Terminal, base station, and communication method

Abstract

The present invention improves the efficiency of radio communication. This terminal is provided with: a control circuit that, on the basis of a first rank value at the current time point and a second rank value at the immediately previous time point, controls input, to a terminal-side artificial intelligence model that generates a report of channel state information, of an intermediate output from the terminal-side artificial intelligence model at a time point in the past; and a transmission circuit that transmits the report of the channel state information generated using the intermediate output.

Description

Terminal, base station and communication method

[0001] The present disclosure relates to a terminal, a base station, and a communication method.

[0002] In recent years, the expansion and diversification of wireless services has led to the expectation of rapid development of the Internet of Things (IoT). Mobile communications are now being used in a wide range of applications, from smartphones and other information terminals to automobiles, homes, home appliances, and industrial equipment. To support this diversification, significant improvements in the performance and functionality of mobile communication systems are required, addressing various requirements, such as increased system capacity, an increased number of connected devices, and low latency. Fifth-generation mobile communication systems (5G) boast high-capacity and ultra-high-speed data transfer (eMBB: enhanced Mobile Broadband), massive machine-type communication (mMTC: massive Machine-Type Communication), and ultra-reliable and low-latency communication (URLLC), providing flexible wireless communications to meet diverse needs.

[0003] RP-221348, “Revised SID: Study on Artificial Intelligence (AI) / Machine Learning (ML) for NR Air Interface,”Qualcomm (Moderator), June 2022.C. K. Wen, W. T. Shih, and S. Jin, “Deep learning for massive MIMO CSI feedback,” IEEE Wireless Communications Letters, Vol.7, No.5, October 2018.W. Liu, W. Tian, H. Xiao, S. Jin, X. Liu, J. Shen, “EVCsiNet: Eigenvector-based CSI feedback under 3GPP link-level channels,” IEEE Wireless Communications Letters, Vol.10, No.12, December 2021.H. Xiao, Z. Wang, D. Li, W. Tian, X. Liu, S. Jin, J. Shen, Z. Zhang, N. Yang, “AI enlightens wireless communication: A transformer backbone for CSI feedback,” China Communications, June 2022.RP-234039, “New WID on Artificial Intelligence (AI) / Machine Learning (ML) for NR Air Interface,” Qualcomm (Moderator), December 2023.R1-2400546, “Discussion on two-sided AI / ML model based CSI compression,” Xiaomi, February 26th - March 1st, 2024.3GPP TS38.211, “NR Physical channels and modulation (Release 18),” March 2024.3GPP TS38.212, “NR Multiplexing and channel coding (Release 18),” March 2024.3GPP TS38.213, “NR Physical layer procedures for control (Release 18),” March 2024.3GPP TS38.214, “NR Physical layer procedures for data (Release 18),” March 2024.

[0004] However, there is room for improvement in the efficiency of wireless communications.

[0005] Non-limiting embodiments of the present disclosure contribute to providing a terminal, a base station, and a communication method that can improve the efficiency of wireless communication.

[0006] A terminal according to one embodiment of the present disclosure includes a control circuit that controls input of an intermediate output from an artificial intelligence model on the terminal side at a past time to an artificial intelligence model on the terminal side that generates a channel state information report based on a first rank value at the current time and a second rank value at the previous time, and a transmitting circuit that transmits the channel state information report generated using the intermediate output.

[0007] These comprehensive or specific aspects may be realized as a system, an apparatus, a method, an integrated circuit, a computer program, or a recording medium, or may be realized as any combination of a system, an apparatus, a method, an integrated circuit, a computer program, and a recording medium.

[0008] According to an embodiment of the present disclosure, it is possible to improve the efficiency of wireless communication.

[0009] Further advantages and benefits of one embodiment of the present disclosure will become apparent from the specification and drawings. Such advantages and / or benefits may be provided by some embodiments and features described in the specification and drawings, respectively, but not necessarily all of them may be provided to obtain one or more identical features.

[0010] Channel State in the spatial and frequency domains using Machine Learning (ML) / Artificial Intelligence (AI) technology FIG. 1 shows an example of CSI compression in the time domain using AI / ML technology. FIG. 2 shows an example of an AI / MI model that performs CSI compression in the spatial domain, time domain, and frequency domain. FIG. 3 shows an example of an AI / MI model that performs CSI compression in the spatial domain, time domain, and frequency domain. FIG. 4 shows an example of setting an AI / MI model for each rank. FIG. 5 shows an example of CSI compression in the spatial domain, time domain, and frequency domain with an AI / MI model set for each rank. FIG. 6 shows an example of setting an AI / MI model for each layer. FIG. 7 shows an example of CSI compression in the spatial domain, time domain, and frequency domain with an AI / MI model set for each layer. FIG. 8 shows an example of updating AI / ML models for all ranks. Block diagram showing an example configuration of a portion of a base station. Block diagram showing an example configuration of a portion of a terminal. FIG. 9 shows an example of AI / ML processing. FIG. 10 shows an example of AI / ML processing. FIG. 11 shows an example of AI / ML processing. FIG. 12 shows an example of AI / ML processing. FIG. 13 shows an example of AI / ML processing. FIG. 14 shows an example of AI / ML processing. FIG. 15 shows an example of AI / ML processing. FIG. 16 shows an example of data set transfer. Flowchart showing an example of the operation of a terminal. FIG. 17 shows an example of an architecture. Block diagram showing an example configuration of a base station. Block diagram showing an example configuration of a terminal. FIG. 18 shows an example of an exemplary architecture of a 3GPP NR system. Diagram of an exemplary functional division in O-RAN

[0011] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings.

[0012] The 3rd Generation Partnership Project (3GPP), an international standardization organization, is working on the specification of New Radio (NR) as one of the 5G radio interfaces. The basic functions of eMBB and URLLC were specified in Release 15, and from Release 16 onwards, URLLC will be extended to include Industrial IoT, Vehicle-to-Everything (V2X), and non-terrestrial networks (NTN) including satellites. The 3GPP extended specifications will also be called "5G-Advanced" from Release 18.

[0013] Furthermore, advances in artificial intelligence (AI) (hereinafter also referred to as "AI / ML") technologies such as machine learning (ML) have been remarkable, and the application of AI / ML to mobile communications is also being considered. 3GPP is also studying and standardizing the application of AI / ML to various applications, and has begun studying the application of AI / ML to wireless interfaces since Release 18 (see, for example, Non-Patent Document 1). Representative use cases for utilizing AI / ML technologies include, for example, channel state information (CSI) feedback, beam control, and position estimation.

[0014] In NR, for example, an access scheme based on Orthogonal Frequency Division Multiplexing (OFDM) is adopted for downlink transmission. Furthermore, Multiple-Input Multiple Output (MIMO) is adopted to improve communication quality or increase data rates. To effectively utilize MIMO performance, closed-loop control is introduced, in which a terminal (e.g., user equipment (UE)) estimates the channel (e.g., measures and generates CSI) and feeds back the CSI from the terminal to a base station (e.g., gNB). The base station performs downlink communication by applying, for example, transmit precoding based on the fed-back CSI. CSI feedback is expected to achieve high performance with a smaller amount of control information.

[0015] In 3GPP Release 18, the application of AI / ML technology to the radio interface has been discussed. For example, a method for compressing CSI information in the spatial and frequency domains (hereinafter also referred to as the space-frequency domain) using AI / ML technology for CSI feedback (hereinafter also referred to as "CSI compression") is being considered. For example, machine learning (ML) can automatically extract features by training an AI / ML model (e.g., an artificial intelligence model) such as a neural network using a huge amount of training data. CSI compression can utilize an autoencoder, an algorithm primarily used for dimensional compression and reconstruction of image data. For example, as shown in Figure 1, a CSI matrix represented by a two-dimensional space-frequency domain is considered as an image. An encoder used for compression in the autoencoder is used for CSI compression processing on the terminal side, and a decoder used for reconstruction is used for CSI reconstruction processing on the base station or network (e.g., also referred to as "base station / network") side (see, for example, Non-Patent Document 2 or Non-Patent Document 3).

[0016] In addition, for CSI compression using ML, an algorithm called Transformer, which compresses the information of an entire sentence in a way that takes context into account in natural language processing, can also be used (see, for example, Non-Patent Document 4).

[0017] Furthermore, in 3GPP Release 19, in order to improve the performance of CSI compression in the space-frequency domain or reduce the amount of processing, CSI compression that uses information in the time domain (e.g., temporal domain) in addition to the space-frequency domain is also being considered (see, for example, Non-Patent Document 5). One example of using time-domain information for CSI compression is a method in which past CSI information is used in both an encoder that compresses CSI in a terminal and a decoder that reconstructs CSI on the base station / network side.

[0018] For example, in CSI compression using past CSI information in both the encoder of the terminal and the decoder of the base station / network, as shown in FIG. 2, the input to the AI / ML model (encoder or decoder) used for CSI compression at a certain time (e.g., time t) is the current CSI matrix (or precoding matrix) (Input(V t ) and the output from the AI / ML model in the past (e.g., time t-1) (past accumulated CSI). This allows for more efficient CSI compression, as the CSI feedback is differentially informed with respect to previous CSI reports, for example.

[0019] Rank adaptation is widely used in MIMO systems for mobile wireless communications, such as NR. Rank adaptation adaptively switches the number of multiplexed signals (rank) in MIMO transmission according to the state of the wireless channel (e.g., reception quality (e.g., SINR: Signal to Interference and Noise Ratio) or correlation of fading fluctuations between antennas), enabling robust and highly spectrally efficient transmission against time-varying wireless channels.

[0020] In the method using past CSI information, for example, an algorithm called Convolutional Long Short Term Memory (LTSM) (ConvLSTM: Convolutional LTSM), which is a type of recurrent neural network (RNN) that is also used for video frame prediction, may be used to construct an AI / ML model by combining LSTM with a machine learning algorithm for CSI compression (e.g., Transformer) as shown in FIG. 3 (see, for example, Non-Patent Document 6). The input of the AI / ML model for CSI compression in the spatial, time, and frequency domains (hereinafter also referred to as the space-time-frequency domains) includes both CSI information (or precoding information) at the current time and accumulated information from the past (past / accumulated CSI). The output of the encoder is compressed CSI (or precoding) information (CSI feedback C). t ), and the intermediate output of the LSTM part of the encoder is accumulated information (past / accumulated CSI) that is used as an input (additional input) to the AI / ML model used for CSI compression at the next time point. Similar to the encoder, the decoder restores the CSI (or precoding) information based on both the compressed CSI (or precoding) information and past accumulated information (past / accumulated CSI) calculated on the decoder side.

[0021] When rank adaptation is applied, the rank (or the number of layers) may change over time because it depends on the state of the wireless propagation path (channel). For example, when an AI / ML model is configured to apply inference processing for each rank or layer, if the rank or the number of layers changes over time, the accumulated information from the past (past / accumulated CSI) may not be continuously updated, as shown in Figure 4. In this case, performance degradation may occur in space-time-frequency domain CSI compression using past CSI information.

[0022] FIG. 5 shows an example of setting an AI / ML model for each rank in CSI compression in the space-frequency domain. FIG. 6 shows an example of extending the AI / ML model setting shown in FIG. 5 to CSI compression in the space-time-frequency domain, where the rank value changes over time.

[0023] As shown in Figures 5 and 6, when an AI / ML model is set for each rank, for example, an AI / ML model is trained for each rank, and in the inference process, rank-specific inference processing (e.g., encoder processing or decoder processing) is performed using the AI / ML model corresponding to the rank determined in the rank selection.

[0024] Fig. 6 shows an example in which the rank value calculated from the state of CSI information at time t-1 and the state of CSI information at time t is 2, and the rank value calculated from the state of CSI information at time t+1 is 1. For example, as shown in Fig. 6, at time t, an AI / ML model corresponding to rank 2 is used, and the input to the AI / ML model is the CSI information H t , and the accumulated information (past / accumulated CSI) S from time t-1 for the AI / ML model corresponding to rank 2 t-1 (2) Also, at time t, the output of the encoder is the compressed CSI (or precoding) information c t The intermediate output of the encoder is the accumulated information S t (2) is.

[0025] Here, as shown in Fig. 6, the rank value calculated from the state of CSI information at time t+1 is 1. In this case, the input to the AI / ML model corresponding to rank 1 is the CSI information H t+1 , and the accumulated information S from time t for the AI / ML model corresponding to rank 1 t (1) However, as shown in Figure 6, at the previous time t, the accumulated information S for the AI / ML model corresponding to rank 2 t (2)is output, but the accumulated information S for the AI / ML model corresponding to rank 1 t (1) is not output, and the AI / ML model corresponding to rank 1 used at time t+1 cannot use the accumulated information from time t.

[0026] Figure 7 shows an example of setting an AI / ML model for each layer in CSI compression in the space-frequency domain, and Figure 8 shows an example of extending the AI / ML model shown in Figure 7 to CSI compression in the space-time-frequency domain, showing a case where the rank value (or the number of layers) changes over time.

[0027] 7 and 8, when an AI / ML model is set for each layer, for example, an AI / ML model is trained for each layer, and in the inference process, an individual inference process (e.g., encoder process or decoder process) is performed for the layer corresponding to the rank determined in the rank selection. Alternatively, for example, a unified AI / ML model common to multiple layers may be trained, and the layer-specific inference process may be performed using the unified AI / ML model.

[0028] 8 shows an example in which the rank value calculated from the state of the CSI information at time t−1 and the state of the CSI information at time t+1 is 2, and the rank value calculated from the state of the CSI information at time t is 1. For example, as shown in FIG. 8, at time t, an AI / ML model corresponding to layer 1 is used, and the input to the AI / ML model is the CSI information H t , and accumulated information S from time t-1 for the AI / ML model corresponding to layer 1 t-1 (1) Also, at time t, the output of the encoder is the compressed CSI (or precoding) information c t The intermediate output of the encoder is the accumulated information S for the AI / ML model corresponding to Layer 1. t (1) is.

[0029] Here, as shown in Fig. 8, the rank value calculated from the state of the CSI information at time t+1 is 2. In this case, the input to the AI / ML model corresponding to layer 1 is the CSI information H t+1 , and accumulated information S from time t for the AI / ML model corresponding to Layer 1 t (1) On the other hand, the input to the AI / ML model corresponding to Layer 2 includes the CSI information H t+1 , and accumulated information S from time t for the AI / ML model corresponding to layer 2 t (2) However, as shown in Figure 8, at the previous time t, the accumulated information S for the AI / ML model corresponding to layer 2 is t (2) Since the output is not available, the AI / ML model corresponding to Layer 2 cannot use the accumulated information from time t.

[0030] In this way, the performance of CSI compression may be degraded due to the influence of loss of accumulated information (past / accumulated CSI) caused by a change in rank or the number of layers over time in rank adaptation.

[0031] One method for eliminating the impact of missing stored information due to temporal changes in rank or number of layers caused by the rank adaptation described above is to separate chains of stored information in the time domain between ranks or layers, as shown in FIG. 9 , so that the terminal and base station / network update the AI / ML models for all ranks or layers regardless of the rank selection result at each time. FIG. 9 illustrates an example of setting an AI / ML model for each rank. In this method, for example, only the chain of stored information corresponding to the rank determined by rank selection or the AI / ML model of the layer corresponding to the determined rank is actually used in the inference process for CSI compression or reconstruction, but the chain of stored information corresponding to other AI / ML models (AI / ML models not corresponding to the selected rank or layer) is also used to perform the inference process for updating the stored information. This allows past stored information for different ranks or layers to be used and updated even when the rank changes over time.

[0032] However, with this method, inference processing is performed using all AI / ML models, including not only the AI / ML models actually used in the inference processing for CSI compression or reconstruction, but also other AI / ML models that are not actually used. Since the AI / ML models are kept in continuous operation, there is a possibility that the power consumption or memory usage of the terminal and base station / network will increase.

[0033] In one non-limiting embodiment of the present disclosure, a method for eliminating or mitigating the effect of changes in rank or number of layers in rank adaptation in space-time-frequency domain CSI compression using past CSI information is described.

[0034] For example, in an AI / ML model configuration that separates the chain of accumulated information in the time domain between ranks or layers, the terminal and base station / network perform inference processing of an AI / ML model of a rank determined by rank selection or a layer corresponding to the determined rank (e.g., an AI / ML model actually used in inference processing for CSI compression or reconstruction) and update accumulated information (past / accumulated CSI) for the AI / ML model. Furthermore, the terminal and base station / network set (e.g., differentiate) accumulated information (past / accumulated CSI) used as additional input to the AI / ML model in the inference processing based on the rank value at the current time (e.g., time instance) and the rank value at a past time (time instance).

[0035] This makes it possible to prevent inconsistencies from occurring in the chain of accumulated information (past / accumulated CSI) even when the rank or the number of layers changes over time.

[0036] Non-limiting embodiments of the present disclosure will be described below.

[0037] [Overview of Communication System] A communication system according to an aspect of the present disclosure includes, for example, at least one base station and at least one terminal.

[0038] FIG. 10 is a block diagram showing a configuration example of a portion of a base station 100 according to an embodiment of the present disclosure, and FIG. 11 is a block diagram showing a configuration example of a portion of a terminal 200 according to an embodiment of the present disclosure.

[0039] In the base station 100 shown in Fig. 10, a communication unit (e.g., corresponding to a receiving circuit) receives a channel state information report (CSI report) generated by an AI / ML model on the terminal side. A control unit (e.g., corresponding to a control circuit) controls input of intermediate output (e.g., accumulated information) from the AI / ML model on the network side at a past time to an AI / ML model on the network (or base station / network) side that reconstructs the channel state information report based on a first rank value at the current time and a second rank value at the immediately preceding time.

[0040] 11 , a control unit (e.g., corresponding to a control circuit) controls input of an intermediate output from an AI / ML model on the terminal side at a past time to an AI / ML model on the terminal side that generates a channel state information report (CSI report) based on the first rank value at the current time and the second rank value at the immediately preceding time. A communication unit (e.g., corresponding to a transmission circuit) transmits the channel state information report generated using the intermediate output.

[0041] (Embodiment 1) Terminal 200 inputs CSI information measured and generated by terminal 200 to a CSI compression unit (encoder), generates a CSI report (e.g., compressed CSI), and reports the CSI report to base station 100 (e.g., base station / network). The base station / network uses a CSI reconstruction unit (decoder) to reconstruct (restore) the CSI report (compressed CSI) reported from terminal 200. These processes can also be considered as inference processes using AI / ML models (two-sided models) on both the terminal side and the base station / network side.

[0042] Here, a CSI report configuration framework based on at least a subband configuration and a rank notification (e.g., a layer configuration) may be applied to a CSI report (e.g., a compressed CSI report) based on an inference process using an AI / ML model. Furthermore, the CSI report (e.g., a compressed CSI) may be transmitted via uplink control information (UCI) on a physical uplink control channel (PUCCH) or a physical uplink shared channel (PUSCH). For example, a periodic CSI report, an aperiodic CSI report, or a semi-persistent CSI report may be used as the CSI report.

[0043] Terminal 200 may be notified of, for example, information regarding resources for CSI reports of at least one of Periodic CSI, Semi-persistent CSI, and Aperiodic CSI (e.g., information regarding resources such as PUCCH resources and PUSCH resources; also referred to as "CSI report configuration") using Radio Resource Control (RRC) signaling, Medium Access Control (MAC) signaling, Downlink Control Information (DCI), or a combination thereof.

[0044] The CSI report configuration may include, for example, information regarding a reporting period, an offset, etc. Furthermore, the CSI report configuration may include a configuration ID (e.g., CSI-ReportConfigId). The configuration ID may identify parameters such as the type of CSI reporting method and the reporting period. Furthermore, the CSI report configuration may include information indicating which reference signal (e.g., CSI-RS) is used to report the measured CSI.

[0045] The input of the AI / ML model of the CSI compression unit (encoder) includes both CSI information (or precoding information) at the current time and accumulated information (past / accumulated CSI) from past time points, as shown in Figure 2 or 3. The output of the encoder is compressed CSI (or precoding) information, and the intermediate output of the encoder is accumulated information (past / accumulated CSI) to be used as input to the encoder (AI / ML model) at the next time point.

[0046] The CSI reconstruction unit (decoder), like the encoder, may restore the CSI (or precoding) information based on both the compressed CSI (or precoding) information and the accumulated information from the past (past / accumulated CSI) calculated on the decoder side, for example, as shown in FIG. 2 or FIG. 3 .

[0047] In this embodiment, for example, as shown in FIG. 12 , an AI / ML model is set for each rank, and the chain of accumulated information in the time domain (e.g., update of accumulated information (intermediate output of the AI / ML model)) is made independent between ranks. Furthermore, when an AI / ML model is set for each rank, for example, the AI / ML model is trained for each rank, and the terminal 200 and the base station / network may perform rank-specific inference processing (e.g., encoder processing or decoder processing) in the inference processing using the AI / ML model corresponding to the rank determined by rank selection.

[0048] In this embodiment, in the inference process, if the rank value at a certain time (time instance) is unchanged (the same) from the previous time, the terminal 200 and the base station / network input the accumulated information updated and output in the AI / ML model at the previous time to the AI / ML model of the CSI compression unit (encoder) or CSI reconstruction unit (decoder) corresponding to the rank at the certain time as accumulated information (past / accumulated CSI) that is additional input to the encoder or decoder.

[0049] On the other hand, in the inference process, if the rank value at a certain time (time instance) is changed (different) from the previous time, the terminal 200 and the base station / network reset (discard) the accumulated information (for example, do not use the accumulated information) or set the accumulated information to a neutral value (for example, 0) and input it to the AI / ML model of the CSI compression unit (encoder) or the CSI reconstruction unit (decoder). The neutral value may be, for example, a value common between the base station / network and the terminal 200.

[0050] 12, rank 2 is selected at time t-1 and time t, and the AI / ML model corresponding to rank 2 is used. Therefore, since the rank value at time t is not changed from the rank value at time t-1, terminal 200 uses CSI information H t , and the accumulated information S from time t-1 for the AI / ML model corresponding to rank 2 t-1 (2) Enter.

[0051] 12, for example, rank 1 is selected at time t+1, and the AI / ML model corresponding to rank 1 is used. Therefore, since the rank value at time t+1 is changed from the rank value at time t, terminal 200 uses the CSI information H t+1 to reset the accumulated information (or the chain of accumulated information) for the past time, or to input a neutral value.

[0052] In addition, similar to the operation of terminal 200 shown in FIG. 12, the base station / network controls the input of accumulated information to the AI / ML model (use of accumulated information, reset of accumulated information, or use of a neutral value) according to the rank value at each time (time instance).

[0053] According to this embodiment, when the rank value changes over time, the chain of accumulated information is reset or updated without using (resetting) accumulated information (past / accumulated CSI) or by using pre-set accumulated information (neutral value), thereby preventing inconsistencies from occurring in the chain of accumulated information even when the rank or number of layers changes over time.

[0054] Furthermore, in the inference process at a certain time instance, only one AI / ML model corresponding to a specific rank (e.g., a rank determined by rank selection) is executed, thereby preventing an increase in power consumption or memory usage of the terminal 200 and base station / network.

[0055] (Embodiment 2) Terminal 200 inputs CSI information measured and generated by terminal 200 to a CSI compression unit (encoder), generates a CSI report (e.g., compressed CSI), and reports the CSI report to base station 100 (e.g., base station / network). The base station / network uses a CSI reconstruction unit (decoder) to reconstruct (restore) the CSI report (compressed CSI) reported from terminal 200. These processes can also be considered as inference processes using AI / ML models (two-sided models) on both the terminal side and the base station / network side.

[0056] Here, a CSI report configuration framework based on at least a subband configuration and a rank indication (e.g., a layer configuration) may be applied to a CSI report (e.g., a compressed CSI report) based on inference processing using an AI / ML model. Furthermore, the CSI report (e.g., a compressed CSI) may be transmitted in UCI on a PUCCH or a PUSCH. For example, a periodic CSI report, an aperiodic CSI report, or a semi-persistent CSI report may be used as the CSI report.

[0057] Terminal 200 may be notified of, for example, information regarding resources for CSI reports of at least one of Periodic CSI, Semi-persistent CSI, and Aperiodic CSI (e.g., information regarding resources such as PUCCH resources and PUSCH resources (e.g., CSI report configuration)) using RRC signaling, MAC signaling, downlink control information (DCI), or a combination thereof.

[0058] The CSI report configuration may include, for example, information regarding a reporting period, an offset, etc. Furthermore, the CSI report configuration may include a configuration ID (e.g., CSI-ReportConfigId). The configuration ID may identify parameters such as the type of CSI reporting method and the reporting period. Furthermore, the CSI report configuration may include information indicating which reference signal (e.g., CSI-RS) is used to report the measured CSI.

[0059] The input of the AI / ML model of the CSI compression unit (encoder) includes both CSI information (or precoding information) at the current time and accumulated information (past / accumulated CSI) from past time points, as shown in Figure 2 or 3. The output of the encoder is compressed CSI (or precoding) information, and the intermediate output of the encoder is accumulated information (past / accumulated CSI) to be used as input to the encoder (AI / ML model) at the next time point.

[0060] The CSI reconstruction unit (decoder), like the encoder, may restore the CSI (or precoding) information based on both the compressed CSI (or precoding) information and the accumulated information from the past (past / accumulated CSI) calculated on the decoder side, for example, as shown in FIG. 2 or FIG. 3 .

[0061] In this embodiment, as shown in FIG. 13, for example, similar to the first embodiment, an AI / ML model is set for each rank, and the chain of accumulated information in the time domain (e.g., update of accumulated information (intermediate output of the AI / ML model)) is made independent between ranks. Furthermore, when an AI / ML model is set for each rank, for example, the AI / ML model is trained for each rank, and in the inference process, the terminal 200 and the base station / network may perform rank-specific inference processing (e.g., encoder processing or decoder processing) using the AI / ML model corresponding to the rank determined by rank selection.

[0062] In this embodiment, in the inference process, if the rank value at a certain time (time instance) is unchanged (the same) from the previous time, the terminal 200 and the base station / network input the accumulated information updated and output in the AI / ML model at the previous time to the AI / ML model of the CSI compression unit (encoder) or CSI reconstruction unit (decoder) corresponding to the rank at the certain time as accumulated information (past / accumulated CSI) that is additional input to the encoder or decoder.

[0063] On the other hand, when the rank value at a certain time is changed (different) from the previous time in the inference process, the terminal 200 and the base station / network input the intermediate output (e.g., accumulated information) of the inference process at a past time at which the same rank value as the current time was set to the AI / ML model of the CSI compression unit (encoder) or CSI reconstruction unit (decoder) corresponding to the rank at the certain time. For example, the terminal 200 and the base station / network use the accumulated information updated and output by the AI / ML model in the inference process at the latest available time (e.g., the time closest to the current time) among past times at which the same rank value as the current time was set as accumulated information that is an additional input to the encoder or decoder.

[0064] For example, as shown in Fig. 13, at time t-1, rank 2 is selected and the AI / ML model corresponding to rank 2 is used, and at time t, rank 1 is selected and the AI / ML model corresponding to rank 1 is used. Therefore, since the rank value at time t is changed from the rank value at time t-1, terminal 200 uses CSI information H t , and accumulated information S from past time tx for the AI / ML model corresponding to rank 1 t-x (1) Enter.

[0065] 13, for example, rank 2 is selected at time t+1, and the AI / ML model corresponding to rank 2 is used. Therefore, since the rank value at time t+1 is changed from the rank value at time t, terminal 200 uses the CSI information H t+1 , and accumulated information S from the past time t-1 for the AI / ML model corresponding to rank 2 t-1 (2) Enter.

[0066] In addition, similar to the operation of terminal 200 shown in FIG. 13, the base station / network controls the input of accumulated information to the AI / ML model (use of accumulated information from the previous time, or two or more previous times) depending on the rank value at each time (time instance).

[0067] The accumulated information at past times used in this embodiment is not limited to the accumulated information updated and output by the AI / ML model in the inference process at the time closest to the current time. For example, in consideration of the processing delay of the terminal 200, the accumulated information at the latest time available at a time prior to the current time plus a certain processing delay time may be used.

[0068] In the following explanation, the most recent available time (e.g., the time instance closest to the current time) among the past time instances that has the same rank value as the current time instance will be referred to as "time instance A" (or "time A").

[0069] According to this embodiment, when a rank value changes over time, the chain of accumulated information can be updated using accumulated information (past / accumulated CSI) updated and output from an AI / ML model corresponding to a time (e.g., time instance A) with the same available rank value in the past. This prevents inconsistencies from occurring in the chain of accumulated information even when the rank or number of layers changes over time. Furthermore, this embodiment enables CSI compression using time information (past accumulated information), which may potentially achieve superior performance in CSI compression compared to when the accumulated information is set to a neutral value (e.g., 0).

[0070] Furthermore, in the inference process at a certain time instance, only one AI / ML model corresponding to a specific rank (e.g., a rank determined by rank selection) is executed, thereby preventing an increase in power consumption or memory usage of the terminal 200 and base station / network.

[0071] (Embodiment 3) Terminal 200 inputs CSI information measured and generated by terminal 200 to a CSI compression unit (encoder), generates a CSI report (e.g., compressed CSI), and reports the CSI report to base station 100 (e.g., base station / network). The base station / network uses a CSI reconstruction unit (decoder) to reconstruct (restore) the CSI report (compressed CSI) reported from terminal 200. These processes can also be considered as inference processes using AI / ML models (two-sided models) on both the terminal side and the base station / network side.

[0072] Here, a CSI report configuration framework based on at least a subband configuration and a rank indication (e.g., a layer configuration) may be applied to a CSI report (e.g., a compressed CSI report) based on inference processing using an AI / ML model. Furthermore, the CSI report (e.g., a compressed CSI) may be transmitted in UCI on a PUCCH or a PUSCH. For example, a periodic CSI report, an aperiodic CSI report, or a semi-persistent CSI report may be used as the CSI report.

[0073] Terminal 200 may be notified of, for example, information regarding resources for CSI reports of at least one of Periodic CSI, Semi-persistent CSI, and Aperiodic CSI (e.g., information regarding resources such as PUCCH resources and PUSCH resources (e.g., CSI report configuration)) using RRC signaling, MAC signaling, downlink control information (DCI), or a combination thereof.

[0074] The CSI report configuration may include, for example, information regarding a reporting period, an offset, etc. Furthermore, the CSI report configuration may include a configuration ID (e.g., CSI-ReportConfigId). The configuration ID may identify parameters such as the type of CSI reporting method and the reporting period. Furthermore, the CSI report configuration may include information indicating which reference signal (e.g., CSI-RS) is used to report the measured CSI.

[0075] The input of the AI / ML model of the CSI compression unit (encoder) includes both CSI information (or precoding information) at the current time and accumulated information (past / accumulated CSI) from past time points, as shown in Figure 2 or 3. The output of the encoder is compressed CSI (or precoding) information, and the intermediate output of the encoder is accumulated information (past / accumulated CSI) to be used as input to the encoder (AI / ML model) at the next time point.

[0076] The CSI reconstruction unit (decoder), like the encoder, may restore the CSI (or precoding) information based on both the compressed CSI (or precoding) information and the accumulated information from the past (past / accumulated CSI) calculated on the decoder side, for example, as shown in FIG. 2 or FIG. 3 .

[0077] In this embodiment, as shown in, for example, FIGS. 14 and 15 , similarly to the first embodiment, an AI / ML model is set for each rank, and the chain of accumulated information in the time domain (e.g., update of accumulated information (intermediate output of the AI / ML model)) is made independent between ranks. Furthermore, when an AI / ML model is set for each rank, for example, the AI / ML model is trained for each rank, and in the inference process, the terminal 200 and the base station / network may perform rank-specific inference processing (e.g., encoder processing or decoder processing) using the AI / ML model corresponding to the rank determined by rank selection.

[0078] In this embodiment, in the inference process, if the rank value at a certain time (time instance) is unchanged (the same) from the previous time, the terminal 200 and the base station / network input the accumulated information updated and output in the AI / ML model at the previous time to the AI / ML model of the CSI compression unit (encoder) or CSI reconstruction unit (decoder) corresponding to the rank at the certain time as accumulated information (past / accumulated CSI) that is additional input to the encoder or decoder.

[0079] On the other hand, in the inference process, when the rank value at a certain time is changed (different) from the previous time, the terminal 200 and the base station / network set the accumulated information to be input to the AI / ML model depending on whether or not the accumulated information (e.g., accumulated information at a past time) updated and output by the AI / ML model in the inference process of time instance A in embodiment 2 is valid. For example, the terminal 200 and the base station / network may select either the process of embodiment 1 or embodiment 2 depending on whether or not the accumulated information updated and output by the AI / ML model in the inference process of time instance A is valid.

[0080] For example, if the accumulated information updated and output by the AI / ML model in the inference process of time instance A is invalid, the terminal 200 and the base station / network perform the process of embodiment 1 (e.g., resetting the accumulated information (not using the accumulated information) or setting a neutral value (e.g., 0)).

[0081] On the other hand, if the accumulated information updated and output by the AI / ML model in the inference processing of time instance A is valid, terminal 200 and base station / network perform the processing of embodiment 2 (e.g., processing to input the accumulated information updated and output by the AI / ML model in the inference processing of time instance A to the AI / ML model).

[0082] Here, a time window or timer may be applied to determine whether the accumulated information updated and output by the AI / ML model in the inference process of time instance A is valid.

[0083] For example, when a time window is applied, as shown in Fig. 14, if the time interval between time instance A (e.g., time instance #tM) and the current time instance (e.g., time instance #t) is equal to or less than the time interval set as the time window, terminal 200 and the base station / network determine that the accumulated information for time instance A is valid and select the processing of embodiment 2. On the other hand, as shown in Fig. 15, if the time interval between time instance A (e.g., time instance #tM) and the current time instance (e.g., time instance #t) is greater than the time interval set as the time window, terminal 200 and the base station / network determine that the accumulated information for time instance A is invalid and select the processing of embodiment 1.

[0084] Furthermore, for example, when a timer is applied, the terminal 200 may clear (discard) the stored information when a certain time has elapsed since the terminal 200 performed the inference process at a certain time instance (e.g., time instance A). For example, as shown in FIG. 14 , if the terminal 200 at the current time instance (e.g., time instance#t) is storing stored information updated and output by the AI / ML model in the inference process at time instance A (time instance#tM) (e.g., if the time is within a time range set by the timer), the terminal 200 determines that the stored information at time instance A is valid and selects the process of embodiment 2. On the other hand, for example, as shown in FIG. 15 , if the terminal 200 at the current time instance (e.g., time instance#t) is not storing stored information updated and output by the AI / ML model in the inference process at time instance A (time instance#tM) (e.g., if the time is outside the time range set by the timer), the terminal 200 determines that the stored information at time instance A is invalid and selects the process of embodiment 1.

[0085] In addition, the base station / network controls the input of accumulated information to the AI / ML model according to a time window or timer, similar to the operation of the terminal 200 shown in Figures 14 and 15.

[0086] According to this embodiment, while avoiding the influence of past CSI information for which the accumulated information is no longer valid, when the rank value changes over time, if the accumulated information (past / accumulated CSI) updated and output from the AI / ML model corresponding to a time (e.g., time instance A) with the same usable rank value in the past is valid, the accumulated information chain can be updated using the accumulated information. Therefore, even when the rank or the number of layers changes over time, inconsistencies in the accumulated information chain can be prevented.

[0087] Furthermore, in this embodiment, in the inference process at a certain time instance, only one AI / ML model corresponding to a specific rank (e.g., a rank determined by rank selection) is executed, thereby preventing an increase in power consumption or memory usage of the terminal 200 and the base station / network.

[0088] The time window or timer setting value may be set to the same value for all ranks, or may be set to a different value depending on the rank.

[0089] (Variation of Embodiment 3) In Embodiment 3, a case has been described in which a time window or a timer is applied to determine whether the accumulated information updated and output by the AI / ML model in the inference process of time instance A is valid, but other methods may be used to determine whether past accumulated information is valid.

[0090] For example, terminal 200 and base station / network may compare the reception quality (e.g., SINR) at time instance A with the SINR at the current time instance, and determine whether or not previously accumulated information is valid based on the difference in SINR. For example, terminal 200 and base station / network may determine that previously accumulated information is valid if the difference in SINR is equal to or smaller than a threshold, and may determine that previously accumulated information is invalid if the difference in SINR is greater than the threshold.

[0091] Furthermore, for example, terminal 200 and base station / network may determine whether previously accumulated information is valid depending on the range of fluctuation in rank between time instance A and the current time instance. For example, terminal 200 and base station / network may determine that previously accumulated information is valid if the range of fluctuation in rank is equal to or less than a threshold, and may determine that previously accumulated information is invalid if the range of fluctuation in rank is greater than the threshold.

[0092] (Fourth Embodiment) Terminal 200 inputs CSI information measured and generated by terminal 200 to a CSI compression unit (encoder), generates a CSI report (e.g., compressed CSI), and reports the CSI report to base station 100 (e.g., base station / network). The base station / network uses a CSI reconstruction unit (decoder) to reconstruct (restore) the CSI report (compressed CSI) reported from terminal 200. These processes can also be considered as inference processes using AI / ML models (two-sided models) on both the terminal side and the base station / network side.

[0093] Here, a CSI report configuration framework based on at least a subband configuration and a rank indication (e.g., a layer configuration) may be applied to a CSI report (e.g., a compressed CSI report) based on inference processing using an AI / ML model. Furthermore, the CSI report (e.g., a compressed CSI) may be transmitted in UCI on a PUCCH or a PUSCH. For example, a periodic CSI report, an aperiodic CSI report, or a semi-persistent CSI report may be used as the CSI report.

[0094] Terminal 200 may be notified of, for example, information regarding resources for CSI reports of at least one of Periodic CSI, Semi-persistent CSI, and Aperiodic CSI (e.g., information regarding resources such as PUCCH resources and PUSCH resources (e.g., CSI report configuration)) using RRC signaling, MAC signaling, downlink control information (DCI), or a combination thereof.

[0095] The CSI report configuration may include, for example, information regarding a reporting period, an offset, etc. Furthermore, the CSI report configuration may include a configuration ID (e.g., CSI-ReportConfigId). The configuration ID may identify parameters such as the type of CSI reporting method and the reporting period. Furthermore, the CSI report configuration may include information indicating which reference signal (e.g., CSI-RS) is used to report the measured CSI.

[0096] The input of the AI / ML model of the CSI compression unit (encoder) includes both CSI information (or precoding information) at the current time and accumulated information (past / accumulated CSI) from past time points, as shown in Figure 2 or 3. The output of the encoder is compressed CSI (or precoding) information, and the intermediate output of the encoder is accumulated information (past / accumulated CSI) to be used as input to the encoder (AI / ML model) at the next time point.

[0097] The CSI reconstruction unit (decoder), like the encoder, may restore the CSI (or precoding) information based on both the compressed CSI (or precoding) information and the accumulated information from the past (past / accumulated CSI) calculated on the decoder side, for example, as shown in FIG. 2 or FIG. 3 .

[0098] In this embodiment, for example, as shown in FIG. 16 , an AI / ML model is set for each layer, and the chain of accumulated information in the time domain (e.g., update of accumulated information (intermediate output of the AI / ML model)) is made independent between layers. Furthermore, when an AI / ML model is set for each layer, for example, the AI / ML model is trained individually for each layer, and the terminal 200 and the base station / network may perform layer-specific inference processing (e.g., encoder processing or decoder processing) using an AI / ML model corresponding to each layer in inference processing (e.g., layer-specific and rank common). Furthermore, when an AI / ML model is set for each layer, for example, a unified AI / ML model common to the layers is trained, and the terminal 200 and the base station / network may perform layer-specific inference processing using the unified AI / ML model in inference processing (e.g., layer common and rank common). In either case, it is assumed that a unified AI / ML model is trained for all ranks in a specific layer. For example, an AI / ML model is trained individually or commonly for each layer, and an AI / ML model corresponding to a certain layer is commonly used for different ranks. In addition, the input to the AI / ML model of the CSI compression unit (encoder) may be information obtained by pre-processing the measured channel information and converting it into layer-specific information (e.g., precoding information such as eigenvalue vectors).

[0099] In this embodiment, in the inference process, if the rank value at a certain time (time instance) is unchanged (the same) from the previous time, or if the rank value at a certain time is changed to a value smaller than the rank value at the previous time, the terminal 200 and the base station / network input the accumulated information updated and output in the AI / ML model at the previous time to the AI / ML model of the CSI compression unit (encoder) or CSI reconstruction unit (decoder) corresponding to the rank at the certain time as accumulated information (past / accumulated CSI) that is an additional input to the encoder or decoder.

[0100] On the other hand, when, in the inference process, the rank value (e.g., N) at a certain time instance is changed to a value greater than the rank value (e.g., M) at the immediately previous time instance (when N>M), the terminal 200 and the base station / network input the accumulated information updated and output in the AI / ML model corresponding to layers 1 to M at the immediately previous time instance as accumulated information that is additional input to the encoder or decoder, to the AI / ML model corresponding to layers 1 to M. Also, when N>M, the terminal 200 and the base station / network reset (e.g., do not use) the accumulated information for the AI / ML model corresponding to layers (M+1) to N, or set the accumulated information to a neutral value (e.g., 0), and input it to the AI / ML model of the CSI compression unit (encoder) or the CSI reconstruction unit (decoder).

[0101] 16 , at time t−1, rank 2 is selected and an AI / ML model corresponding to layers 1 and 2 is used, and at time t, rank 1 is selected and an AI / ML model corresponding to layer 1 is used. Therefore, since the rank value at time t is changed to a value smaller than the rank value at time t−1, terminal 200 uses CSI information H t , and the accumulated information S that is updated and output in the AI / ML model corresponding to layer 1 at the previous time t-1 t-1 (1) Enter.

[0102] 16, for example, rank 2 is selected at time t+1, and the AI / ML model corresponding to layers 1 and 2 is used. Therefore, since the rank value at time t+1 is changed to a value greater than the rank value at time t, terminal 200 uses CSI information H t+1 , and the accumulated information S t (1) Furthermore, the terminal 200 inputs CSI information Ht+1 to reset the accumulated information (or the chain of accumulated information) for the past time, or to input a neutral value.

[0103] In addition, similar to the operation of terminal 200 shown in FIG. 16, the base station / network controls the input of accumulated information to the AI / ML model (use of accumulated information, reset of accumulated information, or use of a neutral value) according to the rank value at each time (time instance).

[0104] According to this embodiment, even if the rank value changes over time, the AI / ML model corresponding to the lower layer (e.g., layer 1) can continue updating without resetting the chain of accumulated information. Also, in the higher layer (e.g., layer 2), even if the rank value changes over time, the chain of accumulated information is reset or updated without using (resetting) the accumulated information (past / accumulated CSI) or by using preset accumulated information (neutral value), thereby preventing inconsistency in the chain of accumulated information even if the rank or the number of layers changes over time.

[0105] Furthermore, in the inference process at a certain time instance, only the AI / ML model with the number of layers corresponding to a specific rank (e.g., the rank determined by rank selection) is executed, thereby preventing an increase in power consumption or memory usage of the terminal 200 and the base station / network.

[0106] (Embodiment 5) Terminal 200 inputs CSI information measured and generated by terminal 200 to a CSI compression unit (encoder), generates a CSI report (e.g., compressed CSI), and reports the CSI report to base station 100 (e.g., base station / network). The base station / network uses a CSI reconstruction unit (decoder) to reconstruct (restore) the CSI report (compressed CSI) reported from terminal 200. These processes can also be considered as inference processes using AI / ML models (two-sided models) on both the terminal side and the base station / network side.

[0107] Here, a CSI report configuration framework based on at least a subband configuration and a rank indication (e.g., a layer configuration) may be applied to a CSI report (e.g., a compressed CSI report) based on inference processing using an AI / ML model. Furthermore, the CSI report (e.g., a compressed CSI) may be transmitted in UCI on a PUCCH or a PUSCH. For example, a periodic CSI report, an aperiodic CSI report, or a semi-persistent CSI report may be used as the CSI report.

[0108] Terminal 200 may be notified of, for example, information regarding resources for CSI reports of at least one of Periodic CSI, Semi-persistent CSI, and Aperiodic CSI (e.g., information regarding resources such as PUCCH resources and PUSCH resources (e.g., CSI report configuration)) using RRC signaling, MAC signaling, downlink control information (DCI), or a combination thereof.

[0109] The CSI report configuration may include, for example, information regarding a reporting period, an offset, etc. Furthermore, the CSI report configuration may include a configuration ID (e.g., CSI-ReportConfigId). The configuration ID may identify parameters such as the type of CSI reporting method and the reporting period. Furthermore, the CSI report configuration may include information indicating which reference signal (e.g., CSI-RS) is used to report the measured CSI.

[0110] The input of the AI / ML model of the CSI compression unit (encoder) includes both CSI information (or precoding information) at the current time and accumulated information (past / accumulated CSI) from past time points, as shown in Figure 2 or 3. The output of the encoder is compressed CSI (or precoding) information, and the intermediate output of the encoder is accumulated information (past / accumulated CSI) to be used as input to the encoder (AI / ML model) at the next time point.

[0111] The CSI reconstruction unit (decoder), like the encoder, may restore the CSI (or precoding) information based on both the compressed CSI (or precoding) information and the accumulated information from the past (past / accumulated CSI) calculated on the decoder side, for example, as shown in FIG. 2 or FIG. 3 .

[0112] In this embodiment, for example, as shown in FIG. 17 , an AI / ML model is set for each layer, and the chain of accumulated information in the time domain (e.g., update of accumulated information (intermediate output of the AI / ML model)) is made independent between layers. Furthermore, when an AI / ML model is set for each layer, for example, the AI / ML model is trained individually for each layer, and the terminal 200 and the base station / network may perform layer-specific inference processing (e.g., encoder processing or decoder processing) using an AI / ML model corresponding to each layer in inference processing (e.g., layer-specific and rank common). Furthermore, when an AI / ML model is set for each layer, for example, a unified AI / ML model common to the layers is trained, and the terminal 200 and the base station / network may perform layer-specific inference processing using the unified AI / ML model in inference processing (e.g., layer common and rank common). In either case, it is assumed that a unified AI / ML model is trained for all ranks in a specific layer. For example, an AI / ML model is trained individually or commonly for each layer, and an AI / ML model corresponding to a certain layer is commonly used for different ranks. In addition, the input to the AI / ML model of the CSI compression unit (encoder) may be information obtained by pre-processing the measured channel information and converting it into layer-specific information (e.g., precoding information such as eigenvalue vectors).

[0113] In this embodiment, in the inference process, if the rank value at a certain time (time instance) is unchanged (the same) from the previous time, or if the rank value at a certain time is changed to a value smaller than the rank value at the previous time, the terminal 200 and the base station / network input the accumulated information updated and output in the AI / ML model at the previous time to the AI / ML model of the CSI compression unit (encoder) or CSI reconstruction unit (decoder) corresponding to the rank at the certain time as accumulated information (past / accumulated CSI) that is an additional input to the encoder or decoder.

[0114] On the other hand, in the inference process, when the rank value (e.g., N) at a certain time (time instance) is changed to a value greater than the rank value (e.g., M) at the previous time (when N>M), the terminal 200 and the base station / network input the accumulated information that is updated and output in the AI / ML model corresponding to layers 1 to M at the previous time as accumulated information that is additional input to the encoder or decoder for the AI / ML model corresponding to layers 1 to M.

[0115] Furthermore, when N>M, the terminal 200 and the base station / network input intermediate outputs (e.g., accumulated information) in the inference process at a past time for which a rank value equal to or greater than the current time is set for the AI / ML models corresponding to layers (M+1) to N to the AI / ML model of the CSI compression unit (encoder) or CSI reconstruction unit (decoder) corresponding to the rank at a certain time. For example, the terminal 200 and the base station / network use accumulated information updated and output by the AI / ML model in the inference process at the latest available time (e.g., the time closest to the current time) among past times for which a rank value equal to or greater than the current time is set, as accumulated information that is an additional input to the encoder or decoder.

[0116] 17 , at time t−1, rank 2 is selected and an AI / ML model corresponding to layers 1 and 2 is used, and at time t, rank 1 is selected and an AI / ML model corresponding to layer 1 is used. Therefore, since the rank value at time t is changed to a value smaller than the rank value at time t−1, terminal 200 uses CSI information H t , and the accumulated information S that is updated and output in the AI / ML model corresponding to layer 1 at the previous time t-1 t-1 (1) Enter.

[0117] 17, for example, rank 2 is selected at time t+1, and the AI / ML model corresponding to layers 1 and 2 is used. Therefore, since the rank value at time t+1 is changed to a value greater than the rank value at time t, terminal 200 uses the CSI information H t+1 , and the accumulated information S t (1) Furthermore, the terminal 200 inputs the CSI information H t+1 is input, and the accumulated information S t-1 (2) Enter.

[0118] In addition, similar to the operation of terminal 200 shown in FIG. 17, the base station / network controls the input of accumulated information to the AI / ML model (use of accumulated information from the previous time, or two or more previous times) depending on the rank value at each time (time instance).

[0119] The accumulated information at past times used in this embodiment is not limited to the accumulated information updated and output by the AI / ML model in the inference process at the time closest to the current time. For example, in consideration of the processing delay of the terminal 200, the accumulated information at the latest time available at a time prior to the current time plus a certain processing delay time may be used.

[0120] In the following explanation, the most recent available time (e.g., the time instance closest to the current time) among the past time instances that has a rank value equal to or greater than the current time (current time instance) will be referred to as "time instance B" (or "time B").

[0121] According to this embodiment, even if the rank value changes over time, the AI / ML model corresponding to a lower layer (e.g., layer 1) can continue updating the chain of accumulated information without resetting it. Furthermore, even if the rank value changes over time, in a higher layer (e.g., layer 2), the chain of accumulated information can be updated using accumulated information (past / accumulated CSI) updated and output from an AI / ML model corresponding to a time (e.g., time instance B) at which the same available rank value is set in the past. Therefore, even if the rank or the number of layers changes over time, inconsistencies in the chain of accumulated information can be prevented. Furthermore, according to this embodiment, CSI compression using time information (past accumulated information) is possible, which may enable superior performance in CSI compression compared to setting the accumulated information to a neutral value (e.g., 0).

[0122] Furthermore, in the inference process at a certain time instance, only the AI / ML model with the number of layers corresponding to a specific rank (e.g., the rank determined by rank selection) is executed, thereby preventing an increase in power consumption or memory usage of the terminal 200 and the base station / network.

[0123] Sixth Embodiment Terminal 200 inputs CSI information measured and generated by terminal 200 to a CSI compression unit (encoder), generates a CSI report (e.g., compressed CSI), and reports the CSI report to base station 100 (e.g., base station / network). The base station / network uses a CSI reconstruction unit (decoder) to reconstruct (restore) the CSI report (compressed CSI) reported from terminal 200. These processes can also be considered as inference processes using AI / ML models (two-sided models) on both the terminal side and the base station / network side.

[0124] Here, a CSI report configuration framework based on at least a subband configuration and a rank indication (e.g., a layer configuration) may be applied to a CSI report (e.g., a compressed CSI report) based on inference processing using an AI / ML model. Furthermore, the CSI report (e.g., a compressed CSI) may be transmitted in UCI on a PUCCH or a PUSCH. For example, a periodic CSI report, an aperiodic CSI report, or a semi-persistent CSI report may be used as the CSI report.

[0125] Terminal 200 may be notified of, for example, information regarding resources for CSI reports of at least one of Periodic CSI, Semi-persistent CSI, and Aperiodic CSI (e.g., information regarding resources such as PUCCH resources and PUSCH resources (e.g., CSI report configuration)) using RRC signaling, MAC signaling, downlink control information (DCI), or a combination thereof.

[0126] The CSI report configuration may include, for example, information regarding a reporting period, an offset, etc. Furthermore, the CSI report configuration may include a configuration ID (e.g., CSI-ReportConfigId). The configuration ID may identify parameters such as the type of CSI reporting method and the reporting period. Furthermore, the CSI report configuration may include information indicating which reference signal (e.g., CSI-RS) is used to report the measured CSI.

[0127] The input of the AI / ML model of the CSI compression unit (encoder) includes both CSI information (or precoding information) at the current time and accumulated information (past / accumulated CSI) from past time points, as shown in Figure 2 or 3. The output of the encoder is compressed CSI (or precoding) information, and the intermediate output of the encoder is accumulated information (past / accumulated CSI) to be used as input to the encoder (AI / ML model) at the next time point.

[0128] The CSI reconstruction unit (decoder), like the encoder, may restore the CSI (or precoding) information based on both the compressed CSI (or precoding) information and the accumulated information from the past (past / accumulated CSI) calculated on the decoder side, for example, as shown in FIG. 2 or FIG. 3 .

[0129] In this embodiment, for example, an AI / ML model is set for each layer, and the chain of accumulated information in the time domain (e.g., update of accumulated information (intermediate output of the AI / ML model)) is made independent between layers. Furthermore, when an AI / ML model is set for each layer, for example, the AI / ML model is trained individually for each layer, and the terminal 200 and the base station / network may perform layer-specific inference processing (e.g., encoder processing or decoder processing) using an AI / ML model corresponding to each layer in inference processing (e.g., layer-specific and rank common). Furthermore, when an AI / ML model is set for each layer, for example, a unified AI / ML model common to the layers is trained, and the terminal 200 and the base station / network may perform layer-specific inference processing using the unified AI / ML model in inference processing (e.g., layer common and rank common). In either case, it is assumed that a unified AI / ML model is trained for all ranks in a specific layer. For example, an AI / ML model is trained individually or commonly for each layer, and an AI / ML model corresponding to a certain layer is commonly used for different ranks. In addition, the input to the AI / ML model of the CSI compression unit (encoder) may be information obtained by pre-processing the measured channel information and converting it into layer-specific information (e.g., precoding information such as eigenvalue vectors).

[0130] In this embodiment, in the inference process, if the rank value at a certain time (time instance) is unchanged (the same) from the previous time, or if the rank value at a certain time is changed to a value smaller than the rank value at the previous time, the terminal 200 and the base station / network input the accumulated information updated and output in the AI / ML model at the previous time to the AI / ML model of the CSI compression unit (encoder) or CSI reconstruction unit (decoder) corresponding to the rank at the certain time as accumulated information (past / accumulated CSI) that is an additional input to the encoder or decoder.

[0131] On the other hand, in the inference process, when the rank value (e.g., N) at a certain time (time instance) is changed to a value greater than the rank value (e.g., M) at the previous time (when N>M), the terminal 200 and the base station / network input the accumulated information that is updated and output in the AI / ML model corresponding to layers 1 to M at the previous time as accumulated information that is additional input to the encoder or decoder for the AI / ML model corresponding to layers 1 to M.

[0132] Furthermore, when N>M, terminal 200 and base station / network set accumulated information to be input to the AI / ML model corresponding to layers (M+1) to N, depending on whether or not accumulated information (e.g., accumulated information at past times) updated and output by the AI / ML model in the inference process of time instance B in embodiment 5 is valid. For example, terminal 200 and base station / network may select either the process of embodiment 4 or embodiment 5 depending on whether or not accumulated information updated and output by the AI / ML model in the inference process of time instance B is valid.

[0133] For example, if the accumulated information updated and output by the AI / ML model in the inference process of time instance B is invalid, the terminal 200 and the base station / network perform the process of embodiment 4 (e.g., resetting the accumulated information (not using the accumulated information) or setting a neutral value (e.g., 0)).

[0134] On the other hand, if the accumulated information updated and output by the AI / ML model in the inference processing of time instance B is valid, terminal 200 and base station / network perform the processing of embodiment 5 (e.g., processing to input the accumulated information updated and output by the AI / ML model in the inference processing of time instance B to the AI / ML model).

[0135] Here, a time window or timer may be applied to determine whether the accumulated information updated and output by the AI / ML model in the inference process of time instance B is valid.

[0136] For example, when a time window is applied, if the time interval between time instance B and the current time instance is equal to or less than the time interval set as the time window, terminal 200 and the base station / network determine that the accumulated information for time instance B is valid and select the processing of embodiment 5. On the other hand, if the time interval between time instance B and the current time instance is greater than the time interval set as the time window, terminal 200 and the base station / network determine that the accumulated information for time instance B is invalid and select the processing of embodiment 4.

[0137] Furthermore, for example, when a timer is applied, the terminal 200 may clear (discard) the stored information when a certain time has elapsed since the terminal 200 performed the inference process at a certain time instance (e.g., time instance B). For example, if the terminal 200 at the current time instance stores stored information that has been updated and output by the AI / ML model in the inference process at time instance B (e.g., within a time range set by the timer), the terminal 200 determines that the stored information at time instance B is valid and selects the processing of embodiment 5. On the other hand, for example, if the terminal 200 at the current time instance does not store stored information that has been updated and output by the AI / ML model in the inference process at time instance B (e.g., outside the time range set by the timer), the terminal 200 determines that the stored information at time instance B is invalid and selects the processing of embodiment 4.

[0138] In addition, the base station / network controls the input of accumulated information to the AI / ML model according to a time window or timer, similar to the operation of the terminal 200.

[0139] According to this embodiment, even if the rank value changes over time, the AI / ML model corresponding to a lower layer (e.g., layer 1) can continue updating the chain of accumulated information without resetting it. Furthermore, in a higher layer (e.g., layer 2), while avoiding the influence of past CSI information for which accumulated information is no longer valid, if the rank value changes over time, the chain of accumulated information can be updated using valid accumulated information (past / accumulated CSI) updated and output from the AI / ML model corresponding to a time (e.g., time instance B) with the same usable rank value in the past. Therefore, even if the rank or the number of layers changes over time, inconsistencies in the chain of accumulated information can be prevented.

[0140] Furthermore, in this embodiment, in the inference process at a certain time (time instance), only the AI / ML model with the number of layers corresponding to a specific rank (for example, a rank determined by rank selection) is executed, which can prevent an increase in power consumption or memory usage of the terminal 200 and the base station / network.

[0141] The set value of the time window or timer may be set to the same value across all ranks, or may be set to different values ​​depending on the rank. Furthermore, the time window or timer does not have to be set for a lower layer (e.g., layer 1). For example, the time window or timer may be set for a layer other than layer 1.

[0142] (Variation of Embodiment 6) In Embodiment 6, a case has been described in which a time window or a timer is applied to determine whether the accumulated information updated and output by the AI / ML model in the inference process of time instance B is valid, but other methods may be used to determine whether past accumulated information is valid.

[0143] For example, terminal 200 and base station / network may compare the reception quality (e.g., SINR) at time instance B with the SINR at the current time instance, and determine whether or not previously accumulated information is valid based on the difference in SINR. For example, terminal 200 and base station / network may determine that previously accumulated information is valid if the difference in SINR is equal to or less than a threshold, and may determine that previously accumulated information is invalid if the difference in SINR is greater than the threshold. Note that the threshold for the difference in SINR may be a common value for all layers, or may be an individual value for each layer.

[0144] Furthermore, for example, terminal 200 and base station / network may determine whether previously accumulated information is valid depending on the fluctuation range of rank between time instance B and the current time instance. For example, terminal 200 and base station / network may determine that previously accumulated information is valid if the fluctuation range of rank is equal to or less than a threshold, and may determine that previously accumulated information is invalid if the fluctuation range of rank is greater than the threshold. Note that the threshold for the fluctuation range of rank may be a common value for all layers, or may be a value individual to each layer.

[0145] Seventh Embodiment Terminal 200 inputs CSI information measured and generated by terminal 200 to a CSI compression unit (encoder), generates a CSI report (e.g., compressed CSI), and reports the CSI report to base station 100 (e.g., base station / network). The base station / network uses a CSI reconstruction unit (decoder) to reconstruct (restore) the CSI report (compressed CSI) reported from terminal 200. These processes can also be considered as inference processes using AI / ML models (two-sided models) on both the terminal side and the base station / network side.

[0146] Here, a CSI report configuration framework based on at least a subband configuration and a rank indication (e.g., a layer configuration) may be applied to a CSI report (e.g., a compressed CSI report) based on inference processing using an AI / ML model. Furthermore, the CSI report (e.g., a compressed CSI) may be transmitted in UCI on a PUCCH or a PUSCH. For example, a periodic CSI report, an aperiodic CSI report, or a semi-persistent CSI report may be used as the CSI report.

[0147] Terminal 200 may be notified of, for example, information regarding resources for CSI reports of at least one of Periodic CSI, Semi-persistent CSI, and Aperiodic CSI (e.g., information regarding resources such as PUCCH resources and PUSCH resources (e.g., CSI report configuration)) using RRC signaling, MAC signaling, downlink control information (DCI), or a combination thereof.

[0148] The CSI report configuration may include, for example, information regarding a reporting period, an offset, etc. Furthermore, the CSI report configuration may include a configuration ID (e.g., CSI-ReportConfigId). The configuration ID may identify parameters such as the type of CSI reporting method and the reporting period. Furthermore, the CSI report configuration may include information indicating which reference signal (e.g., CSI-RS) is used to report the measured CSI.

[0149] The input of the AI / ML model of the CSI compression unit (encoder) includes both CSI information (or precoding information) at the current time and accumulated information (past / accumulated CSI) from past time points, as shown in Figure 2 or 3. The output of the encoder is compressed CSI (or precoding) information, and the intermediate output of the encoder is accumulated information (past / accumulated CSI) to be used as input to the encoder (AI / ML model) at the next time point.

[0150] The CSI reconstruction unit (decoder), like the encoder, may restore the CSI (or precoding) information based on both the compressed CSI (or precoding) information and the accumulated information from the past (past / accumulated CSI) calculated on the decoder side, for example, as shown in FIG. 2 or FIG. 3 .

[0151] In this embodiment, for example, as shown in FIG. 18 , an AI / ML model is set for each layer, and the chain of accumulated information in the time domain (e.g., updates to accumulated information (intermediate output) of the AI / ML model) is made independent between layers. Furthermore, when an AI / ML model is set for each layer, for example, an AI / ML model is trained individually for each rank layer, and the terminal 200 and the base station / network may perform layer-specific inference processing (e.g., encoder processing or decoder processing) using an AI / ML model corresponding to each layer of each rank in inference processing (e.g., layer-specific and rank-specific). Furthermore, when an AI / ML model is set for each layer, for example, a common, unified AI / ML model is trained for each rank layer, and the terminal 200 and the base station / network may perform layer-specific inference processing for each rank using the unified AI / ML model in inference processing (e.g., layer common and rank specific). In either case, it is assumed that individual AI / ML models are trained for different ranks in a specific layer. For example, an AI / ML model is trained individually for each layer or commonly and individually for different ranks. In addition, the input to the AI / ML model of the CSI compression unit (encoder) may be information obtained by pre-processing the measured channel information and converting it into layer-specific information (e.g., precoding information such as eigenvalue vectors).

[0152] In this embodiment, in the inference process, if the rank value at a certain time (time instance; in the example of FIG. 18, time instance #t) is unchanged (the same) from the previous time (time instance #t-1 in the example of FIG. 18), terminal 200 and base station / network input the accumulated information updated and output in the AI / ML model at the previous time as accumulated information (past / accumulated CSI) that is additional input to the encoder or decoder to the AI / ML model of the CSI compression unit (encoder) or CSI reconstruction unit (decoder) corresponding to the layer of the rank at the certain time.

[0153] On the other hand, in the inference process, if the rank value at a certain time (time instance; in the example of FIG. 18, time instance #t+1) is changed (different) from the rank value at the previous time (in the example of FIG. 18, time instance #t), the terminal 200 and the base station / network may apply any of the following processes (any of Options 1 to 3).

[0154] <Option 1> The terminal 200 and the base station / network reset the stored information (e.g., do not use the stored information) or set the stored information to a neutral value (e.g., 0) and input it to the AI / ML model of the CSI compression unit (encoder) or the CSI reconstruction unit (decoder).

[0155] <Option 2> The terminal 200 and the base station / network use the accumulated information (past accumulated information) updated and output by the AI / ML model in the inference process of time instance A as accumulated information that is additional input to the encoder or decoder.

[0156] Note that the accumulated information of past times used in this embodiment is not limited to accumulated information updated and output by the AI / ML model in the inference process of time instance A (the time closest to the current time). For example, taking into account a processing delay in the terminal 200, accumulated information of the latest time available at a time (time instance) prior to the current time plus a certain processing delay time may be used.

[0157] <Option 3> The terminal 200 and the base station / network set the accumulated information to be input to the AI / ML model, depending on whether the accumulated information (e.g., accumulated information at past times) updated and output by the AI / ML model in the inference process for time instance A in Option 2 is valid. For example, the terminal 200 and the base station / network may select either the process of Option 2 or Option 1 depending on whether the accumulated information updated and output by the AI / ML model in the inference process for time instance A is valid.

[0158] For example, if the accumulated information updated and output by the AI / ML model in the inference process of time instance A is invalid, the terminal 200 and the base station / network perform processing in Option 1 (e.g., resetting the accumulated information (not using the accumulated information) or setting a neutral value (e.g., 0)).

[0159] On the other hand, if the accumulated information updated and output by the AI / ML model in the inference processing of time instance A is valid, terminal 200 and base station / network perform processing of Option 2 (e.g., processing to input the accumulated information updated and output by the AI / ML model in the inference processing of time instance A to the AI / ML model).

[0160] Here, a time window or timer may be applied to determine whether the accumulated information updated and output by the AI / ML model is valid in the inference process of time instance A. The set value of the time window or timer may be the same across all ranks and all layers, or may be different depending on the layer, different values ​​may be set depending on the rank, or different values ​​may be set depending on the combination of layer and rank.

[0161] Options 1 to 3 have been explained above.

[0162] According to this embodiment, when the rank value changes over time, the chain of accumulated information can be reset or updated without using (resetting) accumulated information or by using preset accumulated information (neutral value), and the chain of accumulated information can be updated using accumulated information updated and output from an AI / ML model corresponding to a time when the same rank value was available in the past. Therefore, even if the rank or the number of layers changes over time, it is possible to prevent inconsistencies from occurring in the chain of accumulated information.

[0163] Furthermore, in the inference process at a certain time instance, only the AI / ML model with the number of layers corresponding to a specific rank (e.g., the rank determined by rank selection) is executed, which can prevent an increase in power consumption of the terminal 200 and the base station / network.

[0164] (Variation of Embodiment 7) In Option 3 of Embodiment 7, a case has been described in which a time window or a timer is applied to determine whether the accumulated information updated and output by the AI / ML model in the inference process of time instance A is valid, but other methods may be used to determine whether past accumulated information is valid.

[0165] For example, terminal 200 and base station / network may compare the reception quality (e.g., SINR) at time instance A with the SINR at the current time instance, and determine whether or not previously accumulated information is valid based on the difference in SINR. For example, terminal 200 and base station / network may determine that previously accumulated information is valid if the difference in SINR is equal to or smaller than a threshold, and may determine that previously accumulated information is invalid if the difference in SINR is greater than the threshold. Note that the threshold for the difference in SINR may be a common value for all layers, or may be an individual value for each layer.

[0166] Furthermore, for example, terminal 200 and base station / network may determine whether previously accumulated information is valid based on the fluctuation range of rank between time instance A and the current time instance. For example, terminal 200 and base station / network may determine that previously accumulated information is valid if the fluctuation range of rank is equal to or less than a threshold, and may determine that previously accumulated information is invalid if the fluctuation range of rank is greater than the threshold. Note that the threshold for the fluctuation range of rank may be a common value for all layers, or may be a separate value for each layer.

[0167] The first to seventh embodiments have been described above.

[0168] Thus, in a non-limiting example of the present disclosure, terminal 200 controls input of past accumulated information (past / accumulated CSI) to a terminal-side AI / ML model that generates a CSI report based on the rank value at the current time and the rank value at the immediately preceding time, and transmits a CSI report generated using the accumulated information. Furthermore, base station 100 receives a CSI report generated by inference processing of the terminal-side AI / ML model, and controls input of past accumulated information to a base station / network-side AI / ML model that reconstructs a CSI report based on the rank value at the current time and the rank value at the immediately preceding time.

[0169] As a result, even if the rank or number of layers changes over time in rank adaptation, for example, the terminal 200 and the base station / network can suppress inconsistencies in the chain of past accumulated information (past / accumulated CSI) input to the AI / ML model, thereby improving the performance of CSI compression.

[0170] Furthermore, in a non-limiting example of the present disclosure, inference processing for CSI compression or reconstruction is performed using only the AI / ML model that is actually used at each time instance. This prevents an increase in power consumption or memory usage of the terminal 200 and the base station / network without having to continuously operate all AI / ML models.

[0171] Therefore, according to non-limiting examples of the present disclosure, the efficiency of wireless communication can be improved.

[0172] (Other Embodiment 1) In the above-described embodiment, the time window or timer is not limited to the case where the rank value is changed, but can also be applied when the rank value is not changed. For example, when a UCI drop or layer omission occurs on the terminal side, past accumulated information (past / accumulated CSI) may not always be available even when the rank value is not changed. Therefore, even when the rank value is not changed, if accumulated information that is an additional input to the AI / ML model for the corresponding rank or layer cannot be obtained from the previous time instance, the operation when the rank value is changed in Embodiments 3, 6, or 7 (Option 3) may be applied.

[0173] (Another embodiment 2) When an AI / ML model is used on the terminal 200 side and the base station / network side, learning may be coordinated between vendors. When learning is coordinated between vendors, for example, as shown in Fig. 19, a method is possible in which a base station / network (gNB / NW) learns an encoder and decoder model, and transfers a data set used and generated in the learning to the terminal 200 (UE), and the terminal 200 uses the acquired data set to learn an encoder model (or a reference decoder model) on the terminal side.

[0174] In this method, the data set transferred from the base station / network to the terminal 200 may include, for example, the following data:

[0175] The dataset for encoder model training may include, for example, time-series data for each of target CSI (CSI input to the encoder), rank value, and CSI feedback (encoder output corresponding to compressed CSI).

[0176] The data set for training the reference decoder model may include, for example, time-series data of the rank value, the CSI feedback (encoder output corresponding to compressed CSI), and the reconstructed target CSI (decoder output). After training the reference decoder model, the terminal 200 may train the encoder model using the reference decoder model.

[0177] The dataset for training the encoder model and the reference decoder model may include, for example, time series data for the target CSI (CSI input to the encoder), rank value, CSI feedback (encoder output corresponding to compressed CSI), and reconstructed target CSI (decoder output).

[0178] In either case, at least time series data of the selected rank value (rank fluctuation) may be included in the data set exchanged from the base station / network to terminal 200. By including the rank value in the data set, it is possible to avoid inconsistencies in rank selection between the encoder and decoder.

[0179] [Example of Operation of Base Station / Network and Terminal 200] FIG. 20 is a flowchart showing an example of operation of terminal 200.

[0180] The terminal 200 performs CSI measurement (S101) and selects a rank (S102).

[0181] The terminal 200 determines whether the rank at the current time (time instance) is the same as the rank at the immediately previous time (S103).

[0182] If the rank at the current time is the same as the rank at the immediately previous time (if it is not changed) (S103: Yes), the terminal 200 uses the accumulated information updated and output in the AI / ML model at the immediately previous time as additional input to the AI / ML model (S104).

[0183] On the other hand, if the rank at the current time is not the same as the rank at the immediately previous time (if it has been changed) (S103: No), the terminal 200 uses the accumulated information as additional input to the AI / ML model by varying the accumulated information according to the rank, for example, according to the operation example of the first, second, third, or seventh embodiment (S105).

[0184] The base station / network may set accumulated information to be additionally input to the AI / ML model in a manner similar to the example operation of the terminal 200 described above.

[0185] Also, in the fourth, fifth and sixth embodiments, the base station / network and terminal 200 may similarly set accumulated information to be additionally input to the AI / ML model.

[0186] [Regarding Architecture] Fig. 21 shows an example of an architecture including a function for processing AI / ML. Note that part of the network architecture may have the configuration shown in Fig. 21. Furthermore, the function for processing AI / ML may be included in an Access and Mobility Management Function (AMF) or a RAN (gNB).

[0187] [Configuration of Base Station] Fig. 22 is a block diagram showing an example configuration of base station 100. In Fig. 22, base station 100 has a control unit 101, a signal generation unit 102, a transmission unit 103, a reception unit 104, an extraction unit 105, a demodulation unit 106, and a decoding unit 107.

[0188] At least one of the control unit 101, the signal generation unit 102, the extraction unit 105, the demodulation unit 106, and the decoding unit 107 shown in Fig. 22 may be included in the control unit shown in Fig. 10. Also, at least one of the transmission unit 103 and the reception unit 104 shown in Fig. 22 may be included in the communication unit shown in Fig. 10.

[0189] The control unit 101, for example, determines control information related to CSI measurement and reporting of the terminal 200 and outputs the determined information to the signal generation unit 102. The control information related to CSI measurement and reporting may include, for example, the configuration of a reference signal for CSI measurement, such as CSI-RS, and information related to a CSI report configuration for the terminal 200 to report CSI. Furthermore, the control unit 101 may output the CSI report input from the decoding unit 107 to a function for processing AI / ML (for example, a data collection function or a learning function of an AI / ML model). The function for processing AI / ML may be included in the base station 100 or in a node different from the base station 100.

[0190] Furthermore, the control unit 101 determines, for example, information for the terminal 200 to receive a downlink signal, and outputs the determined information to the signal generation unit 102. The information for the terminal 200 to receive a downlink signal may include, for example, information on resource allocation of a downlink data channel (e.g., PDSCH: Physical Downlink Shared Channel) or a downlink control channel (e.g., PDCCH: Physical Downlink Control Channel), and information on a coding and modulation scheme (e.g., MCS: Modulation and Coding Scheme).

[0191] Furthermore, the control unit 101 determines, for example, information used by the terminal 200 to transmit an uplink signal (e.g., a CSI report), and outputs the determined information to the signal generation unit 102, the extraction unit 105, the demodulation unit 106, and the decoding unit 107. The information used by the terminal 200 to transmit an uplink signal may include, for example, information on resource allocation of an uplink data channel (e.g., a Physical Uplink Shared Channel (PUSCH)) or an uplink control channel (e.g., a Physical Uplink Control Channel (PUCCH)), information on a coding / modulation scheme (e.g., MCS), and information on a CSI report.

[0192] The signal generation unit 102 generates a data signal or a control signal bit sequence using, for example, information input from the control unit 101, and applies encoding as necessary. The signal generation unit 102 also modulates the encoded bit sequence to generate a modulated signal (for example, a symbol sequence), and maps the modulated signal to the radio resource specified by the control unit 101. The signal generation unit 102 outputs the mapped signal to the transmission unit 103.

[0193] The transmitting unit 103 performs, for example, OFDM or other transmission waveform generation processing on the signal input from the signal generating unit 102. Furthermore, in the case of OFDM transmission using a cyclic prefix (CP), for example, the transmitting unit 103 performs an inverse fast Fourier transform (IFFT) processing on the signal and adds a CP to the signal after the IFFT. Furthermore, the transmitting unit 103 performs, for example, RF processing such as D / A conversion or up-conversion on the signal and transmits the radio signal to the terminal 200 via an antenna.

[0194] The receiving unit 104 performs RF processing such as downconvert or A / D conversion on an uplink signal received from the terminal 200 via an antenna. In addition, in the case of OFDM transmission, the receiving unit 104 performs Fast Fourier Transform (FFT) processing on the received signal, and outputs the resulting frequency domain signal to the extracting unit 105.

[0195] The extraction unit 105 extracts, for example, based on information input from the control unit 101, a radio resource portion from which an uplink signal (for example, a PUSCH or a PUCCH) is transmitted, from the received signal input from the receiving unit 104, and outputs the extracted radio resource portion to the demodulation unit 106.

[0196] The demodulation unit 106 demodulates the uplink signal (for example, PUSCH or PUCCH) input from the extraction unit 105, for example, based on information input from the control unit 101. The demodulation unit 106 outputs the demodulation result to the decoding unit 107, for example.

[0197] The decoding unit 107 performs error correction decoding on the uplink signal (for example, PUSCH or PUCCH) based on, for example, the information input from the control unit 101 and the demodulation result input from the demodulation unit 106, and obtains a decoded received bit sequence. For example, if the decoded received bit sequence includes a CSI report from the terminal 200, the decoding unit 107 outputs the information to the control unit 101.

[0198] [Terminal Configuration] Fig. 23 is a block diagram showing an exemplary configuration of a terminal 200 according to an embodiment of the present disclosure. For example, in Fig. 23, the terminal 200 includes a receiving unit 201, an extracting unit 202, a demodulating unit 203, a decoding unit 204, a control unit 205, a signal generating unit 206, and a transmitting unit 207.

[0199] At least one of the extraction unit 202, demodulation unit 203, decoding unit 204, control unit 205, and signal generation unit 206 shown in Fig. 23 may be included in the control unit shown in Fig. 11. Also, at least one of the reception unit 201 and transmission unit 207 shown in Fig. 23 may be included in the communication unit shown in Fig. 11.

[0200] The receiving unit 201 receives, for example, a downlink signal (e.g., a downlink data signal or a downlink control signal) from the base station 100 via an antenna, and performs RF processing such as downconvert or A / D conversion on the radio received signal to obtain a received signal (baseband signal). Furthermore, when receiving an OFDM signal, the receiving unit 201 performs FFT processing on the received signal to convert it into the frequency domain. The receiving unit 201 outputs the received signal to the extracting unit 202.

[0201] The extraction unit 202 extracts a radio resource portion that may include a downlink control signal from the received signal input from the receiving unit 201, based on, for example, information related to the radio resource of the downlink control signal input from the control unit 205, and outputs the extracted radio resource portion to the demodulation unit 203. Furthermore, based on information related to the radio resource of the data signal input from the control unit 205, the extraction unit 202 extracts a radio resource portion that includes a downlink data signal from the received signal, and outputs the extracted radio resource portion to the demodulation unit 203. Furthermore, the extraction unit 202 extracts a radio resource portion that includes a reference signal for CSI measurement (e.g., CSI-RS) from the received signal, and outputs the extracted radio resource portion to the control unit 205.

[0202] The demodulation unit 203 demodulates the signal (for example, PDCCH or PDSCH) input from the extraction unit 202 based on information input from the control unit 205 , for example, and outputs the demodulation result to the decoding unit 204 .

[0203] The decoding unit 204 performs error correction decoding of the PDCCH or PDSCH using, for example, information input from the control unit 205 and the demodulation result input from the demodulation unit 203, and obtains, for example, a control signal or a downlink data signal. The decoding unit 204 outputs the control signal to the control unit 205.

[0204] The control unit 205 identifies information related to downlink transmission based on, for example, information obtained from a control signal input from the decoding unit 204, and outputs the information to the extraction unit 202, the demodulation unit 203, and the decoding unit 204. The control unit 205 also identifies information related to uplink transmission based on, for example, information obtained from a control signal input from the decoding unit 204, and outputs the information to the signal generation unit 206. The control unit 205 also performs CSI measurement using the CSI-RS input from the extraction unit 202, generates control information related to CSI reporting based on the CSI measurement results using the above-described method, and outputs the control information to the signal generation unit 206. The control unit 205 may also output information related to CSI reporting to a function that processes AI / ML based on the CSI measurement results.

[0205] The signal generation unit 206 generates an uplink data signal or an uplink control signal based on control information related to the CSI report input from the control unit 205, information related to uplink transmission, or information related to the coded CSI report input from the AI / ML processing function, encodes and modulates the bit string of the generated signal, and maps it to radio resources. The signal generation unit 206 outputs the uplink signal onto which the signal has been mapped to, for example, the transmission unit 207.

[0206] The transmitter 207 generates a transmission signal waveform, such as OFDM, for the signal input from the signal generator 206. Furthermore, in the case of OFDM transmission or DFT-s-OFDM transmission using a CP, for example, the transmitter 207 performs IFFT processing on the signal and adds a CP to the signal after IFFT. Alternatively, when the transmitter 207 generates a single-carrier waveform such as a DFT-s-OFDM waveform, a DFT unit (not shown) may be added before the signal generator 206. Furthermore, the transmitter 207 performs RF processing, such as D / A conversion and up-conversion, on the transmission signal, and transmits the radio signal to the base station 100 via an antenna.

[0207] The above describes the embodiments according to non-limiting examples of the present disclosure.

[0208] [Use Cases] The use cases to which the above-described embodiments or modifications are applied are not limited to CSI compression (CSI feedback enhancement), and can be applied to any use case in which an AI / ML model is deployed on the terminal side and the base station / network side. For example, examples of use cases in which an AI / ML model is deployed on the terminal side and the base station / network side include CSI prediction using an AI / ML model, beam prediction in the spatial or time domain using an AI / ML model, positioning accuracy enhancement using an AI / ML model, and optimization of parameters and configurations of a wireless interface using AI / ML.

[0209] Furthermore, if the processor or AI / ML accelerator is general-purpose for AI / ML processing, AI / ML processing for different use cases may be executed simultaneously, or multiple AI / ML processing for different use cases may be selected.

[0210] [AI / ML Model] In non-limiting examples of the present disclosure, an "AI / ML model (or model)" may include a physical model, a binary model, an executable model, a converted model, a source code model, a non-executable model, a raw model, a logical model, or other model formats.

[0211] In addition, different AI / ML models may be used depending on parameters such as the cell, site, Transmission and Reception Point (TRP), beam, location, terminal movement speed, wireless channel such as multipath, cell congestion status, and transmitted traffic.

[0212] Furthermore, in the present disclosure, the signal / message / signaling used for notification may be a control plane message (e.g., UCI or MAC-CE), an RRC signal, or a notification in DCI, which is physical layer signaling.

[0213] Furthermore, the parameter values ​​used in the above embodiment are merely examples, and other values ​​may be used.

[0214] (Supplementary Note) Information indicating whether the terminal 200 supports the functions, operations, or processes described in each of the above-described embodiments and each supplementary note may be transmitted (or notified) from the terminal 200 to the base station 100, for example, as capability information or capability parameters of the terminal 200.

[0215] The capability information may include an information element (IE) that individually indicates whether or not the terminal 200 supports at least one of the functions, operations, or processes described in the above-described embodiments, modifications, and supplements. Alternatively, the capability information may include an information element that indicates whether or not the terminal 200 supports a combination of any two or more of the functions, operations, or processes described in the above-described embodiments, modifications, and supplements.

[0216] For example, the base station 100 may determine (or decide or assume) functions, operations, or processes that the terminal 200 that transmitted the capability information supports (or does not support) based on the capability information received from the terminal 200. The base station 100 may perform operations, processes, or controls according to the determination result based on the capability information. For example, the base station 100 may control processing related to CSI reporting based on the capability information received from the terminal 200.

[0217] Note that the fact that terminal 200 does not support some of the functions, operations, or processes described in the above-described embodiments, modifications, and supplementary notes may be interpreted as meaning that such some of the functions, operations, or processes are restricted in terminal 200. For example, information or a request regarding such restrictions may be notified to base station 100.

[0218] Information regarding the capabilities or limitations of terminal 200 may, for example, be defined in a standard, or may be implicitly notified to base station 100 in association with information known at base station 100 or information transmitted to base station 100.

[0219] The above has described the embodiments, modifications, and supplementary notes according to a non-limiting example of the present disclosure.

[0220] (Control Signal) In the present disclosure, a downlink control signal (or downlink control information) related to an embodiment of the present disclosure may be, for example, a signal (or information) transmitted in a Physical Downlink Control Channel (PDCCH) of a physical layer, or a signal (or information) transmitted in a Medium Access Control Control Element (MAC CE) or Radio Resource Control (RRC) of a higher layer. Furthermore, the signal (or information) is not limited to being notified by a downlink control signal, but may be predefined in a specification (or standard) or preconfigured in a base station and a terminal.

[0221] In the present disclosure, an uplink control signal (or uplink control information) related to an embodiment of the present disclosure may be, for example, a signal (or information) transmitted in a PUCCH of a physical layer, or a signal (or information) transmitted in a MAC CE or RRC of a higher layer. Furthermore, the signal (or information) is not limited to being notified by an uplink control signal, but may be predefined in a specification (or standard) or preconfigured in a base station and a terminal. Furthermore, the uplink control signal may be replaced with, for example, uplink control information (UCI), 1st stage sidelink control information (SCI), or 2nd stage SCI.

[0222] (Base Station) In an embodiment of the present disclosure, the base station may be a Transmission Reception Point (TRP), a cluster head, an access point, a Remote Radio Head (RRH), an eNodeB (eNB), a gNodeB (gNB), a Base Station (BS), a Base Transceiver Station (BTS), a parent device, a gateway, or the like. In sidelink communication, a terminal may play the role of a base station. Instead of a base station, a relay device that relays communication between an upper node and a terminal may be used. Alternatively, a roadside unit may be used.

[0223] (Uplink / Downlink / Sidelink) An embodiment of the present disclosure may be applied to, for example, any of the uplink, downlink, and sidelink. For example, an embodiment of the present disclosure may be applied to a Physical Uplink Shared Channel (PUSCH), a Physical Uplink Control Channel (PUCCH), or a Physical Random Access Channel (PRACH) in the uplink, a Physical Downlink Shared Channel (PDSCH), a PDCCH, or a Physical Broadcast Channel (PBCH) in the downlink, or a Physical Sidelink Shared Channel (PSSCH), a Physical Sidelink Control Channel (PSCCH), or a Physical Sidelink Broadcast Channel (PSBCH) in the sidelink.

[0224] The PDCCH, PDSCH, PUSCH, and PUCCH are examples of a downlink control channel, a downlink data channel, an uplink data channel, and an uplink control channel, respectively. The PSCCH and PSSCH are examples of a sidelink control channel and a sidelink data channel. The PBCH and PSBCH are examples of a broadcast channel, and the PRACH is an example of a random access channel.

[0225] (Data Channel / Control Channel) An embodiment of the present disclosure may be applied to, for example, either a data channel or a control channel. For example, the channel in an embodiment of the present disclosure may be replaced with any of the data channels PDSCH, PUSCH, and PSSCH, or the control channels PDCCH, PUCCH, PBCH, PSCCH, and PSBCH.

[0226] (Reference Signal) In one embodiment of the present disclosure, a reference signal is, for example, a signal known by both a base station and a mobile station, and may also be called a Reference Signal (RS) or a pilot signal. The reference signal may be any of a Demodulation Reference Signal (DMRS), a Channel State Information - Reference Signal (CSI-RS), a Tracking Reference Signal (TRS), a Phase Tracking Reference Signal (PTRS), a Cell-specific Reference Signal (CRS), or a Sounding Reference Signal (SRS).

[0227] (Time Interval) In one embodiment of the present disclosure, the unit of time resource is not limited to one or a combination of slots and symbols, but may be, for example, a time resource unit such as a frame, a superframe, a subframe, a slot, a time slot, a subslot, a minislot, a symbol, an Orthogonal Frequency Division Multiplexing (OFDM) symbol, a Single Carrier-Frequency Division Multiplexing Access (SC-FDMA) symbol, or another time resource unit. Furthermore, the number of symbols included in one slot is not limited to the number of symbols exemplified in the above-mentioned embodiment, and may be another number of symbols.

[0228] (Frequency Band) An embodiment of the present disclosure may be applied to either a licensed band or an unlicensed band.

[0229] (Communication) An embodiment of the present disclosure may be applied to communication between a base station and a terminal (Uu link communication), communication between terminals (Sidelink communication), or Vehicle to Everything (V2X) communication. For example, the channel in an embodiment of the present disclosure may be replaced with any of PSCCH, PSSCH, Physical Sidelink Feedback Channel (PSFCH), PSBCH, PDCCH, PUCCH, PDSCH, PUSCH, or PBCH.

[0230] An embodiment of the present disclosure may be applied to a terrestrial network, a non-terrestrial network (NTN) using a satellite or a high altitude pseudo satellite (HAPS), or a terrestrial network in which transmission delay is large compared to the symbol length or slot length, such as a network with a large cell size or an ultra-wideband transmission network.

[0231] (SBFD) In ​​one embodiment of the present disclosure, operations on uplink, downlink, and sidelink symbols may be applied to symbols (e.g., SBFD symbols) on which SBFD (Subband Non-Overlapping Full Duplex, Subband Full Duplex) operations or controls are performed. In SBFD symbols, a frequency domain (or frequency resource, frequency band) is divided into multiple frequency domains (e.g., subbands, RB sets, subbands, or sub-BWPs (Bandwidth Parts)). A terminal transmits and receives in different directions (e.g., downlink or uplink) in units of subbands, which are the divided domains. In SBFD symbols, a terminal may transmit and receive in one direction, either uplink or downlink, but not in the other direction. On the other hand, a base station may be capable of transmitting and receiving on both the uplink and downlink simultaneously. SBFD symbols may have a smaller frequency domain available for downlink use than symbols that transmit and receive only downlink use. Also, SBFD symbols may have a smaller frequency domain available for uplink use than symbols that transmit and receive only uplink use.

[0232] In addition, in the SBFD symbol, a terminal may transmit and receive uplink and downlink simultaneously. In this case, the frequency domain in which the terminal transmits and the frequency domain in which the terminal receives may not be adjacent, but may be separated by a frequency interval (also called a frequency gap).

[0233] In addition, different transmission and reception directions in subband units, which are divided areas, may include transmission and reception of side links.

[0234] (XDD: Cross Division Duplex) In one embodiment of the present disclosure, the operation for uplink, downlink, and sidelink symbols may be applied to symbols (e.g., full duplex symbols) where full duplex operation or control is performed. In a full duplex symbol, both the terminal and the base station can simultaneously transmit and receive on the uplink and downlink. In a full duplex symbol, the terminal and the base station may simultaneously transmit and receive in an available frequency region (or frequency resource, frequency band), or may simultaneously transmit and receive in a partial frequency region (i.e., transmission or reception may be performed in other frequency regions). In this case, the frequency region in which the base station or terminal transmits and receives may not be adjacent, but may have a frequency interval (also called a frequency gap). Furthermore, for the purpose of, for example, reducing interference, either the terminal or the base station may simultaneously transmit and receive (i.e., the other may transmit or receive).

[0235] In addition, full duplex operation may be applied to an operation in which a terminal can simultaneously transmit and receive sidelinks, or to an operation in which a terminal can simultaneously transmit and receive sidelinks and uplinks or downlinks.

[0236] (Antenna Port) In one embodiment of the present disclosure, an antenna port refers to a logical antenna (antenna group) consisting of one or more physical antennas. For example, an antenna port does not necessarily refer to a single physical antenna, but may refer to an array antenna consisting of multiple antennas. For example, the number of physical antennas that an antenna port is composed of is not specified, and the antenna port may be specified as the smallest unit by which a terminal station can transmit a reference signal. Furthermore, an antenna port may also be specified as the smallest unit by which a weighting of a precoding vector is multiplied.

[0237] <5G NR System Architecture and Protocol Stack> The 5G NR system architecture generally assumes an NG-RAN (Next Generation - Radio Access Network) including gNBs. The gNBs provide UE-side termination of the NG radio access user plane (SDAP / PDCP / RLC / MAC / PHY) and control plane (RRC) protocols. The gNBs are connected to each other via an Xn interface. The gNBs are also connected to the Next Generation Core (NGC) via a Next Generation (NG) interface, more specifically to the Access and Mobility Management Function (AMF) (e.g., a specific core entity that performs AMF) via an NG-C interface, and to the User Plane Function (UPF) (e.g., a specific core entity that performs UPF) via an NG-U interface. The NG-RAN architecture is shown in Figure 24 (see, for example, 3GPP TS 38.300 v15.6.0, section 4).

[0238] <RRC connection setup and reconfiguration procedure> This shows the NAS part of the interaction between the UE, gNB, and AMF (5GC entity) when the UE transitions from RRC_IDLE to RRC_CONNECTED (see TS 38.300 v15.6.0).

[0239] RRC is a higher layer signaling protocol used to configure the UE and gNB. The AMF prepares UE context data (including, for example, PDU session context, security keys, UE radio capabilities, UE security capabilities, etc.) and sends it to the gNB along with an INITIAL CONTEXT SETUP REQUEST. The gNB then activates AS security together with the UE. This is done by the gNB sending a SecurityModeCommand message to the UE, and the UE responding with a SecurityModeComplete message to the gNB. The gNB then sends an RRCReconfiguration message to the UE, and upon receiving an RRCReconfigurationComplete from the UE, the gNB performs reconfiguration to set up Signaling Radio Bearer 2 (SRB2) and Data Radio Bearer (DRB). For signaling-only connections, the steps related to RRCReconfiguration are omitted because SRB2 and DRB are not set up. Finally, the gNB notifies the AMF that the setup procedure is complete with an INITIAL CONTEXT SETUP RESPONSE.

[0240] Therefore, the present disclosure provides a 5th Generation Core (5GC) entity (e.g., AMF, SMF, etc.) that includes: a control circuit that, upon operation, establishes a Next Generation (NG) connection with a gNodeB; and a transmitter that, upon operation, transmits an initial context setup message to the gNodeB via the NG connection so that a signaling radio bearer between the gNodeB and a user equipment (UE) is set up. Specifically, the gNodeB transmits Radio Resource Control (RRC) signaling, including a resource allocation configuration information element (IE), to the UE via the signaling radio bearer. The UE then transmits in uplink or receives in downlink based on the resource allocation configuration.

[0241] <QoS Control> The 5G Quality of Service (QoS) model is based on QoS flows and supports both QoS flows that require a guaranteed flow bit rate (Guaranteed Bit Rate QoS flows (GBR)) and QoS flows that do not require a guaranteed flow bit rate (non-GBR QoS flows). Thus, at the NAS level, a QoS flow is the finest granularity of QoS classification in a PDU session. A QoS flow is identified within a PDU session by a QoS Flow ID (QFI) carried in an encapsulation header over the NG-U interface.

[0242] For each UE, 5GC establishes one or more PDU sessions. For each UE, the NG-RAN establishes, for example, at least one Data Radio Bearer (DRB) for each PDU session. Additional DRBs for the QoS flows of that PDU session can be configured later (when this is up to the NG-RAN). The NG-RAN maps packets belonging to different PDU sessions to different DRBs. NAS-level packet filters in the UE and 5GC associate UL and DL packets with QoS flows, while AS-level mapping rules in the UE and NG-RAN associate UL and DL QoS flows with DRBs.

[0243] (Open-RAN) The base station described in each embodiment (for example, a 5G NR base station called a gNB) may be configured with three functional modules: a Centralized Unit (CU), a Distributed Unit (DU), and a Radio Unit (RU).

[0244] A CU may be referred to as a centralized node, aggregation node, central station, aggregation station, or centralized unit. A DU may be referred to as an O-RAN Distributed Unit (O-DU), distributed node, distributed station, or distributed unit. An RU may be referred to as an O-RAN Radio Unit (O-RU), radio equipment, radio node, radio station, antenna unit, or radio unit.

[0245] There are several split options for the functional split configuration (or functional split point) between CU, DU, and RU. The term "functional split point" is sometimes referred to as "split," "option," or "split option."

[0246] Examples of "division options" include the following division options 1 to 8. The functions of the base station described in each embodiment may be divided into a CU, a DU, and an RU by any of the following division options 1 to 8. For example, the CU, DU, and RU may be functionally divided, or the functions may be divided only between the CU and DU or only between the DU and RU. (1) Segmentation option 1: Between RRC (radio resource control) and PDCP (2) Segmentation option 2: Between PDCP and RLC (High-RLC) (3) Segmentation option 3: Between High-RLC and Low-RLC (4) Segmentation option 4: Between RLC (Low-RLC) and MAC (High-MAC) (5) Segmentation option 5: Between High-MAC and Low-MAC (6) Segmentation option 6: Between MAC (Low-MAC) and PHY (High-PHY) (7) Segmentation option 7: Between High-PHY and Low-PHY (8) Segmentation option 8: Between PHY (Low-PHY) and RF

[0247] The functional split point between the CU and O-DU may be split option 2. The section between the CU and O-DU is called midhaul, and the F1 interface is specified by 3GPP. The section between the O-DU and O-RU is called fronthaul, and the functional split point may be split option 7-2x, which is adopted as the O-RAN fronthaul specification.

[0248] Figure 25 shows an example of functionally dividing the gNB base station functions into CU, O-DU, and O-RU using Split Option 2 and Split Option 7-2x.

[0249] The CU may have, for example, a radio resource control (RRC) function, a service data adaptation protocol (SDAP) function, and a packet data convergence protocol (PDCP) function.

[0250] The O-DU may include, for example, a radio link control (RLC) function, a MAC function, and a higher physical layer (HIGH-PHY) function. The HIGH-PHY function may include an encoding function, a scrambling function, a modulation function, a layer mapping function, a precoding function, and a resource element (RE) mapping function for downlink (DL) transmission. The HIGH-PHY function may also include a decoding function, a descrambling function, a demodulation function, a layer demapping function, and a resource element (RE) demapping function for uplink (UL) reception.

[0251] The O-RU may have, for example, a LOW-PHY function and an RF function. The LOW-PHY function may also have, for downlink transmission, a beamforming function, an IFFT (Inverse First Fourier Transform) + CP (Cyclic Prefix) assignment function, and a D / A (Digital to Analog) conversion function. The LOW-PHY function may also have, for uplink reception, an A / D (Analog to Digital) conversion function, a CP removal + FFT (First Fourier Transform) function, and a beamforming function.

[0252] In addition, if the O-DU does not have a precoding function, the O-RU may have a precoding function.

[0253] The O-RU may have functionality related to LBT (listen before talk).

[0254] The evolving Common Public Radio Interface (eCPRI) is specified as the communication method between the O-DU and O-RU in Split Option 7-2x. In Split Option 7-2x, eCPRI transmits and receives sampling sequences of the in-phase (I) and quadrature (Q) components of OFDM signals in the frequency domain, as well as information used for beamforming in antennas and time synchronization signals.

[0255] Information transmitted by the signals described in each embodiment (PDCCH, PUCCH, PDSCH, PUSCH, MAC CE, RRC, etc.) may be transmitted between the O-DU and the O-RU via the eCPRI User Plane (U-Plane) or Control Plane (C-Plane).

[0256] When the functions described in each embodiment are performed in the O-RU by functional division, the O-DU may control the O-RU by transmitting information for controlling the functions via a control signal (e.g., eCPRI) between the O-DU and the O-RU.

[0257] When the functions described in each embodiment are performed in the O-DU by functional division, the O-RU may receive the results of the functions performed in the O-DU via a control signal (e.g., eCPRI) and control the O-RU based on the received results.

[0258] The CU, O-DU, and O-RU may be deployed in physically different devices with their respective functions connected by optical fiber or the like, or some or all of their functions may be deployed in the same physical device.

[0259] The CU and O-DU may be logical entities implemented as software running on a server in the cloud or the like as a virtualized RAN (virtual Radio Access Network: vRAN). Also, some or all of the functions of the CU and O-DU may be provided as a virtualized network function (Network Functions Virtualization: NFV) service.

[0260] The transceiver does not have to be a radio transceiver, but may be, for example, a network transceiver, an optical transceiver, etc. The radio resources allocated by the O-DU may be resources for wireless communication between the O-RU and the UE.

[0261] The present disclosure can be realized in software, hardware, or software in conjunction with hardware.

[0262] Each functional block used in the description of the above embodiments may be partially or entirely realized as an LSI, which is an integrated circuit, and each process described in the above embodiments may be partially or entirely controlled by a single LSI or a combination of LSIs. The LSI may be composed of individual chips, or may be composed of a single chip that includes some or all of the functional blocks. The LSI may have data input and output. Depending on the degree of integration, the LSI may also be called an IC, system LSI, super LSI, or ultra LSI.

[0263] The integrated circuit method is not limited to LSI, and may be realized by a dedicated circuit, a general-purpose processor, or a dedicated processor. Also, a field programmable gate array (FPGA) that can be programmed after LSI manufacturing, or a reconfigurable processor that can reconfigure the connections and settings of circuit cells within the LSI, may be used. The present disclosure may be realized as digital processing or analog processing.

[0264] Furthermore, if an integrated circuit technology that can replace LSI emerges due to advances in semiconductor technology or other derivative technologies, it is natural that such technology may be used to integrate functional blocks. The application of biotechnology, etc. is also a possibility.

[0265] The present disclosure may be implemented in any type of apparatus, device, or system (collectively referred to as a communications apparatus) that has a communications function. The communications apparatus may include a radio transceiver and processing / control circuitry. The radio transceiver may include a receiver and a transmitter, or both functions. The radio transceiver (transmitter and receiver) may include a radio frequency (RF) module and one or more antennas. The RF module may include an amplifier, an RF modulator / demodulator, or the like. Non-limiting examples of communication devices include telephones (e.g., cell phones, smartphones), tablets, personal computers (PCs) (e.g., laptops, desktops, notebooks), cameras (e.g., digital still / video cameras), digital players (e.g., digital audio / video players), wearable devices (e.g., wearable cameras, smartwatches, tracking devices), game consoles, digital book readers, telehealth / telemedicine devices, communication-enabled vehicles or mobile transportation (e.g., cars, airplanes, ships), and combinations of the above devices.

[0266] The communication devices are not limited to portable or mobile devices, but also include any kind of non-portable or fixed equipment, devices, and systems, such as smart home devices (such as home appliances, lighting equipment, smart meters or measuring devices, control panels, etc.), vending machines, and any other "things" that may exist on an IoT (Internet of Things) network.

[0267] Communications include data communications via cellular systems, wireless LAN systems, communication satellite systems, and the like, as well as data communications via combinations of these.

[0268] A communications apparatus also includes devices such as controllers and sensors connected or coupled to a communications device that performs the communications functions described in this disclosure, such as controllers and sensors that generate control and data signals used by the communications device to perform the communications functions of the communications apparatus.

[0269] The communication apparatus also includes infrastructure facilities, such as base stations, access points, and any other apparatus, device, or system that communicates with or controls the various apparatuses listed above, but are not limited to these.

[0270] A terminal according to one embodiment of the present disclosure includes a control circuit that controls input of an intermediate output from an artificial intelligence model on the terminal side at a past time to an artificial intelligence model on the terminal side that generates a channel state information report based on a first rank value at the current time and a second rank value at the previous time, and a transmitting circuit that transmits the channel state information report generated using the intermediate output.

[0271] In one embodiment of the present disclosure, when the first rank value and the second rank value are the same, the control circuit inputs the intermediate output of the immediately previous time to the artificial intelligence model.

[0272] In one embodiment of the present disclosure, if the first rank value and the second rank value are different, the control circuit discards the intermediate output or inputs a value common between the terminal and the base station as the intermediate output to the artificial intelligence model.

[0273] In one embodiment of the present disclosure, when the first rank value and the second rank value are different, the control circuit inputs the intermediate output from the past time when the first rank value was set into the artificial intelligence model.

[0274] In one embodiment of the present disclosure, when the first rank value and the second rank value are different, the control circuit sets the intermediate output depending on whether the intermediate output at the past time when the first rank value was set was valid.

[0275] In one embodiment of the present disclosure, if the intermediate output is valid, the control circuit inputs the intermediate output to the artificial intelligence model, and if the intermediate output is not valid, discards the intermediate output or inputs a value common between the terminal and the base station as the intermediate output to the artificial intelligence model.

[0276] In one embodiment of the present disclosure, the artificial intelligence model is set for each rank, the intermediate output input to the artificial intelligence model is updated independently between ranks, and the control circuit performs inference processing individual to the rank using the artificial intelligence model corresponding to the selected rank.

[0277] In one embodiment of the present disclosure, the artificial intelligence model is set for each layer, updates of the intermediate outputs input to the artificial intelligence model are performed independently between layers, and the control circuit uses the artificial intelligence model to perform inference processing individual to each layer.

[0278] In one embodiment of the present disclosure, the artificial intelligence model is trained separately for layers and for different ranks.

[0279] In one embodiment of the present disclosure, the artificial intelligence model is trained commonly for layers and separately for different ranks.

[0280] In one embodiment of the present disclosure, the control circuit inputs the intermediate output from the previous time to the artificial intelligence model when the first rank value and the second rank value are the same or when the first rank value is smaller than the second rank value.

[0281] In one embodiment of the present disclosure, when the first rank value N is greater than the second rank value M, the control circuit inputs the intermediate output from the previous time to the artificial intelligence model corresponding to layer 1 to layer M, and discards the intermediate output or inputs a value common between the terminal and the base station as the intermediate output to the artificial intelligence model corresponding to layer (M+1) to layer N.

[0282] In one embodiment of the present disclosure, when the first rank value N is greater than the second rank value M, the control circuit inputs the intermediate output from the previous time to the artificial intelligence models corresponding to Layer 1 to Layer M, and inputs the intermediate output from the past time at which the first rank value was set to the artificial intelligence models corresponding to Layer (M+1) to Layer N.

[0283] In one embodiment of the present disclosure, when the first rank value N is greater than the second rank value M, the control circuit inputs the intermediate output from the previous time to the artificial intelligence model corresponding to layer 1 to layer M, and sets the intermediate output for the artificial intelligence model corresponding to layer (M+1) to layer N depending on whether the intermediate output from the past time when the first rank value was set is valid.

[0284] In one embodiment of the present disclosure, if the intermediate output is valid, the control circuit inputs the intermediate output to the artificial intelligence model corresponding to layer (M+1) to layer N, and if the intermediate output is not valid, the control circuit discards the intermediate output or inputs a value common between the terminal and the base station as the intermediate output to the artificial intelligence model corresponding to layer (M+1) to layer N.

[0285] In one embodiment of the present disclosure, the artificial intelligence model is set for each layer, updates of the intermediate outputs input to the artificial intelligence model are performed independently between layers, and the control circuit uses the artificial intelligence model to perform inference processing individual to the layer.

[0286] In one embodiment of the present disclosure, the artificial intelligence model is trained separately for each layer, and the artificial intelligence model corresponding to a certain layer is commonly used for different ranks.

[0287] In one embodiment of the present disclosure, the artificial intelligence model is commonly trained for layers, and the artificial intelligence model corresponding to a certain layer is commonly used for different ranks.

[0288] A base station according to one embodiment of the present disclosure includes a receiving circuit that receives a channel state information report generated by a terminal-side artificial intelligence model, and a control circuit that controls input of an intermediate output from a network-side artificial intelligence model at a past time to a network-side artificial intelligence model that reconstructs the channel state information report based on a first rank value at the current time and a second rank value at the previous time.

[0289] In a communication method according to one embodiment of the present disclosure, a terminal controls the input of an intermediate output from an artificial intelligence model on the terminal side at a past time to an artificial intelligence model on the terminal side that generates a channel state information report based on a first rank value at the current time and a second rank value at the previous time, and transmits the channel state information report generated using the intermediate output.

[0290] In a communication method according to one embodiment of the present disclosure, a base station receives a channel state information report generated by a terminal-side artificial intelligence model, and controls the input of intermediate output from the network-side artificial intelligence model at a past time to a network-side artificial intelligence model that reconstructs the channel state information report based on a first rank value at the current time and a second rank value at the previous time.

[0291] The disclosures of the specification, drawings and abstract contained in Japanese Patent Application No. 2024-114902, filed on July 18, 2024, are incorporated herein by reference in their entirety.

[0292] One embodiment of the present disclosure is useful in wireless communication systems.

[0293] 100 Base station 101, 205 Control unit 102, 206 Signal generation unit 103, 207 Transmission unit 104, 201 Reception unit 105, 202 Extraction unit 106, 203 Demodulation unit 107, 204 Decoding unit 200 Terminal

Claims

1. A terminal comprising: a control circuit that controls input of an intermediate output from an artificial intelligence model on the terminal side at a past time to an artificial intelligence model on the terminal side that generates a channel state information report based on a first rank value at the current time and a second rank value at the immediately previous time; and a transmission circuit that transmits the channel state information report generated using the intermediate output.

2. The terminal according to claim 1, wherein the control circuit inputs the intermediate output from the immediately previous time to the artificial intelligence model when the first rank value and the second rank value are the same.

3. The terminal according to claim 2, wherein the control circuit, when the first rank value and the second rank value are different, discards the intermediate output or inputs a value common between the terminal and the base station as the intermediate output to the artificial intelligence model.

4. The terminal according to claim 2, wherein, when the first rank value and the second rank value are different, the control circuit inputs the intermediate output at the past time when the first rank value was set to the artificial intelligence model.

5. The terminal according to claim 2, wherein, when the first rank value and the second rank value are different, the control circuit sets the intermediate output depending on whether the intermediate output at the past time when the first rank value was set was valid or not.

6. The terminal according to claim 5, wherein the control circuit inputs the intermediate output to the artificial intelligence model if the intermediate output is valid, and discards the intermediate output if the intermediate output is not valid, or inputs a value common between the terminal and a base station as the intermediate output to the artificial intelligence model.

7. The terminal described in claim 2, wherein the artificial intelligence model is set for each rank, the intermediate output input to the artificial intelligence model is updated independently between ranks, and the control circuit performs inference processing individual to the rank using the artificial intelligence model corresponding to the selected rank.

8. The terminal according to claim 2, wherein the artificial intelligence model is set for each layer, the intermediate output input to the artificial intelligence model is updated independently between layers, and the control circuit uses the artificial intelligence model to perform inference processing individually for each layer.

9. The terminal according to claim 8, wherein the artificial intelligence model is trained separately for layers and for different ranks.

10. The terminal according to claim 8, wherein the artificial intelligence model is trained commonly for layers and separately for different ranks.

11. The terminal according to claim 1, wherein the control circuit inputs the intermediate output from the immediately previous time to the artificial intelligence model when the first rank value and the second rank value are the same or when the first rank value is smaller than the second rank value.

12. The terminal according to claim 11, wherein, when the first rank value N is greater than the second rank value M, the control circuit inputs the intermediate output of the previous time to the artificial intelligence models corresponding to Layer 1 to Layer M, and discards the intermediate output or inputs a value common between the terminal and a base station as the intermediate output to the artificial intelligence models corresponding to Layer (M+1) to Layer N.

13. The terminal according to claim 11, wherein, when the first rank value N is greater than the second rank value M, the control circuit inputs the intermediate output at the previous time to the artificial intelligence models corresponding to Layer 1 to Layer M, and inputs the intermediate output at the past time at which the first rank value was set to the artificial intelligence models corresponding to Layer (M+1) to Layer N.

14. The terminal according to claim 11, wherein, when the first rank value N is greater than the second rank value M, the control circuit inputs the intermediate output at the previous time to the artificial intelligence models corresponding to Layer 1 to Layer M, and sets the intermediate output for the artificial intelligence models corresponding to Layer (M+1) to Layer N depending on whether the intermediate output at the past time when the first rank value was set is valid.

15. The terminal according to claim 14, wherein the control circuit, if the intermediate output is valid, inputs the intermediate output to the artificial intelligence model corresponding to layer (M+1) to layer N, and, if the intermediate output is not valid, discards the intermediate output or inputs a value common between the terminal and a base station as the intermediate output to the artificial intelligence model corresponding to layer (M+1) to layer N.

16. The terminal according to claim 11, wherein the artificial intelligence model is set for each layer, the intermediate output input to the artificial intelligence model is updated independently between layers, and the control circuit uses the artificial intelligence model to perform inference processing individually for each layer.

17. The terminal according to claim 16, wherein the artificial intelligence model is trained individually for each layer, and the artificial intelligence model corresponding to a certain layer is commonly used for different ranks.

18. The terminal according to claim 16, wherein the artificial intelligence model is trained in common for a layer, and the artificial intelligence model corresponding to a certain layer is used in common for different ranks.

19. A base station comprising: a receiving circuit that receives a channel state information report generated by a terminal-side artificial intelligence model; and a control circuit that controls input of an intermediate output from the network-side artificial intelligence model at a past time to a network-side artificial intelligence model that reconstructs the channel state information report based on a first rank value at the current time and a second rank value at the immediately previous time.

20. A communication method in which a terminal controls the input of an intermediate output from an artificial intelligence model on the terminal side at a past time to an artificial intelligence model on the terminal side that generates a report of channel state information based on a first rank value at the current time and a second rank value at the immediately previous time, and transmits the report of channel state information generated using the intermediate output.

21. A communications method in which a base station receives a report of channel state information generated by an artificial intelligence model on a terminal side, and controls input of intermediate output from the artificial intelligence model on the network side at a past time to an artificial intelligence model on a network side that reconstructs the report of channel state information based on a first rank value at the current time and a second rank value at the immediately previous time.

Images