WO2023140772A1

WO2023140772A1 - Hybrid model-learning solution for csi reporting

Info

Publication number: WO2023140772A1
Application number: PCT/SE2023/050045
Authority: WO
Inventors: Xinlin ZHANG; Chandan PRADHAN; Rakesh Ranjan; Roy TIMO; Mattias Frenne
Original assignee: Telefonaktiebolaget Lm Ericsson (Publ)
Priority date: 2022-01-21
Filing date: 2023-01-18
Publication date: 2023-07-27

Abstract

A wireless device configured to communicate with a network node is provided. The wireless device is configured with a configuration for feature extraction. The wireless device is configured to perform a first channel measurement on a first plurality 5 of antenna ports, and to translate the first channel measurement into a second channel measurement based on the configuration for feature extraction. The wireless device is further configured to encode the second channel measurement using an autoencoder based on an artificial neural network to generate an encoded channel measurement, and to transmit to the network node a first indication of the encoded channel measurement 10 for channel estimation.

Description

HYBRID MODEL-LEARNING SOLUTION FOR CSI REPORTING

TECHNICAL FIELD

The present disclosure generally relates to communication in 3^rd Generation Partnership Project (3GPP) networks, and in particular, a method and system for configuring channel state information (CSI) reporting using a hybrid approach.

BACKGROUND

In recent years, obtaining improvements in the various functionalities of different wireless communication systems has become increasingly important as various network operators seek improved performance in a cost-efficient manner.

The 5^th generation mobile wireless communication system (New Radio (NR)) uses orthogonal frequency division multiplexing (OFDM) with configurable bandwidths and subcarrier spacing to efficiently support a diverse set of use-cases and deployment scenarios. With respect to the 4^th generation system (Long term evolution (LTE)), NR improves deployment flexibility, user throughputs, latency, and reliability. The throughput performance gains are enabled, in part, by enhanced support for Multi-User Multiple Input Multiple Output (MU-MIMO) transmission strategies, where two or more wireless devices (WDs) receive data on the same time-frequency resources, i.e., spatially separated transmissions.

The MU-MIMO transmission strategy is illustrated in FIG. 1 of the drawings including an example of transmission and reception chain for the MU-MIMO operations. The order of modulation and precoding, or demodulation and combining respectively, may differ depending on the implementation of MU-MIMO transmission.

According to the MU-MIMO transmission strategy as shown in FIG. 1, a multi - antenna base station with N_TX antenna ports is simultaneously (on the same OFDM time-frequency resources) transmitting information to several wireless devices (WDs): the sequence

is transmitted to WD(1),

is transmitted to WD(2), and so on. Before modulation and transmission, precoding

is applied to each sequence to mitigate multiplexing interference and the transmissions are spatially separated.

Each WD demodulates its received signal and combines receiver antenna signals to obtain an estimate of a transmitted sequence. This estimate

for WD z can be expressed as (neglecting other interference and noise sources except the MU- MIMO interference) shown in equation 1 below:

The second term of the above-mentioned equation (1) represents the spatial multiplexing interference (due to MU-MIMO transmission) seen by WD(i). The goal for the Network (NW) is to construct the set of precoders to meet a given target.

One such target is to make: the norm large (this norm represents the desired channel gain

towards user i); and the norm i small (this norm represents the interference of user

i’s transmission received by user j).

In other words, the precoder VF® may correlate well with the channel observed by WD(i), whereas it may correlate poorly with the channels observed by other WDs.

To construct precoders VF®, i = 1, . . . ,] that enable efficient MU-MIMO transmissions, the NW (e.g., a network node, such as a base station) needs to obtain detailed information about all the user's (WD’s) downlink channels H (t), i = 1, . . ,] .

In the deployments where full channel reciprocity holds, detailed channel information can be obtained from uplink sounding reference signals (SRS) that are transmitted periodically, or on demand, by active WDs. The NW (e.g., network node) can directly estimate the uplink channel from SRS and, therefore (by reciprocity), the downlink channel H® .

However, the NW (e.g., network node) cannot always accurately estimate the downlink channel from uplink reference signals. Consider the following examples:

In frequency division duplex (FDD) deployments, the uplink and downlink channels use different carriers and, therefore, the uplink channel might not provide enough information about the downlink channel to enable MU-MIMO precoding. - In TDD, the NW (e.g., network node) might only estimate part of the uplink channel using SRS, because the WD typically has fewer TX branches than RX branches (in which case only certain columns of the precoding matrix can be estimated using SRS). This is known as partial channel knowledge.

If the NW (e.g., network node) cannot accurately estimate the full downlink channel from uplink transmissions, then active WDs need to report channel information to the NW (e.g., network node) over the uplink. In LTE and NR, this feedback is achieved by the following signalling protocol:

The NW (e.g., network node) periodically transmits Channel State Information reference signals (CSI-RS) over the downlink using N ports.

The WD estimates the downlink channel for each of the N ports from the transmitted CSI-RS.

The WD reports CSI (e.g., CQI, PMI, RI where PMI is the precoder and RI is the rank of the precoder) to the NW (e.g., network node) over the uplink (control or data) channel.

The NW (e.g., network node) uses the WD’s feedback to select suitable precoders for downlink MU-MIMO transmissions.

In NR, both Type I and Type II CSI reporting is configurable, where the CSI Type II reporting protocol has been specifically designed to enable MU-MIMO operations from uplink WD reports.

The CSI Type II normal reporting mode is based on the specification of sets of Discrete Fourier Transform (DFT) basis functions in a precoder codebook. The WD selects and reports the L DFT vectors from the codebook that best match its channel conditions (like the classical codebook precoding matrix indicator (PMI) from earlier 3GPP releases). The number of DFT vectors L is typically 2 or 4 and is configurable by the NW (e.g., network node). In addition, the WD reports how the L DFT vectors should be combined in terms of relative amplitude scaling and co-phasing.

Algorithms to select L, the L DFT vectors, and co-phasing coefficients are outside the scope of the present disclosure and left for the implementation at WD and NW (e.g., network node) sides. Alternatively, as per the Rel. 16 specification only signalling protocols to enable the above message exchanges are defined.

In accordance with the aforementioned description, “DFT beams” may be used interchangeably with DFT vectors. This use of terminology is appropriate, e.g., whenever the base station (network node) has a uniform planar array with antenna elements separated by half of the carrier wavelength.

An example of the CSI type II normal reporting mode is also illustrated in FIG. 2 of the drawings. The selection and reporting of the L DFT vectors b_n and their relative amplitudes a_n is done in a wideband manner that is, the same beams are used for both polarizations over the entire transmission band. The selection and reporting of the DFT vector co-phasing coefficients are done in a subband manner that is, DFT vector cophasing parameters are determined for each of multiple subsets of contiguous subcarriers. The co-phasing parameters are quantized such that e ^76,77 is taken from either a QPSK or 8PSK signal constellation.

With k denoting a sub-band index, the precoder VF [k] reported by the WD to the NW (e.g., network node) can be expressed in expression (2) as follows:

W_v[k] = _ln b_na_neJ^e M , , , (2)

The Type II CSI report can be used by the NW (e.g., network node) to coschedule multiple WDs on the same OFDM time-frequency resources. For example, the NW (e.g., network node) can select WDs that have reported different sets of DFT vectors with weak correlations. The CSI Type II report enables the WD to report a precoder hypothesis that trades CSI resolution against uplink transmission overhead.

NR. 3GPP Release 15 supports Type II CSI feedback using port selection mode, in addition to the above normal reporting mode. In this case,

The base station (e.g., network node) transmits a CSI-RS port in each one of the beam directions.

The WD does not use a codebook to select a DFT vector (a beam), instead, the WD selects one or multiple antenna ports from the CSI-RS resource of multiple ports.

The Type II CSI feedback using port selection gives the base station (e.g., network node) some flexibility to use non-standardized precoders that are transparent to the WD. For the port-selection codebook, the precoder reported by the WD can be described as expression (3) as follows (excluding eventual normalization): ] = ln ena_ne^jen[k] ... (3)

Here, the vector e is a unit vector with only one non-zero element, which can be viewed as a selection vector that selects a port from the set of ports in the measured CSI- RS resource. The WD thus feeds back which ports it has selected, the amplitude factors, and the co-phasing factors.

Now, a background description will be made by considering a scenario of CSI reporting in NR.

In NR, a WD can be configured with one or multiple CSI Report Settings, each configured by a higher layer parameter CSI-ReportConfig. Each CSI-ReportConfig is associated with a Bandwidth Part (BWP) and contains one or more of the following:

- a CSI resource configuration for channel measurement;

- a CSI-IM resource configuration for interference measurement;

- reporting configuration type, i.e., aperiodic CSI (on Physical Uplink Shared Channel (PUSCH)), periodic CSI (on PUCCH), or semi-persistent CSI on Physical Uplink Control Channel (PUCCH) or PUSCH;

- report quantity specifying what to be reported, such as Rank Indicator (RI), Precoding Matrix Indicator (PMI), Channel Quality Indicator (CQI);

- codebook configuration such as type I or type II CSI;

- frequency domain configuration, i.e., subband vs. wideband CQI or PMI, and subband size; and

- CQI table to be used.

A WD may be configured with one or multiple CSI resource configurations for channel measurement and one or more CSI-IM resources for interference measurement. Each CSI resource configuration for channel measurement may contain one or more NZP CSI-RS resource sets. For each NZP CSI-RS resource set, it may further contain one or more NZP CSI-RS resources. A NZP CSI-RS resource may be periodic, semi -persistent, or aperiodic.

Similarly, each CSI-IM resource configuration for interference measurement may contain one or more CSI-IM resource sets. For each CSI-IM resource set, it may further contain one or more CSI-IM resources. A CSI-IM resource may be periodic, semi- persistent, or aperiodic.

Now, a background description will be made by considering a scenario for Artificial Intelligence/Machine Learning (AI/ML) enhanced CSI reporting by Autoencoders (AEs): Recently, artificial neural network-based autoencoders have shown promising results for compressing downlink MIMO channel estimates for uplink feedback. For example:

- Using AEs to improve the accuracy of reported CSI from the WD to the NW (e.g., network node).

The following contributions to the June 2021 3 GPP TSG RAN Rel- 18 Workshop all support AI/ML-enhanced CSI reporting in a potential Rel-18.

An AE is a type of artificial neural network (NN) that may be used to compress and decompress data, in an unsupervised manner, often with high fidelity. FIG. 3 illustrates a simple fully connected (dense) AE. The AE is divided into two parts: an encoder (used to compress the input data X), and a decoder (used to de-compress the input data).

AEs may have different architectures. For example, AEs can be based on dense NNs, multi-dimensional convolution NNs, variational, recurrent NNs, transformer networks, or any combination thereof. However, all AE architectures possess an encoder-bottleneck-decoder structure as illustrated in FIG. 3.

The size of the codeword (denoted by Y in FIG. 3) of an AE is typically significantly smaller than the size of the input data (denoted by X in FIG. 3). The AE encoder may thus reduce the dimensionality of the input features X down to Y. The decoder part of the AE attempts to invert the encoder and reconstruct X with minimal error, which may be according to some predefined loss function.

FIG. 4 illustrates an example method to illustrate how an AE might be used for AI/ML-enhanced CSI reporting in NR. The WD measures the channel in the downlink using CSI-RS (Step SI). The WD estimates that channel for each subcarrier (sc) from each base station TX antenna and at each WD RX antenna. The estimate can be viewed as a three-dimensional channel matrix. The 3D channel matrix represents the MIMO channel estimated over several subcarriers (SC) and is input to the encoder.

The AE encoder is implemented in the WD (Step S2). The output of the AE encoder (Step S2) is signalled from the WD to the NW over the uplink (Step S3) (e.g., via CSI reporting). The AE decoder (Step S4) is implemented in the NW. The code and/or codeword can be viewed as a learned latent representation of the channel. The quantization layer may be connected at the output of the encoder or directly included in the encoder so that the codeword consists of quantized values that are transmitted to the network node (e.g., gNB, base station, etc.) in a CSI report.

The architecture of an AE (e.g., number of layers, nodes per layer, activation function, etc) typically needs to be numerically optimized for CSI reporting via a process called hyperparameter tuning. Properties of the data (e.g., CSI-RS channel estimates), the channel size, uplink feedback rate, and hardware limitations of the encoder and decoder all need to be considered when optimizing the AE’s architecture.

The weights and biases of an AE (with a fixed architecture) are trained to minimize the reconstruction error (the error between the input X and output X) on some training datasets. For example, the weights and biases can be trained to minimize the mean squared error (MSE) (X — X)². Model training is typically done using some variant of the gradient descent algorithm on a large training data set. To achieve good performance during live operation, the training data set should be representative of the actual data the AE will encounter during live operation.

The process of designing an AE (hyperparameter tuning and model training) can be expensive, consuming significant time, compute, memory, and power resources.

Now, a background description will be made by considering a scenario for CSI reporting by Autoencoders:

AE-based CSI reporting is of interest for 3GPP Rel. 18 because of the following reasons:

- AEs can include non-linear transformations (e.g., activation functions) that help improve compression performance and, therefore, MU-MIMO performance for the same uplink overhead. For example, the normal Type II CSI codebooks in 3GPP Rel. 16 are based on linear DFT transformations and SVD decompositions, which cannot fully exploit redundancies in the channel for compression.

- AEs can be trained to exploit long-term redundancies in the propagation environment and/or site (e.g., antenna configuration) for compression purposes. For example, a particular AE does not need to work well for all possible deployments. Improved compression performance is obtained by learning which channel inputs it needs to (and doesn’t need to) reliably reconstruct at the base station.

- AEs can be trained to compensate for antenna array irregularities, including, for example, non-uniformly spaced antenna elements and non-half wavelength element spacing. The Type II CSI codebooks in Rel. 15 and 16, for example, use a two-dimensional DFT codebook designed for a regular planar array with perfect half-wavelength element spacing.

- AEs can be trained so that the used CSI reporting is more robust against, or updated (e.g., via transfer learning and training) to compensate for partially failing hardware as the massive MIMO product ages. For example, over time one or more of the multiple Tx and Rx radio chains in the massive MIMO antenna arrays at the base station will fail to compromise the effectiveness of Type II CSI feedback.

From the above described all the scenarios of the CSI reporting and the AI/ML enhanced CSI reporting using AEs, the following problems can be considered in terms of AE-based CSI compression solution, the complexity of a large AE network, and CSI compression efficiency and accuracy as described below:

In an AE-based CSI compression solution, the raw channel estimate, e.g., in the antenna-frequency domain, may be directly fed into the AE. In theory, a capable AE should be able to identify the underlying characteristics of the channel and find a good latent representation. However, the introduction of AE for CSI reporting in wireless communication systems may also bring in unavoidable complexity. In a model-based CSI compression solution such as Type II CSI in NR, the raw channel estimate is decomposed into several pre-defined and standardized components. A notable limitation with this model-based approach is the model being fixed and nongeneralizable to new scenarios as it would require new standardization efforts in 3GPP to introduce such extended models. The CSI compression efficiency and accuracy can be greatly affected by the actual channel realization with the fixed model approach used in the current NR.

Therefore, there lies a need for a solution to overcome the aforementioned problems.

SUMMARY

This summary is provided to introduce a selection of concepts, in a simplified format, that are further described in the detailed description of the present disclosure. This summary is neither intended to identify key or essential inventive concepts of the present disclosure nor is it intended for determining the scope of the present disclosure. Using the raw large-dimensional original channel as the input to the AE may result in the AE’s being exceedingly huge in size, computationally intensive, and also craving for very long trainings.

It is thus a problem with complexity to implement a large AE network with a huge number of nodes and layers to get good performance using raw channel data. It is also quite challenging to train such a large network. There is also an issue with the power consumption of such large networks which is an issue for battery-powered user equipment devices. Embodiments of the present disclosure may solve and/or mitigate one or more of the above-described problems.

According to an embodiment, the present subject matter describes a method and system for configuring a hybrid CSI compression approach which consists of two steps. A first step is the model-based dimension reduction, and a second step is AE-based compression. The model-based dimension reduction step compresses the channel according to a pre-defined model, then the compressed channel after the model-based step is fed into the AE for further compression. For de-compression, the de-compressed channel after AE will be further de-compressed according to the pre-defined model. In particular, in the Model-based dimension-reduction step, the input channel (or channel feature) is transformed into a new coarse channel feature space with reduced dimension to identify and extract coarse features of the channel. Further, in the AE-based compression step, the lower-dimension transformed channel is compressed to a number of bits that can be signaled over the uplink control or data channels. This compression results in extracting and signaling the detailed channel features to the network node (e.g., gNB) in an efficient way. Furthermore, the features extracted from the modelbased dimension-reduction step and the AE-based compression step are reported to the network node (gNB), and they are used by the network node (gNB) to perform downlink scheduling functionalities (e.g., precoder selection, WD pairing for MU- MIMO, link adaptation).

According to another implementation for the WD side, the present subject matter describes a method for configuring the hybrid CSI compression approach which consists of the following steps as follows:

The WD may be configured with a standardized model-based coarse feature extraction via a parameter set c_model for CSI reporting. The WD may process the estimated channel on the assigned CSI-RS resource of N ports, into a set of _model feature ports.

The WD may use the set of A_model feature ports as the input to the Al-based processing unit.

The WD may report the output of the Al-based processing unit (h_AE) and when applicable also the model-based feature extraction parameters (b_modei) in the CSI report.

According to yet another implementation for the network node (e.g., gNB) side, the present subject matter describes a method for configuring the hybrid CSI compression approach which consists of the following steps as follows:

The network node (gNB) configures a WD with a CSI report.

The CSI report configuration comprises parameters for a standardized modelbased feature extraction (c_model).

The network node (gNB) receives a CSI report from the WD. The CSI report is derived using the configured model for feature extraction

According to another aspect of the present disclosure, a wireless device for supporting configurations for autoencoding channel state information in a wireless communication network is provided which is configured with a configuration for feature extraction (e.g., by a network node). The wireless device is configured to perform a first channel measurement on a first plurality of antenna ports. The wireless device is configured to translate the first channel measurement into a second channel measurement based on the configuration for feature extraction (e.g., using a model-based processing unit). The wireless device is configured to encode the second channel measurement using an autoencoder based on an artificial neural network to generate an encoded channel measurement. The wireless device is configured to cause transmission to the network node of a first indication of the encoded channel measurement for channel estimation.

According to some embodiments of this aspect, the translating of the first channel measurement into a second channel measurement based on the configuration for feature extraction includes determining a plurality of extracted features based on the configuration for feature extraction, where the second channel measurement is associated with the plurality of extracted features, and the wireless device is further configured to cause transmission to the network node of a second indication of the determined plurality of extracted features for channel estimation.

According to some embodiments of this aspect, the determining of the plurality of extracted features based on the configuration for feature extraction includes determining a first plurality of extracted features for a first rank of the first channel measurement, where the first rank is associated with a first signal-to-noise (SNR) ratio, and determining a second plurality of extracted features for a second rank of the first channel measurement, where the second rank is associated with a second SNR ratio weaker than the first SNR ratio, and the second plurality of extracted features is smaller than the first plurality of extracted features based on the second SNR ratio being weaker than the first SNR ratio.

According to some embodiments of this aspect, the translating of the first channel measurement into a second channel measurement based on the configuration for feature extraction includes one or more of reducing a first number of dimensions of the first channel measurement to a second number of dimensions of the second channel measurement, transforming an antenna-frequency domain of the first channel measurement to a beam-delay domain of the second channel measurement, transforming an antenna-frequency-time domain of the first channel measurement to a beam delay- doppler domain of the second channel measurement, transforming a first number of antenna ports of the first channel measurement to a second number of feature ports of the second channel measurement, the second number being smaller than the first number, reducing a first number of spatial beams of the first channel measurement to a second number of spatial beams of the second channel measurement, the second number of spatial beams being selected based on a set of orthogonal discrete Fourier transform (DFT) basis vectors, and reducing a first number of delay taps of the first channel measurement to a second number of delay taps of the second channel measurement based on the set of orthogonal discrete Fourier transform (DFT) basis vectors.

According to some embodiments of this aspect, the first indication is transmitted to the network node in an uplink CSI report. In some embodiments, the configuration for feature extraction indicates one or more of an indicated plurality of extracted features for channel estimation, at least one first spatial beam for extraction, at least one second spatial beam to be excluded from extraction, a number of spatial beams for extraction, at least one first rank for extraction, at least one second rank to be excluded from extraction, and a number of ranks for extraction.

According to another aspect of the present disclosure, a method implemented in a wireless device for supporting configurations for autoencoding channel state information in a wireless communication network is provided, where the wireless device is configured with a configuration for feature extraction (e.g., by a network node). A first channel measurement is performed on a first plurality of antenna ports. The first channel measurement is translated into a second channel measurement based on the configuration for feature extraction. The second channel measurement is encoded using an autoencoder based on an artificial neural network to generate an encoded channel measurement. A first indication of the encoded channel measurement for channel estimation is transmitted to the network node.

According to some embodiments of this aspect, the translating of the first channel measurement into a second channel measurement based on the configuration for feature extraction includes determining a plurality of extracted features based on the configuration for feature extraction, where the second channel measurement is associated with the plurality of extracted features, and the method further includes causing transmission to the network node of a second indication of the determined plurality of extracted features for channel estimation.

According to some embodiments of this aspect, the translating of the first channel measurement into a second channel measurement based on the configuration for feature extraction includes one or more of reducing a first number of dimensions of the first channel measurement to a second number of dimensions of the second channel measurement, transforming an antenna-frequency domain of the first channel measurement to a beam-delay domain of the second channel measurement, transforming an antenna-frequency-time domain of the first channel measurement to a beam delay- doppler domain of the second channel measurement, transforming a first number of antenna ports of the first channel measurement to a second number of feature ports of the second channel measurement, where the second number is smaller than the first number, reducing a first number of spatial beams of the first channel measurement to a second number of spatial beams of the second channel measurement, where the second number of spatial beams is selected based on a set of orthogonal discrete Fourier transform (DFT) basis vectors, and reducing a first number of delay taps of the first channel measurement to a second number of delay taps of the second channel measurement based on the set of orthogonal discrete Fourier transform (DFT) basis vectors.

According to another aspect of the present disclosure, a network node for supporting configurations for autoencoding channel state information in a wireless communication network is provided. The network node is configured to determine a configuration for feature extraction. The network node is configured to cause transmissions of the configuration for feature extraction to the wireless device. The network node is configured to, responsive to causing transmission of the configuration for feature extraction, receive, from the wireless device, a first indication of an encoded channel measurement. The network node is configured to decode the encoded channel measurement using an autoencoder based on an artificial neural network to generate a decoded channel measurement. The network node is configured to translate the decoded channel measurement into a first channel measurement based on the configuration for feature extraction (e.g., using a model-based processing unit). The network node is configured to perform at least one network node action based on the first channel measurement. According to one or more embodiments of this aspect, the at least one network node action includes determining a downlink precoding matrix based on the first channel measurement, and causing transmission of signalling to the wireless device using the downlink precoding matrix. According to one or more embodiments of this aspect, the network node is further configured to receive from the wireless device a second indication of a determined plurality of extracted features for channel estimation, and the translating of the decoded channel measurement into a first channel measurement based on the configuration for feature extraction is based on the determined plurality of extracted features. According to one or more embodiments of this aspect, the plurality of extracted features includes a first plurality of extracted features for a first rank of the first channel measurement, the first rank being associated with a first signal-to-noise (SNR) ratio, and a second plurality of extracted features for a second rank of the first channel measurement, the second rank being associated with a second SNR ratio weaker than the first SNR ratio, where the second plurality of extracted features is smaller than the first plurality of extracted features based on the second SNR ratio being weaker than the first SNR ratio. According to one or more embodiments of this aspect, the translating of the decoded channel measurement into a first channel measurement based on the configuration for feature extraction includes one or more of increasing a first number of dimensions of the decoded channel measurement to a second number of dimensions of the first channel measurement, transforming a beam-delay domain of the decoded channel measurement to an antenna-frequency domain of the first channel measurement, transforming a beam delay-doppler domain of the decoded channel measurement to an antenna-frequency-time domain of the first channel measurement, transforming a first number of feature ports of the decoded channel measurement to a second number of antenna ports of the first channel measurement, where the first number is smaller than the second number, increasing a first number of spatial beams of the decoded channel measurement to a second number of spatial beams of the first channel measurement based on a set of orthogonal discrete Fourier transform (DFT) basis vectors, and increasing a first number of delay taps of the decoded channel measurement to a second number of delay taps of the first channel measurement based on the set of orthogonal discrete Fourier transform (DFT) basis vectors.

According to one or more embodiments of this aspect, the first indication is received from the wireless device in an uplink CSI report. According to one or more embodiments of this aspect, the configuration for feature extraction indicates one or more of an indicated plurality of extracted features for channel estimation, at least one first spatial beam for extraction, at least one second spatial beam to be excluded from extraction, a number of spatial beams for extraction, at least one first rank for extraction, at least one second rank to be excluded from extraction, and a number of ranks for extraction.

According to another aspect of the present disclosure, a method implemented in a network node for supporting configurations for autoencoding channel state information in a wireless communication network is provided. A configuration for feature extraction is determined. The configuration for feature extraction is transmitted to the wireless device. Responsive to causing transmission of the configuration for feature extraction, a first indication is received, from the wireless device, of an encoded channel measurement. The encoded channel measurement is decoded using an autoencoder based on an artificial neural network to generate a decoded channel measurement. The decoded channel measurement is translated into a first channel measurement based on the configuration for feature extraction. At least one network node action is performed based on the first channel measurement.

According to one or more embodiments of this aspect, the at least one network node action includes determining a downlink precoding matrix based on the first channel measurement, and causing transmission of signalling to the wireless device using the downlink precoding matrix. According to one or more embodiments of this aspect, the method further includes receiving from the wireless device a second indication of a determined plurality of extracted features for channel estimation, and the translating of the decoded channel measurement into a first channel measurement based on the configuration for feature extraction is based on the determined plurality of extracted features. According to one or more embodiments of this aspect, the plurality of extracted features includes a first plurality of extracted features for a first rank of the first channel measurement, the first rank being associated with a first signal-to-noise (SNR) ratio, and a second plurality of extracted features for a second rank of the first channel measurement, the second rank being associated with a second SNR ratio weaker than the first SNR ratio, where the second plurality of extracted features is smaller than the first plurality of extracted features based on the second SNR ratio being weaker than the first SNR ratio. According to one or more embodiments of this aspect, the translating of the decoded channel measurement into a first channel measurement based on the configuration for feature extraction includes one or more of increasing a first number of dimensions of the decoded channel measurement to a second number of dimensions of the first channel measurement, transforming a beam-delay domain of the decoded channel measurement to an antenna-frequency domain of the first channel measurement, transforming a beam delay-doppler domain of the decoded channel measurement to an antenna-frequency-time domain of the first channel measurement, transforming a first number of feature ports of the decoded channel measurement to a second number of antenna ports of the first channel measurement, where the first number is smaller than the second number, increasing a first number of spatial beams of the decoded channel measurement to a second number of spatial beams of the first channel measurement based on a set of orthogonal discrete Fourier transform (DFT) basis vectors, and increasing a first number of delay taps of the decoded channel measurement to a second number of delay taps of the first channel measurement based on the set of orthogonal discrete Fourier transform (DFT) basis vectors.

To further clarify the advantages and features of the present disclosure, a more particular description of the present disclosure will be rendered by reference to specific embodiments thereof, which is illustrated in the appended drawings. It is appreciated that these drawings depict only typical embodiments of the present disclosure and are therefore not to be considered limiting of its scope. The present disclosure will be described and explained with additional specificity and detail with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

These and other features, aspects, and advantages of the present disclosure will become better understood when the following detailed description is read with reference to the accompanying drawings in which like characters represent like parts throughout the drawings, wherein:

FIG. 1 illustrates an MU-MIMO transmission strategy including an example of transmission and reception chain for the MU-MIMO operations;

FIG. 2 illustrates an example of the CSI type II normal reporting mode;

FIG. 3 illustrates an example fully-connected autoencoder;

FIG. 4 illustrates an example method to indicate how an AE might be used for AI/ML-enhanced CSI reporting in NR;

FIG. 5 is a block diagram of a network node communicating with a wireless device over an at least partially wireless connection according to some embodiments of the present disclosure;

FIG. 6 is a flowchart of an example process in a wireless device for the configuration of a hybrid CSI compression and reporting scheme comprising a modelbased dimension-reduction step and an AE-based learned compression step, according to some embodiments of the present disclosure;

FIG. 7 is a flowchart of an example process in a network node for the configuration of a hybrid CSI compression and reporting scheme comprising a modelbased dimension-reduction step and an AE-based learned compression step, according to some embodiments of the present disclosure;

FIG. 8 illustrates a general system architecture for the configuration of a hybrid CSI compression and reporting scheme comprising a model-based dimensionreduction step and an AE-based learned compression step, according to an embodiment of the present disclosure;

FIG. 9 illustrates an example illustration of CSI-RS ports and feature ports of a model-based feature extraction device, according to an embodiment of the present disclosure;

FIG. 10 illustrates an alternate representation of system architecture for the configuration of the hybrid CSI compression and reporting scheme, according to an alternate embodiment of the present disclosure; FIG. 11 illustrates yet another alternate representation of the system architecture for the configuration of the hybrid CSI compression and reporting scheme, according to an alternate embodiment of the present disclosure;

FIG. 12 illustrates yet another alternate representation of the system architecture for the configuration of the hybrid CSI compression and reporting scheme, according to an alternate embodiment of the present disclosure;

FIG. 13 illustrates advantages of the system architecture shown in Figure 5 of the drawings along with an example illustration of CDF plot for NMSE performance for hybrid beam-delay domain approach, according to an embodiment of the present disclosure; and

FIG. 14 illustrates an example cell-edge user for a system-level simulation at 2GHz, in accordance with an embodiment of the present disclosure; and

FIG. 15 illustrates an example mean-user throughput diagram for a system-level simulation at 2GHz, in accordance with an embodiment of the present disclosure.

Further, skilled artisans will appreciate that elements in the drawings are illustrated for simplicity and may not have necessarily been drawn to scale. For example, the flow charts illustrate the method in terms of the most prominent steps involved to help to improve understanding of aspects of the present disclosure. Furthermore, in terms of the construction of the device, one or more components of the device may have been represented in the drawings by conventional symbols, and the drawings may show only those specific details that are pertinent to understanding the embodiments of the present disclosure so as not to obscure the drawings with details that will be readily apparent to those of ordinary skill in the art having the benefit of the description herein.

DETAILED DESCRIPTION

For the purpose of promoting an understanding of the principles of the present disclosure, reference will now be made to the various embodiments and specific language will be used to describe the same. It will nevertheless be understood that no limitation of the scope of the present disclosure is thereby intended, such alterations and further modifications in the illustrated system, and such further applications of the principles of the present disclosure as illustrated therein being contemplated as would normally occur to one skilled in the art to which the present disclosure relates. It will be understood by those skilled in the art that the foregoing general description and the following detailed description are explanatory of the present disclosure and are not intended to be restrictive thereof.

Reference throughout this specification to “an aspect”, “another aspect” or similar language means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present disclosure. Thus, appearances of the phrase “in an embodiment”, “in another embodiment”, and similar language throughout this specification may, but do not necessarily, all refer to the same embodiment.

The terms “comprises”, “comprising”, or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process or method that comprises a list of steps does not include only those steps but may include other steps not expressly listed or inherent to such process or method. Similarly, one or more devices or sub-systems or elements or structures or components proceeded by “comprises... a” does not, without more constraints, preclude the existence of other devices or other sub-systems or other elements or other structures or other components or additional devices or additional subsystems or additional elements or additional structures or additional components.

The term “network node” used herein can be any kind of network node comprised in a radio network which may further comprise any of base station (BS), radio base station, base transceiver station (BTS), base station controller (BSC), radio network controller (RNC), g Node B (gNB), evolved Node B (eNB or eNodeB), Node B, multi -standard radio (MSR) radio node such as MSR BS, multi-cell/multicast coordination entity (MCE), integrated access and backhaul (IAB) node, relay node, donor node controlling relay, radio access point (AP), transmission points, transmission nodes, Remote Radio Unit (RRU) Remote Radio Head (RRH), a core network node (e.g., mobile management entity (MME), self-organizing network (SON) node, a coordinating node, positioning node, Minimizing Drive Testing (MDT) node, etc.), an external node (e.g., 3rd party node, a node external to the current network), nodes in distributed antenna system (DAS), a spectrum access system (SAS) node, an element management system (EMS), etc. The network node may also comprise test equipment. The term “radio node” used herein may be used to also denote a wireless device (WD) or a radio network node. In some embodiments, the non-limiting terms wireless device (WD) and user equipment (UE) are used interchangeably. The wireless device herein can be any type of wireless device capable of communicating with a network node or another wireless device over radio signals, such as wireless device. The wireless device may also be a radio communication device, target device, device to device (D2D) wireless device, machine type wireless device or wireless device capable of machine to machine communication (M2M), low-cost and/or low-complexity wireless device, a sensor equipped with wireless device, Tablet, mobile terminals, smart phone, laptop embedded equipped (LEE), laptop mounted equipment (LME), USB dongles, Customer Premises Equipment (CPE), an Internet of Things (loT) device, a Narrowband loT (NB-IOT) device, a net-zero-energy-consumption device, etc.

Referring now to the drawing figures, in which like elements are referred to by like reference numerals, there is shown in FIG. 5 a system 10 including a network node 12 and a wireless device 14. Network node 12 includes hardware 16 enabling it to communicate with one or more wireless device(s) 14 and/or other network node(s) 12. The hardware 16 may include a radio interface 18 for setting up and maintaining a wired or wireless connection with an interface of a different communication device, which may be formed as or may include, for example, one or more RF transmitters, one or more RF receivers, and/or one or more RF transceivers. The radio interface 18 may be configured to facilitate a connection 20 to a wireless device 14 or another network node 12, which may be a direct and/or indirect connection.

In the embodiment shown, the hardware 16 of the network node 12 further includes processing circuitry 22. The processing circuitry 22 may include a processor 24 and a memory 26. Thus, the network node 12 further has software 28 stored internally in, for example, memory 26, or stored in external memory (e.g., database, storage array, network storage device, etc.) accessible by the network node 12 via an external connection. The software 28 may be executable by the processing circuitry 22. The processing circuitry 22 may be configured to control any of the methods and/or processes described herein and/or to cause such methods, and/or processes to be performed, e.g., by network node 12. Processor 24 corresponds to one or more processors 24 for performing network node 12 functions described herein. The memory 26 is configured to store data, programmatic software code and/or other information described herein. In some embodiments, the software 28 may include instructions that, when executed by the processor 24 and/or processing circuitry 22, causes the processor 24 and/or processing circuitry 22 to perform the processes described herein with respect to network node 12. For example, processing circuitry 22 of the network node 12 may include a network (NW) model-based processing unit 30, which is configured to support configurations for model -based processing of channel state information in a wireless communication network, and may include NW autoencoder unit 32 which is configured to support configurations for autoencoding (and/or decoding) channel state information in a wireless communication network.

Referring still to FIG. 5, the wireless device 14 may have hardware 33 that may include a radio interface 34 configured to set up and maintain connection 20 with a network node 12. The radio interface 34 may be formed as or may include, for example, one or more RF transmitters, one or more RF receivers, and/or one or more RF transceivers. The hardware 33 of the wireless device 14 further includes processing circuitry 36. The processing circuitry 36 may include a processor 38 and memory 40. The processor 38 may be configured to access (e.g., write to and/or read from) memory 40.

The wireless device 14 may further comprise software 42, which is stored in, for example, memory 40 at the wireless device 14, or stored in external memory (e.g., database, storage array, network storage device, etc.) accessible by the wireless device 14. The software 42 may be executable by the processing circuitry 36.

The processing circuitry 36 may be configured to control any of the methods and/or processes described herein and/or to cause such methods, and/or processes to be performed, e.g., by wireless device 14. The processor 38 corresponds to one or more processors 38 for performing wireless device 14 functions described herein. The wireless device 14 includes memory 40 that is configured to store data, programmatic software code and/or other information described herein. In some embodiments, the software 42 may include instructions that, when executed by the processor 38 and/or processing circuitry 36, causes the processor 38 and/or processing circuitry 36 to perform the processes described herein with respect to wireless device 14. For example, the processing circuitry 36 of the wireless device 14 may include a WD model-based processing unit 44 which is configured to support configurations for model-based processing of channel state information in a wireless communication network, and a WD autoencoder unit 46 which is configured to support configurations for autoencoding channel state information in a wireless communication network.

Although FIG. 5 shows various “units” such as NW model-based processing unit 30, NW autoencoder unit 32, WD model-based processing unit 44, and WD autoencoder unit 46 as being within a respective processor, it is contemplated that these units may be implemented such that a portion of the unit is stored in a corresponding memory within the processing circuitry. In other words, the units may be implemented in hardware or in a combination of hardware and software within the processing circuitry.

FIG. 6 is a flowchart of an example process in a wireless device 14 according to some embodiments of the present disclosure for supporting configurations for autoencoding channel state information in a wireless communication network. One or more blocks described herein may be performed by one or more elements of wireless device 14 such as by one or more of processing circuitry 36 (including the WD modelbased processing unit 44 and/or WD autoencoder unit 46), etc., where wireless device 14 is configured with a configuration for feature extraction (e.g., is preconfigured, is configured by network node 12, etc.). Wireless device 14 is configured to perform (Block SI 00) a first channel measurement on a first plurality of antenna ports. Wireless device 14 is configured to translate (Block S102) the first channel measurement into a second channel measurement based on the configuration for feature extraction (e.g., using WD model-based processing unit 44). Wireless device 14 is configured to encode (Block SI 04) the second channel measurement using an autoencoder (e.g., as implemented by WD autoencoder unit 46) based on an artificial neural network to generate an encoded channel measurement. Wireless device 14 is configured to cause transmission (Block SI 06) to the network node 12 of a first indication of the encoded channel measurement for channel estimation.

In some embodiments, the translating of the first channel measurement into a second channel measurement based on the configuration for feature extraction includes determining a plurality of extracted features based on the configuration for feature extraction, where the second channel measurement is associated with the plurality of extracted features, and the wireless device 14 is further configured to cause transmission to the network node 12 of a second indication of the determined plurality of extracted features for channel estimation. In some embodiments, the determining of the plurality of extracted features based on the configuration for feature extraction includes determining a first plurality of extracted features for a first rank of the first channel measurement, the first rank being associated with a first signal-to-noise (SNR) ratio, and determining a second plurality of extracted features for a second rank of the first channel measurement, where the second rank is associated with a second SNR ratio weaker than the first SNR ratio, and the second plurality of extracted features is smaller than the first plurality of extracted features based on the second SNR ratio being weaker than the first SNR ratio.

In some embodiments, the translating of the first channel measurement into a second channel measurement based on the configuration for feature extraction includes one or more of reducing a first number of dimensions of the first channel measurement to a second number of dimensions of the second channel measurement, transforming an antenna-frequency domain of the first channel measurement to a beam-delay domain of the second channel measurement, transforming an antenna-frequency -time domain of the first channel measurement to a beam delay-doppler domain of the second channel measurement, transforming a first number of antenna ports of the first channel measurement to a second number of feature ports of the second channel measurement, the second number being smaller than the first number, reducing a first number of spatial beams of the first channel measurement to a second number of spatial beams of the second channel measurement, the second number of spatial beams being selected based on a set of orthogonal discrete Fourier transform (DFT) basis vectors, and reducing a first number of delay taps of the first channel measurement to a second number of delay taps of the second channel measurement based on the set of orthogonal discrete Fourier transform (DFT) basis vectors.

In some embodiments, the first indication is transmitted to the network node 12 in an uplink CSI report. In some embodiments, the configuration for feature extraction indicates one or more of an indicated plurality of extracted features for channel estimation, at least one first spatial beam for extraction, at least one second spatial beam to be excluded from extraction, a number of spatial beams for extraction, at least one first rank for extraction, at least one second rank to be excluded from extraction, and a number of ranks for extraction.

FIG. 7 is a flowchart of an example process in a network node 12 according to some embodiments of the present disclosure for supporting configurations for autoencoding channel state information in a wireless communication network. One or more blocks described herein may be performed by one or more elements of network node 12 such as by one or more of processing circuitry 22 (including the NW modelbased processing unit 30 and/or NW autoencoder unit 32), etc. Network node 12 is configured to determine (Block SI 08) a configuration for feature extraction. Network node 12 is configured to cause transmission (Block SI 10) of the configuration for feature extraction to the wireless device 14. Network node 12 is configured to, responsive to causing transmission of the configuration for feature extraction, receive (Block SI 12), from the wireless device 14, a first indication of an encoded channel measurement. Network node 12 is configured to decode (Block SI 14) the encoded channel measurement using an autoencoder (e.g., NW autoencoder unit 32) based on an artificial neural network to generate a decoded channel measurement. Network node 12 is configured to translate (Block SI 16) the decoded channel measurement into a first channel measurement based on the configuration for feature extraction (e.g., using NW model-based processing unit 30). Network node 12 is configured to perform (Block SI 18) at least one network node 12 action based on the first channel measurement.

In some embodiments, the at least one network node 12 action includes determining a downlink precoding matrix based on the first channel measurement, and causing transmission of signalling to the wireless device 14 using the downlink precoding matrix. In some embodiments, the network node 12 is further configured to receive from the wireless device 14 a second indication of a determined plurality of extracted features for channel estimation, and the translating of the decoded channel measurement into a first channel measurement based on the configuration for feature extraction is based on the determined plurality of extracted features. In some embodiments, the plurality of extracted features includes a first plurality of extracted features for a first rank of the first channel measurement, the first rank being associated with a first signal -to-noise (SNR.) ratio, and a second plurality of extracted features for a second rank of the first channel measurement, the second rank being associated with a second SNR. ratio weaker than the first SNR. ratio, where the second plurality of extracted features is smaller than the first plurality of extracted features based on the second SNR. ratio being weaker than the first SNR. ratio. In some embodiments, the translating of the decoded channel measurement into a first channel measurement based on the configuration for feature extraction includes one or more of increasing a first number of dimensions of the decoded channel measurement to a second number of dimensions of the first channel measurement, transforming a beam-delay domain of the decoded channel measurement to an antennafrequency domain of the first channel measurement, transforming a beam delay-doppler domain of the decoded channel measurement to an antenna-frequency-time domain of the first channel measurement, transforming a first number of feature ports of the decoded channel measurement to a second number of antenna ports of the first channel measurement, the first number being smaller than the second number, increasing a first number of spatial beams of the decoded channel measurement to a second number of spatial beams of the first channel measurement based on a set of orthogonal discrete Fourier transform (DFT) basis vectors, and increasing a first number of delay taps of the decoded channel measurement to a second number of delay taps of the first channel measurement based on the set of orthogonal discrete Fourier transform (DFT) basis vectors.

In some embodiments, the first indication is received from the wireless device 14 in an uplink CSI report. In some embodiments, the configuration for feature extraction indicates one or more of an indicated plurality of extracted features for channel estimation, at least one first spatial beam for extraction, at least one second spatial beam to be excluded from extraction, a number of spatial beams for extraction, at least one first rank for extraction, at least one second rank to be excluded from extraction, and a number of ranks for extraction.

FIG. 8 illustrates a general system 10 architecture for the configuration of a hybrid CSI compression and reporting scheme comprising a model-based dimensionreduction step and an AE-based learned compression step in the wireless device 14 (e.g., as performed by WD model -based processing unit 44 and/or WD autoencoder unit 46), according to an embodiment of the present disclosure.

In accordance with FIG. 8, wireless device 14 (e.g., via processing circuitry 36) is configured to estimate the DL channel based on the configured DL reference signals (e.g., CSI-RS, DMRS, etc.) (Step SI 19), and produces a channel estimate H, for example, in the antenna-frequency domain. The raw channel H can be expressed per CSI-RS port (TX side), per receive antenna (RX side), per frequency subband, and measured at one or more points in time. Hence, in the most general cases, the channel H is a four-dimensional matrix or tensor. The raw channel estimate H is further processed in the wireless device 14 (e.g., via the WD model-based processing unit 44) using a model-based method, through a function H_model ⁼ /(#) where (■) defines the processing performed in the WD modelbased processing unit 44, which aims to extract (Step S120) certain coarse features of the channel which are the most relevant information for the network node 12 (e.g., a base station network node 12) in order to perform precoding and transmission in the downlink to the wireless device 14 . In this example, a model-based dimension reduction is used.

Such features can be, for example, the number of and direction/angle of dominant propagation paths (e.g., those with the largest energy / lowest propagation loss), and possibly also the delay information associated with each of these dominant paths. The extracted features are encoded by wireless device 14 into bits (e.g., denoted as b_modei) and reported (Step S 121) to the network node 12 (e.g., a gNB), as part of the uplink CSI report. After feature extraction (Step S120), a compressed channel W_modc| is ^als° produced.

In general, H_modei may h^ave reduced dimension compared to the raw channel estimate H. However, this does not need to always hold, in some cases, the model -based dimension reduction step can maintain the dimension of H, while only transforming the channel to another domain where the transformed channel, H_model, is easier for the WD autoencoder unit 46 to compress. Such an example could be to transform the channel from the antenna-frequency domain to the beam-delay domain where the channel representation is sparser. Another example would be to transform the channel from the antenna-frequency -time domain to the beam-delay-doppler domain.

The output from the model-based dimension reduction step, H_model, is then fed into the WD autoencoder unit 46 for further compression (Step S122). Another set of features, e.g., fine details of the channel, will be extracted in this step, such features can be, for example, small-scale fading channel coefficients, and eigenvectors and values of the channel. The extracted features (excluding information that is encoded by the WD autoencoder unit 46 and/or step) are encoded information bits (e.g., denoted as h_AE)) and to be reported to network node 12 (e.g., gNB), as a part of the uplink CSI report (Step S124) (e.g., via connection 20, which may be an air interface).

On the network node 12 side, the network node receives the encoded information bits and the extracted features, and performs AE-based decompression and/or AE-based channel reconstruction (Step S126), and performs model-based channel reconstruction and/or model -based dimension expansion (Step S128), which outputs a channel estimate, H.

In particular, in context with the 3GPP specifications, the wireless device 14 is configured using higher layer signaling, such as RRC, to measure the channel H using the N CSI-RS ports of the CSI resource for channel measurement in the higher layer CSI report configuration (CSI-ReportConfig) . The CSI report configuration may be extended to include a set of parameters c_model related to the model-based feature extraction procedure such as the number of spatial domain bases to use how many spatial domain bases after down-selection (i.e., the L value).

The output of the model-based feature extraction step may define a new set of Nmodei variables/values which for simplicity are denoted as feature ports, which may be interpreted as virtual/processed CSI-RS antenna ports. Typically, A_model < /, but not necessarily. The N_model feature ports are fed into the AE (Step S122) for further feature extraction.

In 3 GPP specifications, the feature ports may not be referred to as “ports” but simply as variables. However, the terminology “ports” has been used in the present disclosure since the model-based feature extraction can be interpreted as a transformation of a first set of ports to a second set of ports, without deviating from the scope of the present disclosure.

Each of the N_model feature ports may further include a set of multiple estimated/calculated values (where a value is a complex number). Each value, for example, may be associated with a reporting subband (in case of subband feature extraction is configured by the higher layer parameters in c_model) or one per identified tap delay value.

In one illustrative example, the model-based feature extraction step uses the N = 32 port CSI-RS for each of 1V₃ = 52 configured frequency subbands, performs a spatial domain projection onto L = 8 two-dimensional DFT bases out of a space of 64 possible such bases. In other words, the main 8 directions of the channel that contain most of the received energy are selected from among all possible directions. The raw channel projected onto these L = 8 bases (separately for each subband) is then the N_model = 8 output feature ports, where each such port contains one value for each of the 52 subbands. Accordingly, an example illustration of input (CSI-RS ports) and output (feature ports) of a model -based feature extraction device is also shown in FIG. 9 of the drawings. Those skilled in the art will appreciate that the aforementioned architecture in FIG. 9 should be viewed as a general framework and is not intended to limit the scope of the present disclosure. Within the aforementioned framework, some components may take other forms.

The value A_model, i.e., the number of output feature ports, from the feature extraction unit (e.g., WD model-based processing unit 44) may be configured explicitly in the CSI report configuration, or it can be implicitly derived from the parameters in model •

The value A_model may also depend on the selected rank by the wireless device 14, for example, if the rank is one, then a small value of _model is sufficient. The selected rank may be encoded into b_modei, the output of the feature extraction unit. In this case, the standard may contain a mapping between the reported rank indication and the number of used feature ports ( _model) as the input to the WD autoencoder unit 46. Alternatively, the selected value of A_model can be signed by the WD 14 to the network node 12 (e.g., gNB).

According to an example embodiment of the present disclosure, various procedures from the wireless device 14 (e.g., performed by processing circuitry 36, including the wireless device autoencoder unit 46 and/or WD model-based processing unit 44) side is described in the steps mentioned below:

- The wireless device 14 receives a CSI report configuration (Step S130) via higher layer signalling (e.g., RRC), from the network node 12 (e.g., a gNB), wherein one or multiple CSI-RS resources of N ports are configured for CSI measurements. Additionally, at least one or multiple of the following are configured:

- the model for feature extraction (c_model);

- the function (■) that specifies the processing in the WD model -based processing unit 44;

- the features that should be extracted from the WD model-based processing unit 44;

- the number of feature ports (N_model) extracted from the WD modelbased processing unit 44; - information for the wireless device 14 to produce quantized output (binary bits) from the encoder at WD autoencoder unit 46 (e.g., the number of bits for quantizing the real and imaginary part (or phase and amplitude) of the soft output from the WD autoencoder unit 46, a quantization function); and

- restrictions on the CSI report (e.g., reportQuantity, rank restriction) a loss function for the WD autoencoder unit 46.

- The wireless device 14 measures and estimates N port channel H (Step S132).

- The wireless device 14 performs (e.g., via processing circuitry 36, including the wireless device autoencoder unit 46 and/or WD model -based processing unit 44) feature extraction (Step SI 34) into /V_model feature ports and outputs any eventual decisions made in this step, such as selecting spatial basis, selecting rank, etc. These are encoded into b_modei-

- The WD autoencoder unit 46 uses (Step SI 36) the /V_model feature ports as the input and outputs the b_AE bits, possibly quantizing the AE output in order to map the soft valued latent feature vectors outputted by the WD autoencoder unit 46 to a stream of hard bits to be included in the CSI report.

Hence, both the encoded bits for the features obtained from both the model-based dimension reduction step (i.e., b_modei) and the AE-based compression step (i.e., b_AE) (e.g., as performed by processing circuitry processing circuitry 36, including the wireless device autoencoder unit 46 and/or WD model-based processing unit 44) are reported by WD 14 to the network node 12 (e.g., gNB) according to the configured CSI report configuration.

According to an embodiment of the present disclosure, future Rel. 18 may describe this as “If the UE is configured with a model-based feature extraction through the parameter set c_model, then the UE shall process the estimated channel on the assigned CSI-RS resource of N ports, into a set of A_model feature ports, which the UE shall further use as the input to the Al-based processing unit. The UE shall report the output of the Al-based processing unit (b_AE) and when applicable also the model-based feature extraction parameters (b_modei) iⁿ the CSI report.”

Further, according to an example embodiment of the present disclosure, various procedures from the network node 12 side, e.g., as performed by processing circuitry 22 (including the NW autoencoder unit 32 and/or NW model-based processing unit 30), are described in steps mentioned below:

- network node 12 (e.g., a gNB) configures the wireless device 14 with a CSI report configuration, CSI-RS resource of N ports for CSI measurements, and a model for feature extraction (c_model) .

- network node 12 receives a CSI report from the wireless device 14 and the received encoded bits in the CSI report that are associated with the features extracted from the AE-based compression step (e.g., as performed by processing circuitry 22, including the NW autoencoder unit 32 and/or NW model-based processing unit 30), i.e., b_AE, and optionally the extraction unit output information b_model.

- network node 12 obtains (e.g., via NW autoencoder unit 32) an estimate for the compressed channel, H_modcE based on b_AE by using the AE-based decoder functionality (e.g., as performed by NW autoencoder unit 32).

- network node 12 decodes/decompresses (e.g., via NW model-based processing unit 30) the coarse features based on the received encoded bits from the modelbased dimension reduction step, i.e., b_modei-

- Then, based on the decoded/decompressed coarse channel features and the estimate H_moiicE the network node 12 obtains (e.g., via processing circuitry 22) a final estimate for the raw channel H.

- network node 12 uses the raw channel from the wireless device 14 to compute (e.g., via processing circuitry 22) the downlink MIMO precoding matrix (per precoding resource group (PRG) of subcarriers or a single wideband precoding matrix).

- network node 12 transmits (e.g., via radio interface 18) Physical Downlink Shared Channel (PDSCH) data to the wireless device 14 using the computed MIMO precoding matrix/matrices.

According to another embodiment of the present disclosure, the input to the model -based dimension reduction step may be some function of H, e.g., u(H), instead of the raw channel estimate H. Such function u(H) may be calculating the eigenvectors or a subset of eigenvectors of H.

According to yet another embodiment of the present disclosure, the channel produced from the model-based dimension reduction step, H_m0licE may instead be some function of H_modc], e.g., (H_modc|). Such function (H_modc|) may be calculating the eigenvectors or a subset of eigenvectors of W_modc|. Accordingly, 6_modc| may be used by network node 12 to produce an estimate of (H_modc|), which will be used for estimating the raw channel.

According to yet another embodiment of the present disclosure, the output from the decoder (e.g., as performed by NW autoencoder unit 32) may be some function of //model and b_modei, e.g., y(/?_modei, b_modei)- For example, such function y(/?_modei, b_model) may be calculating a PMI and hence the output of the function “y” is the PMI.

According to yet another embodiment of the present disclosure, the model-based dimension reduction/decompression step (e.g., as performed by WD model-based processing unit 44 and/or NW model-based processing unit 30) may be layer-dependent, i.e., different transmission layers may be compressed differently.

According to yet another embodiment of the present disclosure, the model-based dimension reduction transformations (e.g., as performed by WD model -based processing unit 44 and/or NW model-based processing unit 30) are standardized and built into the autoencoder(s) (e.g., WD autoencoder unit 46 and/or NW autoencoder unit 32). That is, the network node 12 may configure the wireless device 14 to use an autoencoder functionality (e.g., as performed by WD autoencoder unit 46) with a certain model -based feature extraction built into the WD autoencoder unit 46.

According to yet another implementation of the present disclosure, alternate system architectures for the configuration of the hybrid CSI compression and reporting scheme will be described, according to yet another embodiment of the present disclosure.

According to one another embodiment of the present disclosure, the autoencoding functionality (e.g., as performed by WD autoencoder unit 46 and/or NW autoencoder unit 32) of the system 10 architecture of FIG. 8 can be replaced with other neural network architectures, e.g., a transformer.

According to one another embodiment of the present disclosure, the b_modei could also be input to the WD autoencoder unit 46 to assist the autoencoding for compression. An example illustration of such system 10 architecture is shown in FIG. 10 of the drawings. In particular, FIG. 10 illustrates an alternate representation of system 10 architecture for the configuration of the hybrid CSI compression and reporting scheme, in which channel measurement (Step S138) produces a channel estimate H, which is fed into the model-based feature extraction (e.g., as performed by WD model-based processing unit 44) (Step S140), which produces b_modei which is fed into the AE-based compression step (e.g., as performed by WD autoencoder unit 46) (Step S142). The CSI report is transmitted to the network node 12 (Step S144), which performs AE-based decompression and/or AE-based channel reconstruction (e.g., as performed by NW autoencoder unit 32) (Step 146), and performs model-based dimension expansion and/or model-based channel reconstruction (e.g., as performed by NW model-based processing unit 30) (Step S148).

According to yet another embodiment of the present disclosure, H may also be input the WD autoencoder unit 46 to assist the autoencoding for compression. In this way, the autoencoding may have better performance since no information may be lost, whereas some information may have lost when transforming H to H_modc|. FIG. 11 illustrates yet another alternate representation of the system 10 architecture for the configuration of the hybrid CSI compression and reporting scheme, where H is fed into the WD autoencoder unit 46. In this example, the WD 14 performs a channel measurement (Step SI 50), and performs a model -based feature extraction (e.g., as performed by WD model -based processing unit 44) (Step SI 52). The channel measurement H is fed into the AE-based compression step (e.g., as performed by WD autoencoder unit 46) (Step S 154). The WD transmits the CSI report to the network node 12 (Step SI 56), e.g., via communication channel 20, which performs AE-based decompression and/or AE-based channel reconstruction (e.g., as performed by NW autoencoder unit 30) (Step 158), and performs model-based dimension expansion and/or model-based channel reconstruction (e.g., as performed by NW model-based processing unit 32) (Step SI 60).

According to yet another embodiment of the present disclosure, both 6_modc| ^ar|d H can be fed into the WD autoencoder unit 46 for compression. FIG. 12 illustrates yet another alternate representation of the system 10 architecture for the configuration of the hybrid CSI compression and reporting scheme, where both H and b_modei ^are fed into the WD autoencoder unit 46. In this example, the wireless device 14 performs a channel measurement (Step SI 62), and performs a model-based feature extraction (e.g., as performed by WD model -based processing unit 44) (Step SI 64). Both b_modei ^and H are fed into the AE-based compression step (e.g., as performed by WD autoencoder unit 46) (Step S166). The WD transmits the CSI report to the network node 12 (Step S168), e.g., via connection 20, which performs AE-based decompression and/or AE-based channel reconstruction (e.g., as performed by NW autoencoder unit 32) (Step 170), and performs model-based dimension expansion and/or model-based channel reconstruction (e.g., as performed by NW model -based processing unit 30) (Step SI 72).

According to yet another embodiment of the present disclosure, in a specific case, when the WD model-based processing unit 44, for both feature extraction and channel reconstruction, performs an identity function, i.e., H = ff_model and b_model contains zero bits (accordingly H_modc| ⁼ H), then the system architecture for the configuration of the hybrid CSI compression and reporting scheme reduces to a conventional AE-based CSI compression architecture.

Several embodiments for configuring the model-based dimension-reduction step (e.g., as performed by WD model-based processing unit 44 and/or NW model-based processing unit 30) will be described in accordance with various technical scope as described below:

According to another embodiment of the present disclosure, the model-based dimension-reduction step (e.g., as performed by WD model-based processing unit 44) includes extraction of certain channel features to compress the channel, using a predefined function. The wireless device 14 is configured by the network node 12 (e.g., gNB) via higher layer signalling (e.g., RRC) and a CSI report configuration for the model-based dimension reduction step. The CSI report configuration is extended from previous releases (i.e., before Rel. 18) to include a set of parameters c_model related to the model-based feature extraction procedure. The set of parameters in c_model contains one or more of the following:

The function (■) for the WD model -based processing unit 44.

The features to be extracted from the WD model-based processing unit 44, e.g., spatial beams.

The quantity of each feature to be extracted from the WD model-based processing unit 44, e.g., the number of spatial beams.

Constraints on the CSI report, e.g., avoid report certain spatial beams.

According to yet another embodiment of the present disclosure, the model-based dimension-reduction step (e.g., performed by WD model-based processing unit 44) includes extraction of a number of spatial beams of the channel using a set of selected orthogonal DFT basis vectors. The number of spatial beams selected by the wireless device 14 is configured by the network node 12 via high layer signaling (e.g., RRC) and Indices of the selected DFT basis vectors are further reported to the network node 12 (e.g., gNB).

According to some embodiments, where CSI-RS are beamformed, the DFT basis vectors can instead be a port-selection matrix, which has a single value of “1” per column indicating the selected CSI-RS port, while all “0” elsewhere. Hence, the information encoded into b_modei bits comprise port selection information.

According to yet another embodiment of the present disclosure, the model-based dimension-reduction step (e.g., performed by WD model-based processing unit 44) includes extraction of a number of delay taps of the channel using a set of selected orthogonal basis vectors, such as a DFT basis. The number of delay taps selected by the wireless device 14 is configured by the network node 12 via higher layer signalling (e.g., RRC) and indices of the selected basis vectors are further reported to the network node 12.

According to some embodiments configuring the model-based dimensionreduction step (e.g., performed by WD model-based processing unit 44), a frequency domain (FD) window (or equivalently a number of consecutive delay taps) are selected by the wireless device 14. The length of the FD window (or equivalently the number of consecutive delay taps) selected by the wireless device 14 is configured by the network node 12 via higher layer signalling (e.g., RRC).

According to yet another embodiment configuring the model-based dimensionreduction step, the starting position of the window may not need to be reported, since it may only create a common phase rotation of all the ports, then, b_model ⁼ 0.

According to yet another embodiment, for configuring the model-based dimension-reduction step (e.g., performed by WD model -based processing unit 44), a number of spatial beams and delay taps may be jointly selected by the wireless device 14 and the indices of the selected spatial beams, and the delay taps of the corresponding DFT basis vectors may be reported to the network node 12.

According to yet another embodiment configuring the model-based dimensionreduction step, the channel (or channel feature) compression in SD and/or FD in the model-based dimension reduction step (e.g., performed by WD model-based processing unit 44) may uses some other basis than DFT, e.g., wavelets. The benefit of using such functions could be allowing non-uniform sampling of the spatial domain (SD) and FD, which better fits the characteristics of the propagation channel.

According to yet another embodiment configuring the model-based dimensionreduction step, the model-based dimension reduction step (e.g., performed by WD model -based processing unit 44) could be done using a pre-defined function (■), instead of pre-defined basis functions, e.g., which have been agreed in 3GPP specifications. Such function (■) could be a pre-defined neural network, or one or more pre-defined modelbased functions.

According to yet another embodiment configuring the model-based dimensionreduction step (e.g., performed by WD model-based processing unit 44), the channel representation in the latent space at the output of the encoder could be quantized based on a pre-defined codebook, where the wireless device 14 feeds the index of the quantized latent channel representation back to the network node 12.

According to yet another embodiment configuring the model-based dimensionreduction step (e.g., performed by WD model-based processing unit 44), the network node 12 may configure constraints in the c_model for the model -based processing unit to the WD 14.

According to yet another embodiment configuring the model-based dimensionreduction step (e.g., performed by WD model-based processing unit 44), a codebook subset restriction is configured in c_model. For example, the wireless device 14 may be configured to avoid selecting beams corresponding to certain directions.

According to yet another embodiment configuring the model-based dimensionreduction step (e.g., performed by WD model-based processing unit 44), a rank restriction is configured in the c_model. This may be helpful in terms of reducing the complexity for the autoencoding functionality when network node 12 knows that the higher rank may not be scheduled for certain wireless devices 14.

However, in some cases, where only certain features of the channel are interested in the network node 12. Then, the network node 12 can configure the wireless device 14 to report only those features. According to yet other embodiments configuring the modelbased dimension-reduction step (e.g., performed by WD model-based processing unit 44), the network node 12 configures the wireless device 14 a CSI report configuration, wherein reportQuantity = ‘cri-RI-il-i6’, which means that the wireless device 14 only reports CSI resource indicator (CRI), a rank indicator (RI), selected SD basis vectors (il), and selected FD basis vectors (i6), since only the beam and delay information is needed by the network node 12.

Further, according to the several embodiments of the present disclosure, each of the above parameters for the c_model may be signalled explicitly to the WD 14 or may be determined via some parameter combination, or a mixture of both. When the parameter combination is applied, the corresponding parameters may be determined via the configured index of the parameter combination. Such combination could depend on rank restriction, number of CSI-RS ports, number of SD basis vectors, number of FD basis vectors, etc.

As described above, preferable embodiments of the present disclosure have been described in detail with reference to the accompanying drawings, but the technical scope of the present technology is not limited to these examples. A person skilled in the art may obviously find various alternations and modifications within the technical ideas as set forth in the scope of the appended claims, and it should be understood that they will naturally come under the technical scope of the present disclosure. Also, a few embodiments of the present disclosure can also be combined with any of the other embodiments, where the general architecture is somewhat modified.

The autoencoder compression step (e.g., performed by WD model-based processing unit 44 and/or WD autoencoder unit 46) extracts and labels, in a potentially invertible way, fine channel information (e.g., small-scale fast fading channel coefficients and/or eigen vectors/values) while it may be difficult for the classical model-based approaches to efficiently compress such fine channel details.

In view of the embodiments as described above, the model-based dimensionreduction step (e.g., performed by WD model-based processing unit 44 and/or WD autoencoder unit 46) of the present disclosure provides a plurality of advantages such as:

- It may enable smaller autoencoders (autoencoders with fewer layers and neurons) (e.g., as implemented by WD model-based processing unit 44 and/or WD autoencoder unit 46), by reducing the number of input (and, potentially, output) variables to the autoencoder functionalities (similar reductions may be achieved on the network node 12 side, e.g., NW model-based processing unit 30 and/or NW autoencoder unit 32).

- A smaller autoencoder functionality (e.g., as implemented by WD model-based processing unit 44 and/or WD autoencoder unit 46 on the wireless device 14 side, and/or by NW model-based processing unit 30 and/or NW autoencoder unit 32 on the network node 12 side) simplifies training, retraining, and deployment (fewer hyperparameters and trainable parameters to optimize).

- A smaller autoencoder functionality (e.g., as implemented by WD model-based processing unit 44 and/or WD autoencoder unit 46) is less onerous on the wireless device 14 hardware 31, which positively impacts the wireless device 14 cost.

- A smaller autoencoder functionality (e.g., as implemented by WD model-based processing unit 44 and/or WD autoencoder unit 46) combined with efficient model -based dimension reductions (e.g., DFT transformations) may have smaller latency during inference (e.g., requires less time to compute and report CSI).

- A model-based prepossessing step based on spatial- and frequency-domain DFT codebooks allows the autoencoder functionality (e.g., as implemented by WD model-based processing unit 44 and/or WD autoencoder unit 46) to be agnostic to wireless device 14 implementation-specific details such as Rx-antenna layout, and receiver algorithms (e.g., channel estimation). That is, the autoencoder functionality (e.g., as implemented by WD model-based processing unit 44 and/or WD autoencoder unit 46) input is standardized and consistent across wireless device 14 terminals, which simplifies the autoencoder functionality design, testing, and deployment (e. wireless device 14, chipset and network node 12 vendors may use the same or similar autoencoder functionalities (e.g., as implemented by WD model-based processing unit 44 and/or WD autoencoder unit 46) on wireless devices 14 with different antenna layouts and channel estimation algorithms).

- A model-based prepossessing step (e.g., as implemented by WD model-based processing unit 44 and/or WD autoencoder unit 46) based on spatial and frequency-domain DFT codebooks has less specification impact than a “pure AE” based solution taking a matched filter/channel estimate as input.

- A model-based prepossessing step (e.g., as implemented by WD model-based processing unit 44 and/or WD autoencoder unit 46) based on spatial and frequency-domain DFT codebooks has less specification which may increase the transferability of autoencoder functionalities (e.g., as implemented by WD model-based processing unit 44 and/or WD autoencoder unit 46) between various channel models and propagation environments. FIG. 13 illustrates the advantages of the system architecture described in the present disclosure assuming the example of the system 10 architecture shown in FIG. 8 of the drawings, according to an embodiment of the present disclosure. In particular, FIG. 13 illustrates an example of CDF plotting for Normalized Mean Square Error (NMSE) performance for a hybrid beam-delay domain approach, according to an embodiment of the present disclosure.

In the example shown in FIG. 13 of the drawings, the network node 12 (e.g., gNB) is equipped with a uniform planar array with N_a = 32 antenna ports (4 x 4 array with dual-polarized) with half-wavelength spacing in a vertical and horizontal direction and transmitting over N₃ = 52 subbands (group of subcarriers) to the wireless devices 14 with K = 1 antenna port. Further, assuming that the CSI-RS ports are non- beamformed so that each antenna port can transmit one CSI-RS port. Accordingly, the antenna-frequency domain ‘raw’ channel H have a dimension of 1 X 32 X 52 (K x N_a x A₃). Subsequently, L = 8 spatial beams (per polarization) of the channel, i.e., 8 orthogonal DFT basis vectors per polarization, and M = 12 delay taps, i.e., 12 orthogonal DFT basis vectors, are selected to compress the ‘raw’ channel to the beamdomain channel W_{i odc}| with ^a dimension of 1 X 10 X 12 (K X L X M). The 8 orthogonal DFT basis vectors per polarization for spatial beams and 12 orthogonal DFT basis vectors for delay -taps can be encoded using b_model =

52 bits, where the same spatial beams for both the polarization is considered. Furthermore, it is possible to further reduce the reporting overhead 6_modc|. Accordingly, the beam-delay domain channel W_modc| is given as input to the autoencoder functionality (e.g., as implemented by WD model-based processing unit 44 and/or WD autoencoder unit 46), which generates a channel representation in latent space with dimension 1 x 54, where each element is encoded using 4 bits, generating b_AE = 210 bits at the output of the encoder. Accordingly, the total bits required is 6_modc| + ^AE = 268 bits. The Normalized Mean Square Error (NMSE) performance with beam-delay domain compression is compared to two models where the antenna-frequency ‘raw’ channel H is directly given to the autoencoder functionality (e.g., as implemented by WD modelbased processing unit 44 and/or WD autoencoder unit 46) as input, generating total b_AE = 432 and 352 bits. It can be observed from FIG. 13 that the hybrid model -based dimension reduction helps in achieving better performance (in terms of mean NMSE) and a lower feedback overhead compared to autoencoding with b _E = 352 and b _E = 432 bits.

Further, FIG. 14 and FIG. 15 illustrate a Cell-edge user throughput and a meanuser throughput diagram, respectively, for an urban macro (Uma) system-level simulation at 2GHz, in accordance with an embodiment of the present disclosure. Assuming the network node 12 (e.g., radio interface 18) has 32 antenna ports (4 rows, 4 columns, cross-polarized), and an omnidirectional wireless device 14, explicit CSI assumes perfect channel information at the network node 12 (an upper bound) as shown in FIG. 14 and FIG. 15. Further, as shown in FIG. 14 and FIG. 15, Rel. 10 Type is an NR benchmark and AE1 is based on a convolutional neural network and operates directly on the UE’s channel estimate (perfect) in the antenna-frequency domain. AE2 uses the DFT model-based dimension reduction step described above (the total overhead for AE2 is 210 + 52 = 268 bits, where 52 bits are used for the model -based dimension reduction).

While specific language has been used to describe the present subject matter, any limitations arising on account thereto, are not intended. As would be apparent to a person in the art, various working modifications may be made to the method in order to implement the inventive concept as taught herein. The drawings and the forgoing description give examples of embodiments. Those skilled in the art will appreciate that one or more of the described elements may well be combined into a single functional element. Alternatively, certain elements may be split into multiple functional elements. Elements from one embodiment may be added to another embodiment.

The below table includes a list of abbreviations and acronyms that are being used in the description of the present disclosure:

Abbreviation Explanation

3GPP 3rd Generation Partnership Project

AE Auto Encoder

CE Channel Estimate

DFT Discrete Fourier Transform

ML Machine Learning

MU-MIMO Multi-User Multiple Input, Multiple Output

NMSE Normalized Mean Square Error

PCA Principal Component Analysis

PMI Precoder Matrix Indicator RI Rank Indicator

Al Artificial Intelligence

BS Base station

CIR Channel impulse response

CSI Channel state information

CSI-RS Channel state information reference signal

DC Dual Connectivity

DL Downlink eNB Evolved NodeB

E-UTRAN Evolved Universal Terrestrial Radio Access Network gNB A radio base station in NR.

LTE Long term evolution

MIMO Multiple Input Multiple Output

ML Machine Learning

NR New radio

NW Network

0AM Operation and Maintenance

RAN Radio access network

RL Reinforcement Learning

SINR Signal to interference and noise ratio

SN Secondary node

UE User equipment

UL Uplink

It will be appreciated by persons skilled in the art that the embodiments described herein are not limited to what has been particularly shown and described herein above. In addition, unless mention was made above to the contrary, it should be noted that all of the accompanying drawings are not to scale. A variety of modifications and variations are possible in light of the above teachings without departing from the scope of the following claims.

Claims

What is Claimed:

1. A wireless device (14) configured to communicate with a network node (12), the wireless device (14) being configured with a configuration for feature extraction, the wireless device (14) comprising processing circuitry (36) configured to: perform a first channel measurement on a first plurality of antenna ports; translate the first channel measurement into a second channel measurement based on the configuration for feature extraction; encode the second channel measurement using an autoencoder based on an artificial neural network to generate an encoded channel measurement; and cause transmission to the network node (12) of a first indication of the encoded channel measurement for channel estimation.

2. The wireless device (14) of Claim 1, wherein the translating of the first channel measurement into a second channel measurement based on the configuration for feature extraction includes determining a plurality of extracted features based on the configuration for feature extraction, the second channel measurement being associated with the plurality of extracted features; and the processing circuitry (36) being further configured to cause transmission to the network node (12) of a second indication of the determined plurality of extracted features for channel estimation.

3. The wireless device (14) of any one of Claims 1 and 2, wherein the determining of the plurality of extracted features based on the configuration for feature extraction includes: determining a first plurality of extracted features for a first rank of the first channel measurement, the first rank being associated with a first signal -to-noise (SNR) ratio; and determining a second plurality of extracted features for a second rank of the first channel measurement, the second rank being associated with a second SNR ratio weaker than the first SNR ratio, the second plurality of extracted features being smaller than the first plurality of extracted features based on the second SNR ratio being weaker than the first SNR ratio.

4. The wireless device (14) of any one of Claims 1-3, wherein the translating of the first channel measurement into a second channel measurement based on the configuration for feature extraction includes one or more of: reducing a first number of dimensions of the first channel measurement to a second number of dimensions of the second channel measurement; transforming an antenna-frequency domain of the first channel measurement to a beam-delay domain of the second channel measurement; transforming an antenna-frequency -time domain of the first channel measurement to a beam-delay-doppler domain of the second channel measurement; transforming a first number of antenna ports of the first channel measurement to a second number of feature ports of the second channel measurement, the second number being smaller than the first number; reducing a first number of spatial beams of the first channel measurement to a second number of spatial beams of the second channel measurement, the second number of spatial beams being selected based on a set of orthogonal discrete Fourier transform (DFT) basis vectors; and reducing a first number of delay taps of the first channel measurement to a second number of delay taps of the second channel measurement based on the set of orthogonal discrete Fourier transform (DFT) basis vectors.

5. The wireless device (14) of any one of Claims 1-4, wherein the first indication is transmitted to the network node (12) in an uplink CSI report.

6. The wireless device (14) of any one of Claims 1-5, wherein the configuration for feature extraction indicates one or more of: an indicated plurality of extracted features for channel estimation; at least one first spatial beam for extraction; at least one second spatial beam to be excluded from extraction; a number of spatial beams for extraction; at least one first rank for extraction; at least one second rank to be excluded from extraction; and a number of ranks for extraction.

7. A method implemented in a wireless device (14) configured to communicate with a network node (12), the wireless device (14) being configured with a configuration for feature extraction, the method comprising: performing (Block SI 00) a first channel measurement on a first plurality of antenna ports; translating (Block SI 02) the first channel measurement into a second channel measurement based on the configuration for feature extraction; encoding (Block SI 04) the second channel measurement using an autoencoder based on an artificial neural network to generate an encoded channel measurement; and causing transmission (Block S106) to the network node (12) of a first indication of the encoded channel measurement for channel estimation.

8. The method of Claim 7, wherein the translating of the first channel measurement into a second channel measurement based on the configuration for feature extraction includes determining a plurality of extracted features based on the configuration for feature extraction, the second channel measurement being associated with the plurality of extracted features; and the method further comprising causing transmission to the network node (12) of a second indication of the determined plurality of extracted features for channel estimation.

9. The method of any one of Claims 7 and 8, wherein the determining of the plurality of extracted features based on the configuration for feature extraction includes: determining a first plurality of extracted features for a first rank of the first channel measurement, the first rank being associated with a first signal -to-noise (SNR) ratio; and determining a second plurality of extracted features for a second rank of the first channel measurement, the second rank being associated with a second SNR ratio weaker than the first SNR ratio, the second plurality of extracted features being smaller than the first plurality of extracted features based on the second SNR ratio being weaker than the first SNR ratio.

10. The method of any one of Claims 7-9, wherein the translating of the first channel measurement into a second channel measurement based on the configuration for feature extraction includes one or more of: reducing a first number of dimensions of the first channel measurement to a second number of dimensions of the second channel measurement; transforming an antenna-frequency domain of the first channel measurement to a beam-delay domain of the second channel measurement; transforming an antenna-frequency -time domain of the first channel measurement to a beam-delay-doppler domain of the second channel measurement; transforming a first number of antenna ports of the first channel measurement to a second number of feature ports of the second channel measurement, the second number being smaller than the first number; reducing a first number of spatial beams of the first channel measurement to a second number of spatial beams of the second channel measurement, the second number of spatial beams being selected based on a set of orthogonal discrete Fourier transform (DFT) basis vectors; and reducing a first number of delay taps of the first channel measurement to a second number of delay taps of the second channel measurement based on the set of orthogonal discrete Fourier transform (DFT) basis vectors.

11. The method of any one of Claims 7-10, wherein the first indication is transmitted to the network node (12) in an uplink CSI report.

12. The method of any one of Claims 7-11, wherein the configuration for feature extraction indicates one or more of: an indicated plurality of extracted features for channel estimation; at least one first spatial beam for extraction; at least one second spatial beam to be excluded from extraction; a number of spatial beams for extraction; at least one first rank for extraction; at least one second rank to be excluded from extraction; and a number of ranks for extraction.

13. A network node (12) configured to communicate with a wireless device

(14), the network node (12) comprising processing circuitry (22) configured to: determine a configuration for feature extraction; cause transmission of the configuration for feature extraction to the wireless device (14); responsive to causing transmission of the configuration for feature extraction, receive, from the wireless device (14), a first indication of an encoded channel measurement; decode the encoded channel measurement using an autoencoder based on an artificial neural network to generate a decoded channel measurement; translate the decoded channel measurement into a first channel measurement based on the configuration for feature extraction; and perform at least one network node (12) action based on the first channel measurement.

14. The network node (12) of Claim 13, wherein the at least one network node (12) action includes: determining a downlink precoding matrix based on the first channel measurement; and cause transmission of signaling to the wireless device (14) using the downlink precoding matrix.

15. The network node (12) of any one of Claims 13 and 14, wherein the processing circuitry (22) is further configured to receive from the wireless device (14) a second indication of a determined plurality of extracted features for channel estimation; and the translating of the decoded channel measurement into a first channel measurement based on the configuration for feature extraction being based on the determined plurality of extracted features.

16. The network node (12) of any one of Claims 13-15, wherein the plurality of extracted features includes: a first plurality of extracted features for a first rank of the first channel measurement, the first rank being associated with a first signal -to-noise (SNR) ratio; and a second plurality of extracted features for a second rank of the first channel measurement, the second rank being associated with a second SNR ratio weaker than the first SNR ratio, the second plurality of extracted features being smaller than the first plurality of extracted features based on the second SNR ratio being weaker than the first SNR ratio.

17. The network node (12) of any one of Claims 13-16, wherein the translating of the decoded channel measurement into a first channel measurement based on the configuration for feature extraction includes one or more of increasing a first number of dimensions of the decoded channel measurement to a second number of dimensions of the first channel measurement; transforming a beam-delay domain of the decoded channel measurement to an antenna-frequency domain of the first channel measurement; transforming a beam-delay-doppler domain of the decoded channel measurement to an antenna-frequency -time domain of the first channel measurement; transforming a first number of feature ports of the decoded channel measurement to a second number of antenna ports of the first channel measurement, the first number being smaller than the second number; increasing a first number of spatial beams of the decoded channel measurement to a second number of spatial beams of the first channel measurement based on a set of orthogonal discrete Fourier transform (DFT) basis vectors; and increasing a first number of delay taps of the decoded channel measurement to a second number of delay taps of the first channel measurement based on the set of orthogonal discrete Fourier transform (DFT) basis vectors.

18. The network node (12) of any one of Claims 13-17, wherein the first indication is received from the wireless device (14) in an uplink CSI report.

19. The network node (12) of any one of Claims 13-18, wherein the configuration for feature extraction indicates one or more of an indicated plurality of extracted features for channel estimation; at least one first spatial beam for extraction; at least one second spatial beam to be excluded from extraction; a number of spatial beams for extraction; at least one first rank for extraction; at least one second rank to be excluded from extraction; and a number of ranks for extraction.

20. A method implemented in a network node (12) configured to communicate with a wireless device (14), the method comprising: determining (Block SI 08) a configuration for feature extraction; causing transmission (Block SI 10) of the configuration for feature extraction to the wireless device (14); responsive to causing transmission of the configuration for feature extraction, receiving (Block SI 12), from the wireless device (14), a first indication of an encoded channel measurement; decoding (Block SI 14) the encoded channel measurement using an autoencoder based on an artificial neural network to generate a decoded channel measurement; translating (Block SI 16) the decoded channel measurement into a first channel measurement based on the configuration for feature extraction; and performing (Block SI 18) at least one network node (12) action based on the first channel measurement.

21. The method of Claim 20, wherein the at least one network node (12) action includes: determining a downlink precoding matrix based on the first channel measurement; and causing transmission of signaling to the wireless device (14) using the downlink precoding matrix.

22. The method of any one of Claims 20 and 21, wherein the method further comprises receiving from the wireless device (14) a second indication of a determined plurality of extracted features for channel estimation; and the translating of the decoded channel measurement into a first channel measurement based on the configuration for feature extraction being based on the determined plurality of extracted features.

23. The method of any one of Claims 20-22, wherein the plurality of extracted features includes: a first plurality of extracted features for a first rank of the first channel measurement, the first rank being associated with a first signal -to-noise (SNR) ratio; and a second plurality of extracted features for a second rank of the first channel measurement, the second rank being associated with a second SNR ratio weaker than the first SNR ratio, the second plurality of extracted features being smaller than the first plurality of extracted features based on the second SNR ratio being weaker than the first SNR ratio.

24. The method of any one of Claims 20-23, wherein the translating of the decoded channel measurement into a first channel measurement based on the configuration for feature extraction includes one or more of: increasing a first number of dimensions of the decoded channel measurement to a second number of dimensions of the first channel measurement; transforming a beam-delay domain of the decoded channel measurement to an antenna-frequency domain of the first channel measurement; transforming a beam-delay-doppler domain of the decoded channel measurement to an antenna-frequency -time domain of the first channel measurement; transforming a first number of feature ports of the decoded channel measurement to a second number of antenna ports of the first channel measurement, the first number being smaller than the second number; increasing a first number of spatial beams of the decoded channel measurement to a second number of spatial beams of the first channel measurement based on a set of orthogonal discrete Fourier transform (DFT) basis vectors; and increasing a first number of delay taps of the decoded channel measurement to a second number of delay taps of the first channel measurement based on the set of orthogonal discrete Fourier transform (DFT) basis vectors.

25. The method of any one of Claims 20-24, wherein the first indication is received from the wireless device (14) in an uplink CSI report.

26. The method of any one of Claims 20-25, wherein the configuration for feature extraction indicates one or more of: an indicated plurality of extracted features for channel estimation; at least one first spatial beam for extraction; at least one second spatial beam to be excluded from extraction; a number of spatial beams for extraction; at least one first rank for extraction; at least one second rank to be excluded from extraction; and a number of ranks for extraction.