Csi feedback with ai-csi and conventional csi

Abstract

An apparatus configured to process, based on signaling received from a base station, a Channel State Information (CSI) configuration comprising CSI Reference Signal (CSI-RS) resources to be measured, measure the CSI-RS resources and generate CSI feedback based on measurement results for the CSI-RS resources, wherein the CSI feedback comprises CSI part 1 and CSI part 2 feedback encoded as a latent space representation.

Description

Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1 CSI Feedback with AI-CSI and Conventional CSI Inventors: Weidong Yang, Chunxuan Ye, Dawei Zhang, Haitong Sun, Huaning Niu, Oghenekome Oteri and Wei Zeng Priority / Incorporation By Reference

[0001] This application claims priority to U.S. Provisional Application Serial No. 63 / 605,186 filed on December 1, 2023, entitled “CSI Feedback with AI-CSI and Conventional CSI,” the entirety of which is incorporated by reference herein. Background

[0002] User equipment (UE) may transmit feedback information to a base station for various purposes including indicating the quality of a channel between the UE and the base station. For example, a UE operating in a Fifth Generation (5G) New Radio (NR) network may report channel state information (CSI) feedback to the network. There is an interest in determining whether artificial intelligence (AI) may be used on either the UE side or the network side to improve CSI feedback. Summary

[0003] Some example embodiments are related to an apparatus having processing circuitry coupled to a memory, the processing circuitry configured to process, based on signaling received from a base station, a Channel State Information (CSI) configuration comprising CSI Reference Signal (CSI-RS) resources to be measured, measure the CSI-RS resources and generate CSI feedback based on measurement results for the CSI-RS resources, wherein the CSI feedback comprises CSI part 1 and CSI part 2 feedback encoded as a latent space representation.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0004] Other example embodiments are related to an apparatus having processing circuitry coupled to a memory, the processing circuitry configured to process, based on signaling received from a user equipment (UE), Channel State Information (CSI) feedback comprising CSI encoded as a latent space representation and process the CSI feedback using an artificial intelligence (AI) / machine learning (ML) model to reconstruct a precoder used to encode the CSI feedback.

[0005] Still further example embodiments are related to a method for processing, based on signaling received from a base station, a Channel State Information (CSI) configuration comprising CSI Reference Signal (CSI-RS) resources to be measured, measuring the CSI-RS resources and generating CSI feedback based on measurement results for the CSI-RS resources, wherein the CSI feedback comprises CSI part 1 and CSI part 2 feedback encoded as a latent space representation, wherein the CSI feedback is generated using CSI eType II processing using an artificial intelligence (AI) / machine learning (ML) model. Brief Description of the Drawings

[0006] Fig. 1 shows an example network arrangement according to various example embodiments.

[0007] Fig. 2 shows an example UE according to various example embodiments.

[0008] Fig. 3 shows an example base station according to various example embodiments.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0009] Fig. 4 shows a first example embodiment of a CSI feedback scheme using an auto-encoder according to various example embodiments.

[0010] Fig. 5 shows a second example embodiment of a CSI feedback scheme using CSI eType II according to various example embodiments.

[0011] Fig. 6 shows a third example embodiment of a CSI feedback scheme using a combination of a generalized auto- encoder and a conventional CSI processing such as a CSI (e)Type II codebook according to various example embodiments.

[0012] Fig. 7 shows a fourth example embodiment of a CSI feedback scheme using a generalized auto-encoder with CSI eType II and AI / ML processing according to various example embodiments.

[0013] Fig. 8 shows a fifth example embodiment of a CSI feedback scheme using a generalized auto-encoder with conventional CSI processing such as based on a CSI eType II codebook and AI / ML according to various example embodiments.

[0014] Fig. 9 shows an example method for reporting CSI feedback according to various example embodiments. Detailed Description

[0015] The example embodiments may be further understood with reference to the following description and the related appended drawings, wherein like elements are provided with the sameAttorney Docket No. 30134 / 89802 Ref. No. P63913WO1 reference numerals. The example embodiments relate to a user equipment (UE) providing CSI feedback to a network.

[0016] The example embodiments are described with regard to a user equipment (UE). However, reference to a UE is merely provided for illustrative purposes. The example embodiments may be utilized with any electronic component that may establish a connection to a network and is configured with the hardware, software, and / or firmware to exchange information and data with the network. Therefore, the UE as described herein is used to represent any electronic component.

[0017] The example embodiments are also described with reference to a 5G New Radio (NR) network. However, the example embodiments may also be implemented in other types of networks, including but not limited to legacy cellular networks (e.g., Long Term Evolution (LTE)), future evolutions of the cellular protocol (e.g., 5G advanced, 6G, etc.), or any other type of network.

[0018] The example embodiments are described with reference to implementing artificial intelligence (AI) / machine learning (ML) models on the UE side and / or the network side (e.g., base station) for encoding and / or decoding CSI feedback to the network. Before an AI / ML model is deployed at a UE or base station, AI / ML model training, testing, etc., may be conducted at the UE, at a UE side server, a network side server, a Test Equipment (TE) vendor server, a third party server such as a cloud based server, etc.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0019] Throughout this description, it may be described that certain operations are performed by one or more AI / ML models or. There are many different types of machine learning models. For example, AI / ML models may include classifiers, regression models, etc. Other examples of machine learning models may include multitask learning models (MTL) that can perform both classification, regression and other tasks. The AI / ML models may be based on the use of one or more of: a non-linear hierarchical algorithm, a neural network, a convolutional neural network, a recurrent neural network, a long short-term memory network, a multi-dimensional convolutional network, a memory network, a transformer network, a fully convolutional network, a gated recurrent network, gradient boosting techniques, random forest techniques, etc. The resulting AI / ML models described below may include some or all of the above machine learning components or any other type of machine learning model that may be applied to determine the expected outcome of the encoding / decoding process.

[0020] The example embodiments provide various manners for a UE to provide CSI feedback to a network. The example embodiments include various encoding and decoding techniques including legacy like encoding and corresponding artificial intelligence (AI) / machine learning (ML) decoding techniques, encoding using legacy like techniques and AI / ML encoding and corresponding AI / ML decoding and encoding using legacy like techniques with per CSI component AI / ML encoding and corresponding per CSI component AI / ML decoding. The example embodiments also describe various manners of training the example AI / ML models described herein. These and other example embodiments are described in greater detail below.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0021] Fig. 1 shows an example network arrangement 100 according to various example embodiments. The example network arrangement 100 includes a UE 110. The UE 110 may be any type of electronic component that is configured to communicate via a network, e.g., mobile phones, tablet computers, desktop computers, smartphones, embedded devices, wearables, Internet of Things (IoT) devices, etc. An actual network arrangement may include any number of UEs being used by any number of users. Thus, the example of one UE 110 is merely provided for illustrative purposes.

[0022] The UE 110 may be configured to communicate with one or more networks. In the example of the network arrangement 100, the network with which the UE 110 may wirelessly communicate is a 5G NR radio access network (RAN) 120. The UE 110 may also communicate with other types of networks (e.g., 5G cloud RAN, a next generation RAN (NG-RAN), a legacy cellular network, etc.) and the UE 110 may also communicate with networks over a wired connection. With regard to the example embodiments, the UE 110 may establish a connection with the 5G NR RAN 120. Therefore, the UE 110 may have a 5G NR chipset to communicate with the NR RAN 120.

[0023] The 5G NR RAN 120 may be portions of a cellular network that may be deployed by a network carrier (e.g., Verizon, AT&T, T-Mobile, etc.). The RAN 120 may include cells or base stations that are configured to send and receive traffic from UEs that are equipped with the appropriate cellular chip set. In this example, the 5G NR RAN 120 includes the gNB 120A and the gNB 120B. However, reference to a gNB is merely provided for illustrative purposes, any appropriate base station or cellAttorney Docket No. 30134 / 89802 Ref. No. P63913WO1 may be deployed (e.g., Node Bs, eNodeBs, HeNBs, eNBs, gNBs, gNodeBs, macrocells, microcells, small cells, femtocells, etc.).

[0024] Any association procedure may be performed for the UE 110 to connect to the 5G NR RAN 120. For example, as discussed above, the 5G NR RAN 120 may be associated with a particular network carrier where the UE 110 and / or the user thereof has a contract and credential information (e.g., stored on a SIM card). Upon detecting the presence of the 5G NR RAN 120, the UE 110 may transmit the corresponding credential information to associate with the 5G NR RAN 120. More specifically, the UE 110 may associate with a specific cell (e.g., gNB 120A).

[0025] The network arrangement 100 also includes a cellular core network 130, the Internet 140, an IP Multimedia Subsystem (IMS) 150, and a network services backbone 160. The cellular core network 130 manages the traffic that flows between the cellular network and the Internet 140. The IMS 150 may be generally described as an architecture for delivering multimedia services to the UE 110 using the IP protocol. The IMS 150 may communicate with the cellular core network 130 and the Internet 140 to provide the multimedia services to the UE 110. The network services backbone 160 is in communication either directly or indirectly with the Internet 140 and the cellular core network 130. The network services backbone 160 may be generally described as a set of components (e.g., servers, network storage arrangements, etc.) that implement a suite of services that may be used to extend the functionalities of the UE 110 in communication with the various networks.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0026] Fig. 2 shows an example UE 110 according to various example embodiments. The UE 110 will be described with regard to the network arrangement 100 of Fig. 1. The UE 110 may represent any electronic device and may include a processor 205, a memory arrangement 210, a display device 215, an input / output (I / O) device 220, a transceiver 225, and other components 230. The other components 230 may include, for example, an audio input device, an audio output device, a battery that provides a limited power supply, a data acquisition device, ports to electrically connect the UE 110 to other electronic devices, sensors to detect conditions of the UE 110, etc.

[0027] The processor 205 may be configured to execute a plurality of engines for the UE 110. For example, the engines may include a CSI feedback engine 235 for performing operations related to reporting CSI feedback to a network. The operations include, but are not limited to, encoding CSI feedback into a latent space representation using legacy like encoding, encoding CSI feedback into a latent space representation using legacy like encoding combined with AI / ML model encoding, encoding CSI feedback using AI / ML model encoding on a per CSI component basis and training the AI / ML models, generating CSI feedback arrangement information to aid the network to perform AI / ML model decoding and / or assemble the reconstructed precoder from decoding / parsing on a CSI component basis. Each of these example operations will be described in more detail below.

[0028] The above referenced engine being an application (e.g., a program) executed by the processor 205 is only an example. The functionality associated with the engines may also be represented as a separate incorporated component of the UEAttorney Docket No. 30134 / 89802 Ref. No. P63913WO1 110 or may be a modular component coupled to the UE 110, e.g., an integrated circuit with or without firmware. For example, the integrated circuit may include input circuitry to receive signals and processing circuitry to process the signals and other information. The engines may also be embodied as one application or separate applications. In addition, in some UEs, the functionality described for the processor 205 is split among two or more processors such as a baseband processor and an applications processor. The example embodiments may be implemented in any of these or other configurations of a UE.

[0029] The memory arrangement 210 may be a hardware component configured to store data related to operations performed by the UE 110. The display device 215 may be a hardware component configured to show data to a user while the I / O device 220 may be a hardware component that enables the user to enter inputs. The display device 215 and the I / O device 220 may be separate components or integrated together such as a touchscreen.

[0030] The transceiver 225 may be a hardware component configured to establish a connection with the 5G NR-RAN 120, an LTE-RAN (not pictured), a legacy RAN (not pictured), a WLAN (not pictured), etc. Accordingly, the transceiver 225 may operate on a variety of different frequencies or channels (e.g., set of consecutive frequencies). The transceiver 225 includes circuitry configured to transmit and / or receive signals (e.g., control signals, data signals). Such signals may be encoded with information implementing any one of the methods described herein. The processor 205 may be operably coupled to the transceiver 225 and configured to receive from and / or transmit signals to the transceiver 225. The processor 205 may beAttorney Docket No. 30134 / 89802 Ref. No. P63913WO1 configured to encode, decode and / or process signals (e.g., signaling from a base station of a network) for implementing any one of the methods described herein.

[0031] Fig. 3 shows an example base station 300 according to various example embodiments. The base station 300 may represent the gNB 120A, the gNB 120B or any other access node through which the UE 110 may establish a connection and manage network operations.

[0032] The base station 300 may include a processor 305, a memory arrangement 310, an input / output (I / O) device 315, a transceiver 320, and other components 325. The other components 325 may include, for example, an audio input device, an audio output device, a battery, a data acquisition device, ports to electrically connect the base station 300 to other electronic devices and / or power sources, etc.

[0033] The processor 305 may be configured to execute a plurality of engines for the base station 300. For example, the engines may include a CSI feedback engine 330 for performing operations related to a UE reporting CSI feedback to the base station 300. The operations include, but are not limited to, The operations include, but are not limited to, decoding CSI feedback from a latent space representation using an AI / ML model to reconstruct a precoder used for the CSI, decoding CSI feedback using AI / ML models on a per CSI component basis and training the AI / ML models, parsing CSI feedback arrangement information to perform AI / ML model decoding and / or assembling the reconstructed precoder from decoding / parsing on a CSIAttorney Docket No. 30134 / 89802 Ref. No. P63913WO1 component basis. Each of these example operations will be described in more detail below.

[0034] The memory 310 may be a hardware component configured to store data related to operations performed by the base station 300. The I / O device 315 may be a hardware component or ports that enable a user to interact with the base station 300. The transceiver 320 may be a hardware component configured to exchange data with the UE 110 and any other UE in the network arrangement 100.

[0035] The transceiver 320 may be a hardware component configured to exchange data with the UE 110 and any other UE in the network arrangement 100. The transceiver 320 may operate on a variety of different frequencies or channels (e.g., set of consecutive frequencies). The transceiver 320 includes circuitry configured to transmit and / or receive signals (e.g., control signals, data signals). Such signals may be encoded with information implementing any one of the methods described herein. The processor 305 may be operably coupled to the transceiver 320 and configured to receive from and / or transmit signals to the transceiver 320. The processor 305 may be configured to encode, decode and / or process signals (e.g., signaling from a UE) for implementing any one of the methods described herein.

[0036] There are various types of CSI feedback. For example, one type of CSI feedback is based on a CSI Type I codebook where spatial beams and coarsely quantized combination coefficients are used. In another example, another type of CSI feedback is based on a CSI Type II codebook where quantized spatial beams,Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1 non-zero coefficient locations, and quantized linear combination coefficients are used. The set of linear combination coefficients for Type II is much larger than in Type I and the time domain formulation of Type II reduces overhead by exploiting the time-domain correlation.

[0037] In further examples, a type of CSI feedback may be based on a Rel-15 Type II portion selection codebook, a NR Rel- 16 enhanced CSI Type II codebook that exploits frequency domain correlation among precoders at subbands, a NR Rel-16 enhanced Type II portion selection codebook, a NR Rel-17 further enhanced Type II portion selection codebook, a NR Rel-18 enhanced Type II codebook for Coherent Joint Transmission (CJT), a NR Rel-18 enhanced Type II port selection codebook for CJT, a NR Rel-18 enhanced Type II codebook for predicted Pre-coding Matrix Indicator (PMI), a NR Rel-18 enhanced Type II portion selection codebook for predicted PMI, a combination of NR Rel-18 enhanced Type II (port selection) codebook for predicted PMI and Rel-18 enhanced Type II (port selection) codebook for CJT, etc.

[0038] For these types of CSI feedback an auto-encoder may be used at a frame / image level (e.g., over a 32 x 13 matrix), where E() is the encoder, D() is the decoder, P is the input precoding matrix, L is the latency space representation, and P’ is the recovered precoding matrix.

[0039] Another type of CSI feedback may be a Rel-16 CSI eType II codebook. In this example, the precoding matrix may have a high dimension (e.g., 32 x 13) resulting in a quantizer having many different constellations even with quantization if no structure is enforced on the precoding matrix (eachAttorney Docket No. 30134 / 89802 Ref. No. P63913WO1 constellation for example can be a 32 x 13 matrix). However, for eType II, the channel information may be exploited by constraining the spatial beam selection, the tap selection (which is also called Frequency Domain (FD) component selection in some literature, and the number of non-zero coefficients because not all of the elements in the precoding matrix may be used. Thus, the constellation number may be roughly calculated (assuming the number of non-zero coefficients are equally split among two polarizations) as follows: ^^0 / 2 ^^^^ ^^^^ 2^^−1^^^^−−1^^^^^^^−^^1^^^^^^^^^ ∙ 2 ∙ ^^ ^^

[0040] The for spatial beamselection, the second combination for tap selection, the third combination for NZ coefficients in polarization with the strongest coefficient and the fourth combination for the NZ coefficients in the other polarization. The 2 accounts for the strongest coefficient that may be located in the first polarization (e.g., 45º) or the second polarization (e.g., −45º). A is the number of constellations in the amplitude / phase quantization (e.g., 4-bits for phase and 3 bits for amplitude ==> A=128), and B is the reference amplitude quantization for the “weaker” polarization (B = 16).

[0041] For all Type I / Type II codebooks, including enhancements over Rel-16 to Rel-18, the procedure to reconstruct or recover a precoder at the network side from the UE CSI feedback is detailed in the 3GPP specifications (e.g., Technical Specification (TS) 38.214, TS 38.331, TS 38.321, etc.). For example, consider two base stations that may be provided by different network equipment vendors but have identical configurations for CSI feedback (e.g., both are configured withAttorney Docket No. 30134 / 89802 Ref. No. P63913WO1 Rel-16 enhanced Type II codebook with identical configurations for their operation). These two base stations provide these CSI feedback configurations to their respective UEs. If the respective UEs feedback identical CSI reports to the two base stations, it would be expected that exactly the same precoder is reconstructed or recovered from the CSI feedback by both base stations. Of course, modification to the reconstructed precoder is possible or can be motivated, e.g., out of implementation considerations or need for multi-user massive input / massive output (MU-MIMO) pairing.

[0042] In mathematics, a fixed-point theorem asserts that given certain conditions on a function F, F is guaranteed to possess at least one fixed point (a point x where F(x) = x). Quantization is an essential aspect of CSI feedback. Taking eType II as an example, CSI feedback consists of spatial beam quantization (q_1,q_2 ) selection and L spatial beam selection, tap quantization including Tap (FD component) selection, NZ coefficient selection (bitmap selection) and coefficient quantization.

[0043] For eType II CSI feedback, a normalization step may be performed for the frequency domain precoders 2L x N3 so the precoders are phase-rotated according to the strongest spatial beam (e.g., the top one out of 2L spatial beams). This concentrates power on a single tap to result in a more parsimonious time-domain representation.

[0044] For example, let the Ntx x N3 frequency domain precoder be P, where Ntx= 2N1N2, N1 / N2are for the antenna configurationsAttorney Docket No. 30134 / 89802 Ref. No. P63913WO1 in the cross-pool antenna array. N3 is the number of subbands for PMI feedback. The precoding process may be represented by: ^^(^^) = ^^ (^^ (IDFT2(^^(^^ ^^ ⋅ ^^)))),

[0045] where B = N(⋅) is for the normalization step, IDFT2(⋅) converts the frequency domain precoding matrix into a time-domain representation and the subscript 2 indicates the IDFT is performed along the 2nd dimension of the matrix, Q(⋅)is for the amplitude and phase quantization, T(⋅)is for FD component selection and NZ coefficient selection, which effectively sets taps to zeroes if they are either not selected by FD component selection or not selected by NZ coefficient selection even if covered by FD component selection. Similar to an autoencoder, the UE side processing by the encoder function is denoted as E(⋅).

[0046] This formulation is one out of many possibilities, e.g., the projection to spatial beams may be handled through operations over the subband covariance matrices, the order of quantization and FD component selection and NZ coefficient selection can be swapped or changed, etc. The example processing procedure is illustrative, and conducive to the comparison with that for an auto-encoder.

[0047] On the base station side, the frequency domain precoding matrix may be generated by: ^^ (^^) = ^^ ⋅ DFT (^^) = ^^ ⋅ (^^ ^^ ⋅ ^^) ),Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0048] ,where DFT2(⋅) is for the DFT performed along the 2nd dimension of the matrix, after that the W2's for N3 subbands are generated, and multiplication by B over the resultant W2's generates W1W2, i.e., the frequency domain precoder. Similar to the autoencoder formulation, the network side processing is denoted as a decoder D(⋅).

[0049] We can define the CSI processing as ^^ = ^^(^^), where^^(^^) = ^^(^^(^^)), and x is a complex matrix at Ntx x N3, and W is aat Ntxx N3. The processing steps may depend on P,e.g., FD component selection can be data dependent. In contrast, for an autoencoder, each processing step may be data independent.

[0050] It can be verified that with the normalization operation N(⋅), every output of eType II processing is a fixedpoint of the CSI processing, i.e., ^^(^^(^^)) = ^^(^^), for any ^^ ∈^^^^^^^^×^^3. Due to the use of thewe also have^^(^^) = ^^(^^^^), where ^^ is any complex number. In general, we have:^^1)

[0051] , where ^^^^amplitude complex number.

[0052] Then the whole space ^^^^^^^^×^^3or a cube in the wholespace, or a sphere ∥ ^^ ∥< 1,  ^^ ∈ ^^^^^^^^×^^3 are divided / partitioned intoregions with each region mapping to a fixed point or a constellation. It can be seen that the eType II CSI feedback with the spatial beam selection, FD component selection,Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1 coefficient selection, the strongest coefficient selection, etc., and the NZ coefficient quantization, provides a label to a fixed point. Formulated in another way, the eType II CSI feedback can be characterized by partitioned regions and their corresponding labels.

[0053] The connection between eType II CSI feedback and auto- encoder based CSI feedback starts to emerge. Rather than using a neural network to generate labels or a latent space representation, a CSI processing procedure as captured by: ^^(^^) = ^^(^^(^^)) = ^^ ⋅ DFT2(^^ (^^ (IDFT2(^^(^^ ^^ ⋅ ^^))))

[0054] E(x) is used to generate a label; and = ^^) is afixed point of the function ^^(^^).

[0055] For other types of MIMO codebooks, e.g., from Rel-15 to Rel-18 as described previously, or a MIMO codebook supporting both CJT and predicted PMI, similar analysis may be used and fixed points with the corresponding encoding / decoding procedures may be demonstrated.

[0056] Let DA and EA be the decoder and encoder, respectively, for the autoencoder in AI / ML for CSI feedback (the subscript "A" is for "Autoencoder"), which may subsume preprocessing and / or postprocessing steps. Let ^^^^be the operation of normalization on frequency domain precoders according to the strongest antenna port or the strongest spatial beam (out of Ntx) if a transformed domain is used. Then, let ^^^^(^^) = ^^^^ (^^^^(^^^^(^^)))Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0057] ^^^^(^^) captures both the UE side processing that generates a latent space representation from an AI / ML encoder and network side processing that reconstructs a precoder from the latent space representation generated by the UE with an AI / ML decoder. ^^^^(^^^^(^^)) generates a label in response to x.Ideally, ^^^^(^^ = be a fixed point of the function ^^^^(^^).

[0058] With that, the analog between conventional CSI feedback such as Rel-16 eType-II and AI / ML autoencoder iscomplete: On the sphere of ∥ ^^ ∥≤ 1,  ^^ ∈ ^^^^^^^^×^^3, regions are defined,and each region is mapped to a fixed point which is associated with a latent space representation or label. The similarity in formulation for eType II CSI processing above can be observed.

[0059] If the UE side processing to generate CSI feedback, which can be captured by a general encoder can be treated as a black box, then in terms of functionality in generating latent space representations or labels, which can be accomplished by eType II processing, or eType II like processing, or neural network processing (CNN-encoder or transformer-encoder), no difference can be seen externally outside the black box. Thus, if the network designs a decoder to recover the target CSI from the latent space representation or label, via a procedure which can follow eType II processing, eType II like processing, or neural network processing (fully-connected network-decoder, CNN- decoder or transformer-decoder). Then, for Type 3 training, a first company may provide a training dataset with input CSI and their corresponding labels, which may be generated by eType II CSI feedback (e.g., from Rel-16, Rel-17, Rel-18, etc.), or any variation. This may be used to generate an AI / ML decoder such as transformer decoder with which the target CSI can be recovered.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0060] For eType II or eType II CSI like feedback, instead of a single label for an entry for input CSI (input precoder), there can be a pair, a triplet and in general a set of labels. Each label may correspond to a single block for encoding. In one example, putting it in a more concrete setup, the signaling other than coefficient quantization can be aggregated into sub- label 1, and coefficient quantization can be aggregated into sub-label 2. Compared to the case of using a single label (e.g., the label is a sequence of 60 bits) resulting from concatenation of those two sub-labels (e.g., sub-label-1 is a sequence of 20 bits, sub-label-2 is a sequence of 40 bits) with which a single AI network is required to recover the target CSI from the single label (60 bits for example) if the concatenation operation at the UE side is unknown at the network side; with sub-labels, then two smaller AI / ML models at the network side may be trained to recovered parts of the CSI feedback separately if the network is provided information concerning the composition of sub-labels (e.g., from a received sequence of 60 bits, the network is aware the first 20 bits are for sub-label 1, and the later 40 bits are for sub-label 2). In this case, the composition from those parts to derive the target CSI is disclosed to the network side

[0061] From this it can be seen that the quantized precoding matrices are “fixed points” in the precoding matrix space with respect to the eType-II procedure: f(f(P)) = f(P)

[0062] The CSI eType-II may include feedback that comprises CSI part 1 and CSI part 2 feedback.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0063] There may also be CSI feedback based on an input precoder. An input precoder may include a beamforming matrix. The example embodiments may consider using the input precoder and providing CSI feedback based on the input precoder.

[0064] Fig. 4 shows a first example embodiment of a CSI feedback scheme 400 using an auto-encoder according to various example embodiments. In this example embodiment, the encoder 410 may be implemented on the UE 110 and the decoder 420 may be implemented on the base station 300 (e.g., gNB 120A).

[0065] The encoder 410 may use various AI / machine learning (ML) techniques to encode an input precoder such as a transformer (implementing a trained model), a convolution neural network (CNN), etc. As shown in Fig. 4, the encoder 410 may use the AI / ML techniques to encode the input precoder into a latent space representation that is the CSI feedback sent to the base station 300.

[0066] The base station 300 implements a decoder 420 that has a corresponding transformer, CNN, etc. that decodes the latent space representation back into the reconstructed precoder. An issue with the auto-encoder of the CSI feedback scheme 400 is that the transformer or CNN may implement a large and complicated model to generate accurate results for both the encoding and decoding. Thus, the UE 110 may dedicate a large amount of storage and processing power to implement the encoder 410.

[0067] Fig. 5 shows a second example embodiment of a CSI feedback scheme 500 using CSI eType II according to variousAttorney Docket No. 30134 / 89802 Ref. No. P63913WO1 example embodiments. In this example embodiment, the encoder 510 may be implemented on the UE 110 and the decoder 520 may be implemented on the base station 300 (e.g., gNB 120A).

[0068] The encoder 510 may use CSI eType II processing to encode an input precoder. As shown in Fig. 5, the encoder 510 may encode the input precoder into the CSI part 1 and CSI part 2 feedback that is the CSI feedback sent to the base station 300.

[0069] The base station 300 implements a decoder 520 that implements decoding and parsing operations such as those defined in 3GPP Technical Specification 38.214 to reconstruct the precoder. That is, the decoder 520 has access to the CSI part 1 and CSI part 2 feedback and may use this to reconstruct the precoder. An issue with the CSI eType II encoding and decoding of Fig. 5 is that it does not implement any AI / ML techniques that may improve the CSI feedback performance.

[0070] Fig. 6 shows a third example embodiment of a CSI feedback scheme 600 using a combination of a generalized auto- encoder and a conventional CSI processing such as a CSI (e)Type II codebook according to various example embodiments. In this example embodiment, the encoder 610 may be implemented on the UE 110 and the decoder 620 may be implemented on the base station 300 (e.g., gNB 120A).

[0071] In the example of Fig. 6, the encoder 610 performs a combination of the encoding techniques described above. For example, the encoder 610 processes the input precoder using the CSI eType II processing techniques described above. However, the CSI part 1 and CSI part 2 feedback is represented as a latentAttorney Docket No. 30134 / 89802 Ref. No. P63913WO1 space representation that is the CSI feedback to the base station 300. In this example, the CSI part 1 and CSI part 2 that is represented in the latent space representation may not have the underlying structure of the CSI part 1 and CSI part 2, e.g., the latent space representation is a bit sequence without any underlying meaning.

[0072] The base station 300 may implement a decoder 620 that uses AI / ML techniques to decode the latent space representation back into the reconstructed precoder. Thus, because the UE 110 does not need to implement the AI / ML model in the CSI feedback scheme 600, this solves the issue related to the CSI feedback scheme 400 where the UE 110 may dedicate a large amount of storage and processing power to implement the encoder 410. On the other hand, this also solves the issue with the CSI feedback scheme 500 where the improved performance of AI / ML is not implemented.

[0073] The following provides an example of how the AI / ML model that is implemented by the decoder 620 may be trained and how the model may then perform the inference in operation. A training data set may be generated based on one or more UEs processing the frequency domain precoders according to conventional / legacy CSI processing using one of the MIMO codebooks, or CSI eType II-like processing (e.g., variations from legacy codebooks from Rel-15, Rel-16, Rel-17, Rel-18), e.g., selecting the top K coefficients and generating signaling overhead for the CSI feedback. This signaling overhead may be generated according to combinatorial indexing, a strongest coefficient indication, and coefficient quantization, which may result, for example, in 63 bits for layer 1 feedback.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0074] For example, the UEs may generate the CSI part 1 and CSI part 2 feedback (F) for the CSI eType II as described above for an input precoder (v). Each of the feedback (F) may then be decoded using the CSI eType II decoding and parsing techniques described above for the decoder 520 to generate a reconstructed precoder Q(F). This may be performed for many samples of v’s from UEs to generate a large training data set.

[0075] The training data set may then be used to train the AI / ML model using a layer common / rank common AI / ML decoder. For example, in the training operation, the encoder may perform the CSI eType II processing to generate the CSI part 1 and CSI part 2 feedback but then represent the feedback as the latent space representation. This feedback may then be fed into the AI / ML model of the decoder to generate the reconstructed precoder across the large training data set to train the AI / ML model.

[0076] In some example embodiments, the decoder may be a rank specific AI / ML decoder, rank-common AI / ML decoder, layer specific and rank common AI / ML decoder, or layer specific and rank specific AI / ML decoder, layer common and rank specific AI / ML decoder rather than a layer common / rank common AI / ML decoder depending on the training that is implemented.

[0077] In operation, the trained AI / ML of the transformer- decoder may “remember” the mapping from the latent space representation to the quantized precoder. In these example embodiments, the decoder (e.g., decoder 620) is unaware that the encoder is not implementing a corresponding AI / ML model for encoding as the decoder is trained to reconstruct the precoderAttorney Docket No. 30134 / 89802 Ref. No. P63913WO1 based on the feedback regardless of whether that feedback was generated using an AI / ML model.

[0078] Fig. 7 shows a fourth example embodiment of a CSI feedback scheme 700 using a generalized auto-encoder with CSI eType II and AI / ML processing according to various example embodiments. In this example embodiment, the encoder 710 may be implemented on the UE 110 and the decoder 720 may be implemented on the base station 300 (e.g., gNB 120A).

[0079] In this example, the encoder 710 includes both CSI eType II like processing 715 and AI / ML processing 717. In this example, the encoder 710 processes the input precoder using the CSI eType II processing 715 techniques described above to generate the CSI part 1 and CSI part 2. In this example, the CSI eType II processing 715 may be considered to be pre-processing. The CSI part 1 and CSI part 2 may then be processed through an AI / ML model 717 to represent CSI part 1 and CSI part 2 in a latent space representation that is the CSI feedback to the base station 300.

[0080] The base station 300 may implement a decoder 720 that uses AI / ML techniques to decode the latent space representation back into the reconstructed precoder. Thus, in this example, because the AI / ML model implemented by the UE 110 processes the structured CSI part 1 and CSI part 2 feedback to generate the latent space representation, the AI / ML model may not be as large or complicated as the AI / ML model for directly generating the latent space representation from the input precoder.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0081] The training of the AI / ML models for both the encoder 710 and the decoder 720 may be similar to the training described above for the decoder 620. For example, the training data set may be generated based on one or more UEs processing the frequency domain precoders according to legacy CSI processing using one of the MIMO codebooks, or CSI eType II-like processing (e.g., variations from legacy codebooks from Rel-15, Rel-16, Rel-17, Rel-18), e.g., selecting the top K coefficients and generating signaling overhead for the CSI feedback. This signaling overhead may be generated according to combinatorial indexing, a strongest coefficient indication, and coefficient quantization, which may result, for example, in 63 bits for layer 1 feedback.

[0082] For example, the UEs may generate the CSI part 1 and CSI part 2 feedback (F) for the CSI eType II as described above for an input precoder (v). Each of the feedback (F) may then be decoded using the CSI eType II decoding and parsing techniques described above for the decoder 520 to generate a reconstructed precoder Q(F). This may be performed for many samples from UEs to generate a large training data set.

[0083] The training data set may then be used to train the AI / ML models using a layer common rank common AI / ML-decoder such as a transformer-decoder. For example, in the training operation, the encoder may perform the CSI eType II processing to generate the CSI part 1 and CSI part 2 but then represent the parts as the latent space representation using a UE-side AI / ML model. This feedback may then be fed into the AI / ML model of the decoder to generate the reconstructed precoder across the large training data set to train the AI / ML model.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0084] In some example embodiments, the decoder may be a rank specific AI / ML decoder, rank-common AI / ML decoder, layer specific and rank common AI / ML decoder, or layer specific and rank specific AI / ML decoder, layer common and rank specific AI / ML decoder rather than a layer common / rank common AI / ML- decoder depending on the training that is implemented.

[0085] Fig. 8 shows a fifth example embodiment of a CSI feedback scheme 800 using a generalized auto-encoder with conventional CSI processing such as based on a CSI eType II codebook and AI / ML according to various example embodiments. In this example embodiment, the encoder 810 may be implemented on the UE 110 and the decoders 820-840 may be implemented on the base station 300 (e.g., gNB 120A).

[0086] In this example, the encoder 810 includes both CSI eType II like processing 815 and AI / ML processing 817 and 818. In this example, the encoder 810 processes the input precoder using the CSI eType II processing 815 techniques described above to generate the CSI part 1 and CSI part 2. In this example, the CSI eType II processing 815 may be considered to be pre- processing. The CSI part 1 and CSI part 2 or at least one component of at least one CSI part may then be processed through different AI / ML models 817 and 818 to represent at least one component of CSI part 1 and CSI part 2 as a latent space representation. When multiple AI / ML models are used to process more than one component, this may result in more than one latent space representation. The latent space representation(s) and at least one component which does not go through AI / ML processing may be included in the CSI feedback to the base station 300.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0087] The different AI / ML models 817 and 818 may be related to different aspects or components of the CSI feedback. For example, the AI / ML model 817 may be related to the spatial beam selection and the AI / ML model 818 may be related to the frequency domain (FD) component selection. Other examples of components that may have a separate AI / ML model may include Doppler basis selection or quantization coefficient selection.

[0088] The decoder may implement corresponding AI / ML models 820 and 830 to decode the specific component encoded by the AI / ML models 817 and 818. In addition, the decoder may also implement a parsing engine 840 that receives information directly from the eType II processing 815. The parsing engine 840 may use this information to configure the AI / ML models with information for the model to use. For example, because the parsing engine 840 receives the eType II processing output that includes the CSI part 1 and part 2, the parsing engine may directly know parameters used for the encoding, e.g., the number of NZ coefficients. The parsing engine 840 may provide this information to the AI / ML models 820 and 830 to be used for decoding.

[0089] The training for the AI models 817, 818, 820 and 830 may be similar to the training described above. However, in this example embodiment, the training data sets may be specific to the component of the CSI feedback corresponding to the encoder and decoder, e.g., each encoder / decoder pair are trained separately. For example, if the AI / ML model 817 is related to the spatial beam selection, the AI / ML model 817 and the corresponding AI / ML model 820 are trained separately based onAttorney Docket No. 30134 / 89802 Ref. No. P63913WO1 the spatial beam selection component than the AI / ML model 818 and the corresponding AI / ML model 830 associated with another component of the CSI feedback.

[0090] In the examples described above, various training of the AI / ML models have been described. This training may be one or more of some example types of training. In a first example, the training may be performed on the UE side and the trained decoder may be transferred to the to the network side (e.g., to the base station). In a second example, the training may be performed on the network side and the trained encoder may be transferred to the UE side. In a third example, there may be joint training on the UE side and the network side, where both sides exchange gradient information. In this example, there may be a further subtype under which one side does not update the neural network for a training session, e.g., the network side freezes its neural network.

[0091] In a further example, there may be separate training on the network side and the UE side. The UE may perform a first training where the UE side generates the training dataset with a set for the encoder inputs, and another set for the corresponding latent space representation. There may also be network side first training where the network generates the training dataset with a set for the target CSI, and another set for the corresponding latent space representation.

[0092] Fig. 9 shows an example method 900 for reporting CSI feedback according to various example embodiments. The method 900 is described from the perspective of the UE 110.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0093] In 910, the UE 110 receives a CSI configuration from the base station 300. The CSI configuration may provide various information to the UE 110 to configure the CSI feedback. This may include, for example, the CSI-reference Signal (CSI-RS) resources that are to be measured to provide the CSI feedback.

[0094] In 920, the UE 110 measures the configured CSI-RS resources. The UE 110 may be triggered by the base station 300 to perform the measurements, e.g., via a Downlink Control Information (DCI) trigger.

[0095] In 930, the UE 110 encodes the CSI feedback for transmission to the base station 300. Multiple examples of an encoder and the corresponding encoding operations were described above. The UE 110 may implement any of these example encoders to encode the CSI feedback. In 940, the UE 110 transmits the CSI feedback to the base station 300.

[0096] To facilitate multiple vendor training, e.g., multiple UE vendors with a single network vendor, multiple networks vendors with a single UE vendor, multiple UE vendors and multiple network vendors, to align the mapping between latent space representation and reconstructed precoders among multiple vendors, for each vendor, at least a portion of the training data may be generated through a conventional CSI feedback scheme which can be designated as a reference CSI feedback scheme. If multiple vendors (e.g., UE vendors in a UE first training) choose an identical CSI feedback scheme as their reference CSI feedback scheme (e.g., number of antenna ports and their arrangement (N1 / N2), CSI measurement subband configurations (number of subbands, etc.), paramCombination-r16, etc. isAttorney Docket No. 30134 / 89802 Ref. No. P63913WO1 configured identically), then alignment of the mapping between latent space representation and reconstructed precoders among multiple vendors can be partially achieved if a certain percentage (e.g., less than 100%) of the training data for each individual vendor is generated by a reference CSI feedback scheme, but the rest of the training data is according to its discretion. In the case, where all of the training data is generated by a reference CSI feedback scheme, discussion for Fig. 6 applies. To achieve even better alignment, input data (P’s) to the reference CSI feedback scheme may be shared by multiple vendors, so identical pairs of {E(P), f(P)} are available at multiple vendors for the training of their respective AI / ML models. Similarly, one or more component of CSI part 1 and / or CSI part 2, e.g., identical pairs of {truncation of E(P), f(P)} may be available at multiple vendors for their training of their respective AI / ML models for the design exhibited in Figure 8. Examples

[0097] In a first example, a method, comprising, processing, based on signaling received from a base station, a Channel State Information (CSI) configuration comprising CSI Reference Signal (CSI-RS) resources to be measured by the UE, measuring the CSI- RS resources, encoding CSI feedback based on measurement results for the CSI-RS resources, wherein the CSI feedback comprises CSI part 1 and CSI part 2 feedback encoded as a latent space representation and generating, for transmission to the base station, the CSI feedback.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0098] In a second example, the method of the first example, wherein the CSI feedback is encoded using CSI eType II processing.

[0099] In a third example, the method of the second example, wherein the CSI feedback is further encoded using an artificial intelligence (AI) / machine learning (ML) model.

[0100] In a fourth example, the method of the third example, wherein the AI / ML model is trained at the UE.

[0101] In a fifth example, the method of the third example, wherein the AI / ML model is trained at the base station.

[0102] In a sixth example, the method of the third example, wherein the AI / ML model is jointly trained at the UE and the base station.

[0103] In a seventh example, the method of the second example, wherein the CSI feedback is further encoded using a first artificial intelligence (AI) / machine learning (ML) model corresponding to a first component of the CSI feedback and a second AI / ML model corresponding to a second component of the CSI feedback.

[0104] In an eighth example, the method of the seventh example, wherein the CSI feedback comprises a first portion corresponding to the CSI eType II processing, a second portion corresponding to the first component encoded using the first AI / ML model and a third portion corresponding to the second component encoded using the second AI / ML model.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0105] In a ninth example, the method of the first example, further comprising generating a training data set for an artificial intelligence (AI) / machine learning (ML) model based on processing frequency domain precoders according to CSI eType II processing, selecting a predetermined number of quantization coefficients and generating CSI feedback based on combinatorial indexing and a strongest quantization coefficient.

[0106] In a tenth example, a processor configured to perform any of the method of the first through ninth examples.

[0107] In an eleventh example, a user equipment (UE) configured to perform any of the method of the first through ninth examples.

[0108] In a twelfth example, a method, comprising, processing, based on signals received from a user equipment (UE), Channel State Information (CSI) feedback comprising CSI encoded as a latent space representation and processing the CSI feedback using an artificial intelligence (AI) / machine learning (ML) model to reconstruct a precoder used to encode the CSI feedback.

[0109] In a thirteenth example, the method of the twelfth example, wherein the AI / ML model is trained at the UE.

[0110] In a fourteenth example, the method of the twelfth example, wherein the AI / ML model is trained at the base station.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0111] In a fifteenth example, the method of the twelfth example, wherein the AI / ML model is jointly trained at the UE and the base station.

[0112] In a sixteenth example, the method of the twelfth example, wherein the AI / ML model comprises a plurality of AI / ML models, wherein each of the AI / ML models corresponds to a component of the CSI feedback.

[0113] In a seventeenth example, the method of the sixteenth example, wherein the CSI feedback comprises one or more portions, each portion corresponding to one of the plurality of AI / ML models.

[0114] In an eighteenth example, the method of the twelfth example, further comprising generating a training data set for the AI / ML model based on processing frequency domain precoders according to CSI eType II processing, selecting a predetermined number of quantization coefficients and generating CSI feedback based on combinatorial indexing and a strongest quantization coefficient.

[0115] In a nineteenth example, a processor configured to perform any of the method of the twelfth through eighteenth examples.

[0116] In a twentieth example, a base station configured to perform any of the method of the twelfth through eighteenth examples.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0117] Those skilled in the art will understand that the above-described example embodiments may be implemented in any suitable software or hardware configuration or combination thereof. An example hardware platform for implementing the example embodiments may include, for example, an Intel x86 based platform with compatible operating system, a Windows OS, a Mac platform and MAC OS, a mobile device having an operating system such as iOS, Android, etc. The example embodiments of the above described method may be embodied as a program containing lines of code stored on a non-transitory computer readable storage medium that, when compiled, may be executed on a processor or microprocessor.

[0118] Although this application described various embodiments each having different features in various combinations, those skilled in the art will understand that any of the features of one embodiment may be combined with the features of the other embodiments in any manner not specifically disclaimed or which is not functionally or logically inconsistent with the operation of the device or the stated functions of the disclosed embodiments.

[0119] It is well understood that the use of personally identifiable information should follow privacy policies and practices that are generally recognized as meeting or exceeding industry or governmental requirements for maintaining the privacy of users. In particular, personally identifiable information data should be managed and handled so as to minimize risks of unintentional or unauthorized access or use, and the nature of authorized use should be clearly indicated to users.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1

[0120] It will be apparent to those skilled in the art that various modifications may be made in the present disclosure, without departing from the spirit or the scope of the disclosure. Thus, it is intended that the present disclosure cover modifications and variations of this disclosure provided they come within the scope of the appended claims and their equivalent.

Claims

Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1 What is claimed:

1. An apparatus comprising processing circuitry coupled to a memory, the processing circuitry configured to: process, based on signaling received from a base station, a Channel State Information (CSI) configuration comprising CSI Reference Signal (CSI-RS) resources to be measured; measure the CSI-RS resources; and generate CSI feedback based on measurement results for the CSI-RS resources, wherein the CSI feedback comprises CSI part 1 and CSI part 2 feedback encoded as a latent space representation.

2. The apparatus of claim 1, wherein the CSI feedback is generated using CSI eType II processing.

3. The apparatus of claim 2, wherein the CSI feedback is further generated using an artificial intelligence (AI) / machine learning (ML) model.

4. The apparatus of claim 3, wherein the AI / ML model is trained at a user equipment (UE).

5. The apparatus of claim 3, wherein the AI / ML model is trained at the base station.

6. The apparatus of claim 3, wherein the AI / ML model is jointly trained at a user equipment (UE) and the base station.

7. The apparatus of claim 2, wherein the CSI feedback is further generated using a first artificial intelligence (AI) / machine learning (ML) model corresponding to a firstAttorney Docket No. 30134 / 89802 Ref. No. P63913WO1 component of the CSI feedback and a second AI / ML model corresponding to a second component of the CSI feedback.

8. The apparatus of claim 7, wherein the CSI feedback comprises a first portion corresponding to the CSI eType II processing, a second portion corresponding to the first component encoded using the first AI / ML model and a third portion corresponding to the second component encoded using the second AI / ML model.

9. The apparatus of claim 1, wherein the processing circuitry is further configured to: generate a training data set for an artificial intelligence (AI) / machine learning (ML) model based on the processing circuitry being further configured to: process frequency domain precoders according to CSI eType II processing; select a predetermined number of quantization coefficients; and generate CSI feedback based on combinatorial indexing and a strongest quantization coefficient.

10. An apparatus comprising processing circuitry coupled to a memory, the processing circuitry configured to: process, based on signaling received from a user equipment (UE), Channel State Information (CSI) feedback comprising CSI encoded as a latent space representation; and process the CSI feedback using an artificial intelligence (AI) / machine learning (ML) model to reconstruct a precoder used to encode the CSI feedback.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1 11. The apparatus of claim 10, wherein the AI / ML model is trained at the UE.

12. The apparatus of claim 10, wherein the AI / ML model is trained at a base station.

13. The apparatus of claim 10, wherein the AI / ML model is jointly trained at the UE and a base station.

14. The apparatus of claim 10, wherein the AI / ML model comprises a plurality of AI / ML models, wherein each of the AI / ML models corresponds to a component of the CSI feedback.

15. The apparatus of claim 14, wherein the CSI feedback comprises one or more portions, each portion corresponding to one of the plurality of AI / ML models.

16. The apparatus of claim 10, wherein the processing circuitry is further configured to: generate a training data set for the AI / ML model based on the processing circuitry being further configured to: process frequency domain precoders according to CSI eType II processing; selecting a predetermined number of quantization coefficients; and generate CSI feedback based on combinatorial indexing and a strongest quantization coefficient.Attorney Docket No. 30134 / 89802 Ref. No. P63913WO1 17. A method, comprising: processing, based on signaling received from a base station, a Channel State Information (CSI) configuration comprising CSI Reference Signal (CSI-RS) resources to be measured; measuring the CSI-RS resources; and generating CSI feedback based on measurement results for the CSI-RS resources, wherein the CSI feedback comprises CSI part 1 and CSI part 2 feedback encoded as a latent space representation, wherein the CSI feedback is generated using CSI eType II processing using an artificial intelligence (AI) / machine learning (ML) model.

18. The method of claim 17, wherein the AI / ML model is trained at a user equipment (UE).

19. The method of claim 17, wherein the AI / ML model is trained at the base station.

20. The method of claim 17, wherein the AI / ML model is jointly trained at a user equipment (UE) and the base station.

Images