRI_ACT condition is met, the sender sends a request message for a state transition based on the PRI_ACT condition of the primary model to the receiver. For example, the request message for a state transition based on the PRI_ACT condition of the primary model may be a Request PRI_ACT transition for sender message set to TRUE.

[0342] In S3050, the receiver sends a confirmation message (e.g., Confirm PRI_ACT transition for sender message) for a state transition based on the PRI_ACT condition of the primary model.

[0343] In S3060, the receiver transitions the primary model from the PREPARE state to the ACTIVE state. The receiver transitions the secondary model from the ACTIVE state to the IDLE state.

[0344] Online learning procedure to PREPARE state of secondary model from IDLE for receiver

[0345] Figure 31 illustrates a procedure for a secondary model of a receiver to move from an IDLE state to a PREPARE state according to an embodiment of the present specification.

[0346] When the performance of the primary model, which is a high-performance deep neural network, falls below the expected level, a preparatory operation is performed to replace it with an explainable secondary model. The procedures (S3110 to S3150) related to the operation of the receiver's secondary model transitioning from the IDLE state to the PREPARE state in FIG. 31 correspond to the procedures (S2710 to S2750) related to the operation of the sender's secondary model transitioning from the IDLE state to the PREPARE state in FIG. 27. Therefore, redundant explanation is omitted.

[0347] Online learning procedure to ACTIVE state of secondary model from PREPARE for receiver

[0348] Figure 32 illustrates a procedure in which a secondary model of a receiver according to an embodiment of the present specification moves from a PREPARE state to an ACTIVE state.

[0349] The procedures (S3210 to S3240) related to the operation of the receiver's secondary model transitioning from the PREPARE state to the ACTIVE state based on the SET_ACT condition being satisfied in FIG. 32 correspond to the procedures (S2810 to S2840) related to the operation of the sender's secondary model transitioning from the PREPARE state to the ACTIVE state based on the SET_ACT condition being satisfied in FIG. 28. Therefore, duplicate explanations are omitted.

[0350] Online learning procedure to IDLE state of secondary model from PREPARE for receiver

[0351] Figure 33 illustrates a procedure in which a secondary model of a receiver according to an embodiment of the present specification moves from a PREPARE state to an IDLE state.

[0352] The procedures (S3310 to S3340) related to the operation of the receiver's secondary model transitioning from the PREPARE state to the IDLE state based on the SET_ACT condition not being met in FIG. 33 correspond to the procedures (S2910 to S2940) related to the operation of the sender's secondary model transitioning from the PREPARE state to the IDLE state based on the SET_ACT condition not being met in FIG. 29. Therefore, redundant explanations are omitted.

[0353] Online learning procedure to ACTIVE state of primary model from PREPARE for receiver (or to IDLE state of secondary model from ACTIVE for receiver)

[0354] Figure 34 illustrates a procedure for a secondary model of a receiver to move from an ACTIVE state to an IDLE state according to an embodiment of the present specification.

[0355] In Fig. 34, the procedures (S3410 to S3450) related to the operation of the receiver's primary model transitioning from the PERPARE state to the ACTIVE state (the secondary model transitioning from the ACTIVE state to the IDLE state) based on the PRI_ACT condition being satisfied correspond to the procedures (S3010 to S3060) related to the operation of the sender's primary model transitioning from the PERPARE state to the ACTIVE state (the secondary model transitioning from the ACTIVE state to the IDLE state) based on the PRI_ACT condition being satisfied in Fig. 30. Therefore, duplicate explanations are omitted.

[0356] Priority of online learning procedure for sender and receiver task

[0357] In the operations according to the procedures of FIGS. 27 to 34 described above, the sender may be changed to the receiver and the receiver may be changed to the sender, and the operations according to the procedures may be performed. For example, in the operations according to the procedure of FIG. 27, the operations by the sender may be performed by the receiver, and the operations by the receiver may be performed by the sender.

[0358] In the operations according to the procedures of FIGS. 27 to 34 described above, operations related to the state transition of the receiver's model and operations related to the state transition of the sender's model may be performed simultaneously. In this case, the process / operation related to the state transition of the receiver (or sender) may be performed with priority.

[0359] Hereinafter, with reference to FIG. 35, specific operations are specifically described when the above-described embodiment is applied to a wireless communication-related task (beam tracking task).

[0360] FIG. 35 illustrates a beam tracking task performed according to an embodiment of the present specification.

[0361] Specifically, Fig. 35 illustrates a beam tracking task performed between a base station and a terminal in physical layer communication. An AI / ML regression task can be utilized to determine the distribution of predicted beam indices. It is assumed that the base station transmits a beam reference signal to the terminal to track the beam index. At this time, beam tracking can be divided into two categories: i) transmission beam tracking for determining a transmission beam on the transmitter side, and ii) reception beam tracking for determining a reception beam on the receiver side. The following description focuses on transmission beam tracking.

[0362] The primary AI / ML model that determines the transmission beam is based on a high-performance deep neural network, while the secondary model can be based on a linear regression model. Input features of the AI / ML model at the base station is the feature vector of the terminal, which consists of the terminal's position, velocity, and attitude. can be the distribution information of the predicted beam index. In this online learning, the training samples are assumed to be known to the terminal / base station through other methods.

[0363] Figure 35 (a) shows the vector in the first model Data on some elements of . The first model (deep neural network) shows nonlinear input and output relationships over a large area (i.e., low explainability). Specifically, the x symbol represents a vector Data about is the actual distribution, and the solid line is the corresponding data It shows the prediction of the first model for .

[0364] If online learning is performed so that the channel can have locality, a part of the global model (first model) expressed in (a) of Fig. 35 can be expressed as in (b) of Fig. 35 for the locality domain.

[0365] Figure 35 (b) shows the average response of the predicted beam to the movement of the terminal over time in an area where channel locality is established.

[0366] Based on the fulfillment of the aforementioned SEC_PRE condition, the secondary model and surrogate model of the base station move from the IDLE state to the PREPARE state. In the PREPARE state, training of the secondary model and the surrogate model (knowledge distillation based on the primary model) are performed. During this process, the performance of each model is monitored simultaneously with the training. In Figure 35 (b), the actual distribution is represented by the solid line (35B-1), and the prediction of the deep neural network model is represented by the dotted line (35B-2).

[0367] Here, i) the performance of the first model is significantly lower than that of the second model, and ii) the difference in parameters between the second model and the surrogate model is as large as the difference between the solid line (35B-1) and the dotted line (35B-2) in (b) of Fig. 35. The SEC_ACT condition can be satisfied if the loss of the first model is greater than the loss of the second model + is greater than (in Equation 3) , ii) the difference between the parameters of the second model and the parameters of the surrogate model The SEC_ACT condition can be satisfied if the value is greater than the value.

[0368] When the SEC_ACT condition is met, the base station changes the primary model activated for the beam tracking task to the secondary model. Specifically, the primary model of the base station transitions from the ACTIVE state to the PREPARE state, and the secondary model of the base station transitions from the PREPARE state to the ACTIVE state. The secondary model performs online learning and estimation for the solid line (35B-1) in Fig. 35(b). Since the learning of the secondary model is based on the learning of a linear model, there is an approximate upper limit to the amount of data required for training. Therefore, the learning of the secondary model can be performed quickly within the predicted time based on the amount of data.

[0369] A primary model with degraded performance (e.g., a deep neural network) can be restored by performing extended training with background training data or by downloading a deep neural network model tailored to the channel's locality. If the primary model's performance is restored (e.g., if the PRI_ACT condition based on Equation 5 is met), the primary model transitions from the PREPARE state to the ACTIVE state, and the secondary model transitions from the ACTIVE state to the IDLE state. At this time, the surrogate model can also transition from the IDLE state to the IDLE state according to the secondary model's state transition. Based on the state transitions of the aforementioned models, the primary model can resume performing the transmit beam tracking task.

[0370] For convenience of explanation in Fig. 35, the transmission beam tracking task is described as the standard, but the above-described embodiment can also be extended and applied to the reception beam tracking task. Furthermore, the above-described embodiment can also be applied to other tasks of the physical layer other than beam tracking cast.

[0371] The operations and effects according to the above-described embodiments are summarized as follows.

[0372] When a high-performance neural network, which is the primary AI / ML model, encounters a channel or communication environment that it has not been able to learn, a secondary model with high explanatory power is trained by utilizing the locality of the channel, and knowledge distillation or transfer learning of the primary model is performed on the same model, a surrogate model.

[0373] If model replacement is performed solely based on the temporary performance degradation or restoration of the primary model, frequent model replacement operations can further increase wireless communication latency. In this embodiment, rather than simply switching models when the primary model degrades or recovers, the model activated for a task is replaced from the primary model (secondary model) to the secondary model (primary model) through a comparison with the secondary model and a surrogate model. Therefore, online learning and tasks can be performed continuously while improving the efficiency of model replacement based on more reliable judgment.

[0374] While the secondary model is active and operating, the primary model can successfully perform online learning in the background. When the primary model's performance returns to normal, the primary model resumes performing the task in place of the secondary model. This approach ensures optimal performance for the physical layer's transmit and receive tasks.

[0375] In terms of implementation, the operations according to the embodiments described above (e.g., operations for supporting semantic communication) can be processed by the devices of FIGS. 1 to 5 described above (e.g., processors (202a, 202b) of FIG. 2).

[0376] In addition, the operations (e.g., operations for supporting semantic communication) according to the above-described embodiments may be stored in a memory (e.g., 204a, 204b of FIG. 2) in the form of commands / programs (e.g., instructions, executable codes) for driving at least one processor (e.g., processors (202a, 202b) of FIG. 2).

[0377] The embodiments described below are specifically described with reference to FIG. 36 in terms of the operation of wireless devices (e.g., the first wireless device (200a) and the second wireless device (200b) of FIG. 2). The methods described below are distinguished merely for the convenience of explanation, and it goes without saying that some components of one method may be substituted for some components of another method or may be applied in combination with each other.

[0378] FIG. 36 is a flowchart illustrating a method performed by a first wireless device according to an embodiment of the present specification.

[0379] Referring to FIG. 36, a method performed by a first wireless device in a wireless communication system supporting an AI / ML (Artificial Intelligence / Machine Learning) model according to one embodiment of the present specification includes a step (S3610) of performing learning of a surrogate model and learning of a secondary model based on a value related to a loss of a primary model, and a step (S3620) of changing an activated model from a primary model to a secondary model based on a difference value between a parameter of the surrogate model and a parameter of the secondary model.

[0380] In S3610, the first wireless device performs training of a surrogate model and training of a second model based on a value related to the loss of the first model being greater than a first threshold.

[0381] The above primary model may be an activated model for performing a task related to wireless communication. The surrogate model and the secondary model may have the same model type and structure, but may have different learning-based parameters. In one embodiment, the task related to wireless communication may be related to at least one of reporting channel state information (CSI), measurement of a reference signal, signal modulation, beam tracking, or beam search. However, the task related to wireless communication is not limited to the tasks listed above. The task related to wireless communication may be related to one or more operations performed for each layer (e.g., physical, MAC, and / or RRC) related to wireless communication.

[0382] The learning of the above-mentioned surrogate model may be based on transfer learning of the first model. In one embodiment, the transfer learning may involve knowledge distillation. The learning of the second model may be online learning based on the wireless signal currently being transmitted and received.

[0383] In one embodiment, the training of the surrogate model and the secondary model may be performed in the PREPARE state. Specifically, based on a value related to the performance of the primary model being greater than the first threshold, the state related to the operation of the surrogate model and the secondary model may transition from the IDLE state to the PREPARE state. Training of the surrogate model and training of the secondary model may be performed in the PREPARE state.

[0384] In one embodiment, the first model may include a Deep Neural Network (DNN) model. The first model is not limited to the DNN model and may include models with low explainability as shown in FIG. 19. For example, the first model may include a model based on a support vector machine (SVM). For example, the first model may include a model based on a random forest. For example, the first model may include a model that performs better than a linear model but has lower explainability. The first model is not limited to the models listed above. The first model may refer to a model that performs better but has lower explainability than the second model described below.

[0385] As described below, before changing the model activated for the wireless communication-related task from the primary model to the secondary model, training for the secondary model must be completed. Since the secondary model has a higher explainability than the primary model, the amount of data required for training may be limited. In one embodiment, the size of the training data for training the secondary model may be smaller than the size of the training data for training the primary model.

[0386] In one embodiment, the training data for learning the secondary model may include i) an input and ii) an output corresponding to the input. The input may include a vector related to at least one of a position, a velocity, or a posture of the first wireless device. The output may include information determined based on the vector. The information determined based on the vector may include an index of a beam related to transmission or reception of the first wireless device. For example, the information determined based on the vector may be related to a reception beam (spatial domain receive filter) of the first wireless device or a reception beam (spatial domain transmission filter) of the first wireless device. In this case, the task related to the wireless communication may be a task related to beam tracking of the first wireless device.

[0387] In S3620, the first wireless device changes the model activated for performing the task related to the wireless communication from the first model to the second model based on a value representing the difference between the parameters of the proxy model and the parameters of the second model.

[0388] In one embodiment, the secondary model may include a linear model. The primary model is not limited to the linear model and may include models with high explainability as shown in FIG. 19. For example, the secondary model may include a model based on a decision tree. The secondary model is not limited to the models listed above. The secondary model may refer to a model with higher explainability than the primary model. For example, the secondary model may include a linear model, a kernel model, a graph model, and / or a probabilistic model, excluding a deep neural network.

[0389] In one embodiment, the explainability of the first model may be lower than the explainability of the second model.

[0390] In one embodiment, the value representing the difference between the parameters of the surrogate model and the parameters of the secondary model may be an average value calculated based on the Euclidean distance between the parameters of the surrogate model and the parameters of the secondary model. The average value may be based on mathematical expression 3 (or mathematical expression 4). The Euclidean distance may mean the L2 Norm of mathematical expression 3 (or mathematical expression 4).

[0391] In one embodiment, the change (1st -> 2nd) of the activated model described above can be performed based on the satisfaction of the SEC_ACT condition. This is described in detail below.

[0392] i) a value related to the loss of the first model is greater than a value related to the loss of the second model plus an offset value, and ii) a value representing a difference between a parameter of the surrogate model and a parameter of the second model is greater than a second threshold value:

[0393] The model activated for performing the task related to the above wireless communication may be changed from the first model to the second model. This embodiment may be based on mathematical expression 3 (or mathematical expression 4).

[0394] Based on the above SEC_ACT condition being satisfied, the states related to the operation of the first / second model can be transitioned as follows.

[0395] The state related to the operation of the above first model can transition from the ACTIVE state to the PREPARE state. The state related to the operation of the above second model can transition from the PREPARE state to the ACTIVE state.

[0396] Given that the performance of the primary model, which has a low explanatory probability, is better than that of the secondary model, if the performance of the primary model is restored after the activated model is changed, it is necessary to perform wireless communication-related tasks based on that model again. The operation of changing the activated model from the secondary model to the primary model can be performed based on the PRI_ACT condition being met. This will be described in detail below.

[0397] In one embodiment, based on i) a value associated with the loss of the secondary model being greater than a value associated with the loss of the primary model plus an offset value, and ii) a value representing a difference between a parameter of the surrogate model and a parameter of the secondary model being less than a second threshold value:

[0398] The model activated for performing the task related to the above wireless communication may be changed from the secondary model to the primary model. This embodiment may be based on mathematical expression 5 (or mathematical expression 6).

[0399] Based on the above PRI_ACT condition being satisfied, the states related to the operation of the first / second model can be transitioned as follows.

[0400] The state related to the operation of the above first model can transition from the PREPARE state to the ACTIVE state. The state related to the operation of the above second model can transition from the ACTIVE state to the IDLE state.

[0401] In one embodiment, the method may further comprise transmitting a message related to a change in the activated model to the second wireless device. The message may include parameters of the surrogate model and parameters of the secondary model.

[0402] In one embodiment, the first wireless device may be a sender (e.g., 2010 of FIG. 20, FIGS. 21 to 23, and FIGS. 27 to 34), a receiver (e.g., 2020 of FIG. 20, FIGS. 21 to 23, and FIGS. 27 to 34), a user equipment (UE), or a base station. The second wireless device may be a receiver (e.g., 2020 of FIG. 20, FIGS. 21 to 23, and FIGS. 27 to 34), a sender (e.g., 2010 of FIG. 20, FIGS. 21 to 23, and FIGS. 27 to 34), a base station, or a user equipment (UE).

[0403] In one embodiment, the method may further include a signaling procedure (S2110 and / or S2120) related to the performance of the first wireless device based on FIG. 21. S2110 and / or S2120 may be performed before S3610. For example, (based on a request from the second wireless device), the first wireless device may transmit information related to the performance of the first wireless device to the second wireless device. The information related to the performance of the first wireless device may include information related to a primary model and / or a secondary model supported by the first wireless device (e.g., Table 4).

[0404] In one embodiment, the method may further include signaling procedures (S2210 and / or S2220) related to model setup based on FIG. 22. S2210 and / or S2220 may be performed prior to S3610. For example, (based on a request from the first wireless device) the first wireless device may receive a confirmation message related to the model setup of the first wireless device from the second wireless device.

[0405] In one embodiment, the method may further include a signaling procedure (S2310 and / or S2320) related to model release based on FIG. 23. S2310 and / or S2320 may be performed prior to S3610. For example, (based on a request from the first wireless device) the first wireless device may receive a confirmation message related to model setup of the first wireless device from the second wireless device.

[0406] The operations based on the above-described S3610 and S3620 and FIGS. 21 to 23 can be implemented by the devices of FIGS. 1 to 5. For example, a terminal (e.g., 200a of FIG. 2) can control one or more transceivers (206a) and / or one or more memories (204a) to perform the operations based on the above-described S3610 and S3620 and FIGS. 21 to 23. For example, a base station (e.g., 200b of FIG. 2) can control one or more transceivers (206b) and / or one or more memories (204b) to perform the operations based on the above-described S3610 and S3620 and FIGS. 21 to 23.

[0407] Here, the wireless communication technology implemented in the wireless devices (200a, 200b) of the present specification may include not only LTE, NR, and 6G, but also Narrowband Internet of Things for low-power communication. At this time, for example, NB-IoT technology may be an example of LPWAN (Low Power Wide Area Network) technology, and may be implemented with standards such as LTE Cat NB1 and / or LTE Cat NB2, and is not limited to the above-described names. Additionally or alternatively, the wireless communication technology implemented in the wireless devices (XXX, YYY) of the present specification may perform communication based on LTE-M technology. At this time, for example, LTE-M technology may be an example of LPWAN technology, and may be called by various names such as eMTC (enhanced Machine Type Communication). For example, LTE-M technology can be implemented by at least one of various standards such as 1) LTE CAT 0, 2) LTE Cat M1, 3) LTE Cat M2, 4) LTE non-BL (non-Bandwidth Limited), 5) LTE-MTC, 6) LTE Machine Type Communication, and / or 7) LTE M, and is not limited to the above-described names. Additionally or alternatively, the wireless communication technology implemented in the wireless device (XXX, YYY) of the present specification can include at least one of ZigBee, Bluetooth, and Low Power Wide Area Network (LPWAN) considering low-power communication, and is not limited to the above-described names. For example, ZigBee technology can create PAN (personal area networks) related to small / low-power digital communication based on various standards such as IEEE 802.15.4, and can be called by various names.

[0408] The embodiments described above are combinations of the components and features of the present disclosure in a predetermined form. Each component or feature should be considered optional unless explicitly stated otherwise. Each component or feature may be implemented without being combined with other components or features. Furthermore, it is also possible to combine some components and / or features to form an embodiment of the present disclosure. The order of operations described in the embodiments of the present disclosure may be changed. Some components or features of one embodiment may be included in another embodiment or may be replaced with corresponding components or features of another embodiment. It is self-evident that claims that are not explicitly cited in the scope of the patent may be combined to form an embodiment or may be incorporated as a new claim through a post-application amendment.

[0409] Embodiments according to the present specification may be implemented by various means, for example, hardware, firmware, software, or a combination thereof. In the case of hardware implementation, an embodiment of the present invention may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, etc.

[0410] When implemented via firmware or software, an embodiment of the present specification may be implemented in the form of a module, procedure, function, or the like that performs the functions or operations described above. The software code may be stored in memory and executed by a processor. The memory may be located within or external to the processor and may exchange data with the processor via various known means.

[0411] It will be apparent to those skilled in the art that this specification may be embodied in other specific forms without departing from the essential characteristics thereof. Therefore, the foregoing detailed description should not be construed in any way as limiting but rather as illustrative. The scope of this specification should be determined by a reasonable interpretation of the appended claims, and all changes within the scope of equivalents herein are intended to be included within the scope of this specification.

Claims

1. A method performed by a first wireless device in a wireless communication system supporting an AI / ML (Artificial Intelligence / Machine Learning) model, A step of performing training of a surrogate model and training of a secondary model based on a value related to the loss of the primary model being greater than a first threshold value; The above first model is a model activated to perform a task related to wireless communication, and the above proxy model and the second model are equivalent in model type and structure, but have different learning-based parameters. The learning of the above proxy model is based on the transfer learning of the above first model, and the learning of the above second model is online learning based on the wireless signal currently being transmitted and received; and A method comprising: a step of changing a model activated for performing a task related to wireless communication from the primary model to the secondary model based on a value representing a difference between parameters of the surrogate model and parameters of the secondary model; 2. In paragraph 1, A method characterized in that the above first model includes a deep neural network (DNN) model.

3. In paragraph 1, A method characterized in that the above secondary model includes a linear model, a kernel model, a graph model and / or a probabilistic model excluding a deep neural network.

4. In paragraph 1, A method characterized in that the explainability of the first model is lower than the explainability of the second model.

5. In paragraph 1, A method characterized in that the value representing the difference between the parameters of the surrogate model and the parameters of the secondary model is an average value calculated based on the Euclidean distance between the parameters of the surrogate model and the parameters of the secondary model.

6. In paragraph 1, i) a value related to the loss of the first model is greater than a value related to the loss of the second model plus an offset value, and ii) a value representing a difference between a parameter of the surrogate model and a parameter of the second model is greater than a second threshold value: A method characterized in that the model activated for performing a task related to the wireless communication is changed from the primary model to the secondary model.

7. In paragraph 1, Further comprising the step of transmitting a message related to a change in the activated model to the second wireless device, A method characterized in that the message includes parameters of the surrogate model and parameters of the secondary model.

8. In paragraph 6, A method characterized in that a state related to the operation of the first model transitions from an ACTIVE state to a PREPARE state, and a state related to the operation of the second model transitions from a PREPARE state to an ACTIVE state.

9. In paragraph 1, A method characterized in that, based on a value related to the performance of the first model being greater than the first threshold value, a state related to the operation of the surrogate model and the second model transitions from an IDLE state to a PREPARE state.

10. In paragraph 9, A method characterized in that the learning of the above proxy model and the learning of the second model are performed in the PREPARE state.

11. In paragraph 1, A method characterized in that the above transfer learning is related to knowledge distillation.

12. In paragraph 1, A method characterized in that the task related to the wireless communication is related to at least one of reporting of channel state information, measurement of reference signal, signal modulation, beam tracking, or beam search.

13. In paragraph 1, A method characterized in that the size of the training data for learning the second model is smaller than the size of the training data for learning the first model.

14. In paragraph 13, The training data for learning the above second model includes i) input and ii) output corresponding to the input, The above input comprises a vector related to at least one of the position, velocity or attitude of the first wireless device, A method characterized in that the above output includes information determined based on the above vector.

15. In paragraph 14, A method characterized in that the information determined based on the above vector includes an index of a beam related to transmission or reception of the first wireless device.

16. In paragraph 1, i) a value related to the loss of the second model is greater than a value obtained by adding an offset value to a value related to the loss of the first model, and ii) a value representing a difference between a parameter of the surrogate model and a parameter of the second model is less than a second threshold value; A method characterized in that the model activated for performing the task related to the wireless communication is changed from the secondary model to the primary model.

17. In paragraph 16, A method characterized in that a state related to the operation of the first model transitions from a PREPARE state to an ACTIVE state, and a state related to the operation of the second model transitions from an ACTIVE state to an IDLE state.

18. In paragraph 1, A method characterized in that the first wireless device is a sender, a receiver, a terminal or a base station.

19. In a first wireless device operating in a wireless communication system supporting an AI / ML (Artificial Intelligence / Machine Learning) model, One or more transmitters and receivers; One or more processors controlling the one or more transceivers; and one or more memories operably connectable to said one or more processors and storing instructions that, when executed by said one or more processors, cause said one or more processors to perform operations; The above actions are, A step of performing training of a surrogate model and training of a secondary model based on a value related to the loss of the primary model being greater than a first threshold value; The above first model is a model activated to perform a task related to wireless communication, and the above proxy model and the second model are equivalent in model type and structure, but have different learning-based parameters. The learning of the above proxy model is based on the transfer learning of the above first model, and the learning of the above second model is online learning based on the wireless signal currently being transmitted and received; and A first wireless device comprising: a step of changing a model activated for performing a task related to wireless communication from the primary model to the secondary model based on a value representing a difference between parameters of the surrogate model and parameters of the secondary model; 20. In a device comprising one or more memories and one or more processors functionally connected to the one or more memories, The one or more memories store instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, The above actions are, A step of performing training of a surrogate model and training of a secondary model based on a value related to the loss of the primary model being greater than a first threshold value; The above first model is a model activated to perform a task related to wireless communication, and the above proxy model and the second model are equivalent in model type and structure, but have different learning-based parameters. The learning of the above proxy model is based on the transfer learning of the above first model, and the learning of the above second model is online learning based on the wireless signal currently being transmitted and received; and A device comprising: a step of changing a model activated for performing a task related to wireless communication from the primary model to the secondary model based on a value representing a difference between parameters of the surrogate model and parameters of the secondary model; 21. In one or more non-transitory computer-readable media storing one or more instructions, The one or more instructions, when executed by one or more processors, cause the one or more processors to perform operations; The above actions are, A step of performing training of a surrogate model and training of a secondary model based on a value related to the loss of the primary model being greater than a first threshold value; The above first model is a model activated to perform a task related to wireless communication, and the above proxy model and the second model are equivalent in model type and structure, but have different learning-based parameters. The learning of the above proxy model is based on the transfer learning of the above first model, and the learning of the above second model is online learning based on the wireless signal currently being transmitted and received; and A non-transitory computer-readable medium comprising: a step of changing a model activated for performing a task related to wireless communication from the primary model to the secondary model based on a value representing a difference between parameters of the surrogate model and parameters of the secondary model;

Images