EP4736335A1

EP4736335A1 - Signal detection for mimo

Info

Publication number: EP4736335A1
Application number: EP23942767.7A
Authority: EP
Inventors: Wen Liang QI; Chen Hui YE; Dani Johannes KORPI
Original assignee: Nokia Solutions and Networks Oy
Current assignee: Nokia Solutions and Networks Oy
Priority date: 2023-06-27
Filing date: 2023-06-27
Publication date: 2026-05-06
Also published as: WO2025000243A9; CN121909606A; WO2025000243A1

Abstract

A method comprises obtaining data representing a signal received at a first device transmitted by a second device and channel estimation between the first and second devices, wherein the signal is transmitted over a plurality of layers; generating, by a first sub-model of a machine learning model, shared feature representations for the plurality of layers based on the data and the channel estimation; and generating, by a second sub-model of the machine learning model based on the shared feature representations, respective predictions on bits for the plurality of layers.

Description

SIGNAL DETECTION FOR MIMO

FIELDS
Various example embodiments of the present disclosure generally relate to the field of telecommunication and in particular, to methods, devices, apparatuses and computer readable storage medium for signal detection for Multiple-Input-Multiple-Output (MIMO) .

BACKGROUND

MIMO refers to the type of wireless transmission and reception scheme where both a transmitter and a receiver employ more than one antenna. MIMO allows for spatial diversity to transmit data by use of a plurality of antennas in both uplink (UL) and downlink (DL) directions. The obtained spatial diversity offers a more efficient utilization of the frequency spectrum. Moreover, MIMO can reduce the inter-cell and intra-cell interferences which in turn, leads to more frequency re-use. Therefore, MIMO scheme with very high spectrum efficiency is an important technology of wireless communication systems.
A receiver based on machine learning, which is also referred to as a machine learning receiver, leverages the customized artificial intelligence (AI) /machine learning (ML) techniques to further boost the data transmission from physical layer perspective and is playing the key role in the forthcoming communication. Currently, the System on Chip (SoC) platform compatible with the machine learning receiver solution is under intensive development. A machine learning receiver for MIMO system has been proposed.

SUMMARY

In a first aspect of the present disclosure, there is provided an apparatus. The apparatus comprises at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to: obtain data representing a signal received at a first device transmitted by a second device and channel estimation between the first and second devices, wherein the signal is transmitted over a plurality of layers; generate, by a first sub-model of a machine learning model, shared feature representations for the plurality of layers based on the data and the channel estimation; and generate, by a second sub-model of the machine learning model based on the shared feature representations, respective predictions on bits for the plurality of layers.
In a second aspect of the present disclosure, there is provided a method. The method comprises: obtaining data representing a signal received at a first device transmitted by a second device and channel estimation between the first and second devices, wherein the signal is transmitted over a plurality of layers; generating, by a first sub-model of a machine learning model, shared feature representations for the plurality of layers based on the data and the channel estimation; and generating, by a second sub-model of the machine learning model based on the shared feature representations, respective predictions on bits for the plurality of layers.
In a third aspect of the present disclosure, there is provided an apparatus. The apparatus comprises means for obtaining data representing a signal received at a first device transmitted by a second device and channel estimation between the first and second devices, wherein the signal is transmitted over a plurality of layers; means for generating, by a first sub-model of a machine learning model, shared feature representations for the plurality of layers based on the data and the channel estimation; and means for generating, by a second sub-model of the machine learning model based on the shared feature representations, respective predictions on bits for the plurality of layers.
In a fourth aspect of the present disclosure, there is provided a computer readable medium. The computer readable medium comprises instructions stored thereon for causing an apparatus to perform at least the method according to the second aspect.
It is to be understood that the Summary section is not intended to identify key or essential features of embodiments of the present disclosure, nor is it intended to be used to limit the scope of the present disclosure. Other features of the present disclosure will become easily comprehensible through the following description.

BRIEF DESCRIPTION OF THE DRAWINGS

Some example embodiments will now be described with reference to the accompanying drawings, where:
FIG. 1 illustrates an example communication environment in which example embodiments of the present disclosure can be implemented;
FIG. 2 illustrates an example of an architecture of a machine learning model for signal detection in MIMO system according to some example embodiments of the present disclosure;
FIG. 3 illustrates an example signal processing workflow by using the machine learning model according to some example embodiments of the present disclosure;
FIG. 4A illustrates an example structure for generating the shared feature representations according to some example embodiments of the present disclosure;
FIG. 4B illustrates another example structure for generating the shared feature representations according to some example embodiments of the present disclosure;
FIG. 5 illustrates another example signal processing workflow by using the machine learning model according to some example embodiments of the present disclosure;
FIG. 6 illustrates an example signaling chart for training of the machine learning model according to some example embodiments of the present disclosure;
FIG. 7 illustrates an example signaling chart for deployment of the machine learning model according to some example embodiments of the present disclosure;
FIG. 8 illustrates the evaluation results of some example embodiments of the present disclosure and another solution;
FIG. 9 illustrates a flowchart of a method implemented at a device according to some example embodiments of the present disclosure;
FIG. 10 illustrates a simplified block diagram of a device that is suitable for implementing example embodiments of the present disclosure; and
FIG. 11 illustrates a block diagram of an example computer readable medium in accordance with some example embodiments of the present disclosure.
Throughout the drawings, the same or similar reference numerals represent the same or similar element.

DETAILED DESCRIPTION

Principle of the present disclosure will now be described with reference to some example embodiments. It is to be understood that these embodiments are described only for the purpose of illustration and help those skilled in the art to understand and implement the present disclosure, without suggesting any limitation as to the scope of the disclosure. Embodiments described herein can be implemented in various manners other than the ones described below.
In the following description and claims, unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skills in the art to which this disclosure belongs.
References in the present disclosure to “one embodiment, ” “an embodiment, ” “an example embodiment, ” and the like indicate that the embodiment described may include a particular feature, structure, or characteristic, but it is not necessary that every embodiment includes the particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same embodiment. Further, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is submitted that it is within the knowledge of one skilled in the art to affect such feature, structure, or characteristic in connection with other embodiments whether or not explicitly described.
It shall be understood that although the terms “first, ” “second, ” …, etc. in front of noun (s) and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another and they do not limit the order of the noun (s) . For example, a first element could be termed a second element, and similarly, a second element could be termed a first element, without departing from the scope of example embodiments. As used herein, the term “and/or” includes any and all combinations of one or more of the listed terms.
As used herein, “at least one of the following: <a list of two or more elements>” and “at least one of <a list of two or more elements>” and similar wording, where the list of two or more elements are joined by “and” or “or” , mean at least any one of the elements, or at least any two or more of the elements, or at least all the elements.
As used herein, unless stated explicitly, performing a step “in response to A” does not indicate that the step is performed immediately after “A” occurs and one or more intervening steps may be included.
The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments. As used herein, the singular forms “a” , “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “comprises” , “comprising” , “has” , “having” , “includes” and/or “including” , when used herein, specify the presence of stated features, elements, and/or components etc., but do not preclude the presence or addition of one or more other features, elements, components and/or combinations thereof.
As used in this application, the term “circuitry” may refer to one or more or all of the following:
(a) hardware-only circuit implementations (such as implementations in only analog and/or digital circuitry) and
(b) combinations of hardware circuits and software, such as (as applicable) :
(i) a combination of analog and/or digital hardware circuit (s) with software/firmware and
(ii) any portions of hardware processor (s) with software (including digital signal processor (s) ) , software, and memory (ies) that work together to cause an apparatus, such as a mobile phone or server, to perform various functions) and
(c) hardware circuit (s) and or processor (s) , such as a microprocessor (s) or a portion of a microprocessor (s) , that requires software (e.g., firmware) for operation, but the software may not be present when it is not needed for operation.
This definition of circuitry applies to all uses of this term in this application, including in any claims. As a further example, as used in this application, the term circuitry also covers an implementation of merely a hardware circuit or processor (or multiple processors) or portion of a hardware circuit or processor and its (or their) accompanying software and/or firmware. The term circuitry also covers, for example and if applicable to the particular claim element, a baseband integrated circuit or processor integrated circuit for a mobile device or a similar integrated circuit in server, a cellular network device, or other computing or network device.
As used herein, the term “communication network” refers to a network following any suitable communication standards, such as New Radio (NR) , Long Term Evolution (LTE) , LTE-Advanced (LTE-A) , Wideband Code Division Multiple Access (WCDMA) , High-Speed Packet Access (HSPA) , Narrow Band Internet of Things (NB-IoT) and so on. Furthermore, the communications between a terminal device and a network device in the communication network may be performed according to any suitable generation communication protocols, including, but not limited to, the first generation (1G) , the second generation (2G) , 2.5G, 2.75G, the third generation (3G) , the fourth generation (4G) , 4.5G, the fifth generation (5G) , the sixth generation (6G) communication protocols, and/or any other protocols either currently known or to be developed in the future. Embodiments of the present disclosure may be applied in various communication systems. Given the rapid development in communications, there will of course also be future type communication technologies and systems with which the present disclosure may be embodied. It should not be seen as limiting the scope of the present disclosure to only the aforementioned system.
As used herein, the term “network device” refers to a node in a communication network via which a terminal device accesses the network and receives services therefrom. The network device may refer to a base station (BS) or an access point (AP) , for example, a node B (NodeB or NB) , an evolved NodeB (eNodeB or eNB) , an NR NB (also referred to as a gNB) , a Remote Radio Unit (RRU) , a radio header (RH) , a remote radio head (RRH) , a relay, an Integrated Access and Backhaul (IAB) node, a low power node such as a femto, a pico, a non-terrestrial network (NTN) or non-ground network device such as a satellite network device, a low earth orbit (LEO) satellite and a geosynchronous earth orbit (GEO) satellite, an aircraft network device, and so forth, depending on the applied terminology and technology. In some example embodiments, radio access network (RAN) split architecture comprises a Centralized Unit (CU) and a Distributed Unit (DU) at an IAB donor node. An IAB node comprises a Mobile Terminal (IAB-MT) part that behaves like a UE toward the parent node, and a DU part of an IAB node behaves like a base station toward the next-hop IAB node.
The term “terminal device” refers to any end device that may be capable of wireless communication. By way of example rather than limitation, a terminal device may also be referred to as a communication device, user equipment (UE) , a Subscriber Station (SS) , a Portable Subscriber Station, a Mobile Station (MS) , or an Access Terminal (AT) . The terminal device may include, but not limited to, a mobile phone, a cellular phone, a smart phone, voice over IP (VoIP) phones, wireless local loop phones, a tablet, a wearable terminal device, a personal digital assistant (PDA) , portable computers, desktop computer, image capture terminal devices such as digital cameras, gaming terminal devices, music storage and playback appliances, vehicle-mounted wireless terminal devices, wireless endpoints, mobile stations, laptop-embedded equipment (LEE) , laptop-mounted equipment (LME) , USB dongles, smart devices, wireless customer-premises equipment (CPE) , an Internet of Things (IoT) device, a watch or other wearable, a head-mounted display (HMD) , a vehicle, a drone, a medical device and applications (e.g., remote surgery) , an industrial device and applications (e.g., a robot and/or other wireless devices operating in an industrial and/or an automated processing chain contexts) , a consumer electronics device, a device operating on commercial and/or industrial wireless networks, and the like. The terminal device may also correspond to a Mobile Termination (MT) part of an IAB node (e.g., a relay node) . In the following description, the terms “terminal device” , “communication device” , “terminal” , “user equipment” and “UE” may be used interchangeably.
As used herein, the term “resource, ” “transmission resource, ” “resource block, ” “physical resource block” (PRB) , “uplink resource, ” or “downlink resource” may refer to any resource for performing a communication, for example, a communication between a terminal device and a network device, such as a resource in time domain, a resource in frequency domain, a resource in space domain, a resource in code domain, or any other combination of the time, frequency, space and/or code domain resource enabling a communication, and the like. In the following, unless explicitly stated, a resource in both frequency domain and time domain will be used as an example of a transmission resource for describing some example embodiments of the present disclosure. It is noted that example embodiments of the present disclosure are equally applicable to other resources in other domains.
As used herein, the terms “ML model” , “AI model” or “AI/ML model” may refer to a data driven algorithm that applies AI/ML techniques to generate a set of outputs based on a set of inputs. In the context of the present disclosure, the term “ML model” may be interchangeably with the terms “model” , “AI model” and “AI/ML model” .
As used herein, the term “conventional approach” or the like may refer to an approach for signal detection without a ML model. For example, the conventional approach may include channel estimation, equalization and de-mapping.
Example Environment
FIG. 1 illustrates an example communication environment 100 in which example embodiments of the present disclosure can be implemented. The communication environment 100 may include a first device 110 and a second device 120, which can communicate with each other.
In some example embodiments, the first device 110 may be a network device (for example, gNB) , and the second device120 may be a terminal device (for example, a UE). The second device 120 may communicate with the first device 110 within a coverage of a cell served by the first device 110. In such example embodiments, a link from the first device 110 to the second device 120 is referred to as a DL, and a link from the second device 120 to the first device 110 is referred to as an UL. In DL, the first device 110 is a transmitting (TX) device (or a transmitter) and the second device 120 is a receiving (RX) device (or a receiver) . In UL, the second device 120 is a TX device (or a transmitter) and the first device 110 is a RX device (or a receiver) . Alternatively, in some example embodiments, the first device 110 may be a terminal device and the second device 120 may be a network device.
As shown in FIG. 1, the second device 120 can transmit a signal to the first device 110 and accordingly the first device 110 can receive the signal from the second device 120. In this case, the second device 120 is a TX device and the first device 110 is the RX device. In the example embodiments of the present disclosure, the signal can be transmitted over any suitable channel and carry any suitable information, including but not limited to user data, control information, etc. Embodiments of the present disclosure is not limited in this regard.
The communication environment 100 may be or include a MIMO system. Accordingly, the first device 110 may be equipped with an antenna system, which includes a plurality of antenna elements. The second device 120 may be also equipped with an antenna system which includes a plurality of antenna elements. The antenna system of the first device 110 may include antenna elements to receive the signal transmitted from the antenna system of the second device 120. The antenna system of the first device 110 may also include antenna elements to transmit the signal to the antenna system of the second device120.
In the following, for the purpose of illustration, some example embodiments are described with the first device 110 operating as a network device and the second device 120 operating as a terminal device. However, in some example embodiments, operations described in connection with a terminal device may be implemented at a terminal device or other device, and operations described in connection with a network device may be implemented at a network device or other device.
Communications in the communication environment 100 may be implemented according to any proper communication protocol (s) , comprising, but not limited to, cellular communication protocols of the first generation (1G) , the second generation (2G) , the third generation (3G) , the fourth generation (4G) , the fifth generation (5G) , the sixth generation (6G) , and the like, wireless local network communication protocols such as Institute for Electrical and Electronics Engineers (IEEE) 802.11 and the like, and/or any other protocols currently known or to be developed in the future. Moreover, the communication may utilize any proper wireless communication technology, comprising but not limited to: Code Division Multiple Access (CDMA) , Frequency Division Multiple Access (FDMA) , Time Division Multiple Access (TDMA) , Frequency Division Duplex (FDD) , Time Division Duplex (TDD) , Multiple-Input Multiple-Output (MIMO) , Orthogonal Frequency Division Multiple (OFDM) , Discrete Fourier Transform spread OFDM (DFT-s-OFDM) and/or any other technologies currently known or to be developed in the future.
It is to be understood that the number of devices and their connections shown in FIG. 1 are only for the purpose of illustration without suggesting any limitation. The communication environment 100 may include any suitable number of devices configured to implement example embodiments of the present disclosure.
As briefly mentioned above, the machine learning receiver for the MIMO system has been proposed. The machine learning receiver for MIMO system is targeting at superior data transmission performance by joint block optimization within a neural network. It could replace the conventional channel estimation, equalization, and demodulation in baseband processing functionalities. With the received symbols and raw channel estimates as input, the machine learning receiver may generate bit Log-Likelihood Ratio (LLRs) which may be further used to recover the transmitted source information after channel decoding.
In some solutions, fully convolutional neural network structured deep learning receiver (which is also referred toas deep receiver for short) for MIMO detection of OFDM waveform includes: a Maximum Ratio Combining (MRC) based deep receiver and a multiplicative transformation based Deep receiver.
The MRC based deep receiver takes in two ports of input data with the first port for received data in frequency-time domain after Fast Fourier Transform (FFT) and the second port for interpolated channel estimation in frequency-time domain. The MRC based deep receiver includes three trainable neural networks, including a first neural network for the received data, a second neural network for the interpolated channel estimation, and a third neural network for outputting the bit LLRs.
The two ports of input data are fed to the first and second neural networks, respectively. The first neural network may extend the virtual stream of the channel, and the second neural network may smooth the received data branch for MRC combiner. The essential point of the MRC based deep receiver is the introduction of the untrainable MRC-based transformation, which is similar as the MRC equalizer in the conventional approach. By embedding such untrainable transformation between the trainable neural networks for handling the input data and the third neural network, the MIMO detection of OFDM waveform can be realized.
The multiplicative transformation based deep receiver includes several trainable neural networks, including but not limited to a first neural network for processing a concatenation of raw channel estimation and received data, and a second neural network for outputting the bit LLRs. The key idea of the multiplicative transformation based deep receiver is the introduction of untrainable multiplication which imitates the matrix multiplication to approximate such operation which may exist in the equalization process. Such an operation is also embedded between the trainable first neural network and second neural network.
In another aspect, Discrete Fourier Transform spread Orthogonal Frequency Division Multiplexing (DFT-s-OFDM) has superior performance in terms of spectral efficiency, the sensitivity to frequency selective fading as well as lower Peak-to-Average Power Ratio (PAPR) , which leads to a more efficiency use of the power and low risk of distortion and interference. The DFT-s-OFDM waveform has been adopted in 5G communication system for uplink transmission and is promising as the carrier frequency may further extends to terahertz communication in the future. Thus, supporting DFT-s-OFDM waveform may be required for communication devices, especially for the base station due to the BS serves as the receiver in uplink transmission.
However, the above mentioned solutions are not compatible with the DFT-s-OFDM waveform, which is the required feature for beyond 5G devices. The root cause of the incompatibility is because the fully convolutional structures may not be able to mimic the Inverse Discrete Fourier Transform (IDFT) de-precoding process, which spans across the whole bandwidth.
Moreover, in some other solutions, signal detection for DFT-s-OFDM waveforms in SISO can be performed, but problems exist when applying it for MIMO detection.
Therefore, a new design for the machine learning receiver is required to be compatible with different waveforms (including DFT-s-OFDM waveforms) and enable both the MIMO transmission.
Example architecture of the machine learning receiver
According to some example embodiments of the present disclosure, there is provided a solution for signal detection for MIMO. An ML model for signal detection may be implemented at the RX device. The RX device obtains data representing a signal received at the RX device transmitted by the TX device and channel estimation between the RX and TX devices. The signal may be transmitted over a plurality of layers, which may be also referred to as transmission layers. In other words, the signal may be transmitted in a MIMO system. A first portion of the ML model may be used to generate shared feature representations based on the obtained data and the channel estimation. A second portion of the ML model may be used to generate predictions on bits for each layer of the plurality of layers based on the shared feature representations. The respective predictions on bits for the plurality of layers may be further used to recover the transmitted information after channel decoding.
In this way, the proposed solution can be compatible with different waveforms, especially with the DFT-s-OFDM waveform and enable the MIMO transmission. Much higher gain over the conventional approach can be achieved and the MIMO transmission capability can be promoted.
Example embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.
Reference is now made to FIG. 2, which shows an example of an architecture of the machine learning model for signal detection in the MIMO system according to some example embodiments of the present disclosure. The machine learning model 200 comprises a first sub-model 210 and a second sub-model 220.
The input data to the first sub-model 210 may comprise two parts, that is, data representing a signal received at the first device 110 transmitted by the second device 120 and channel estimation between the first device 110 and the second device 120. The signal is transmitted over a plurality of layers. In the following, the number of the plurality of layers is represented by N_T. The data representing the received signal is also referred to as the received data. The received data may be of any suitable form, for example, data in frequency-time domain after FFT. The channel estimation may be the raw channel estimation or interpolated channel estimation. Embodiments of the present disclosure are not limited in this regard.
The first sub-model 210 may be used to generate the shared feature representations for the plurality of layers based on the received data and the channel estimation. For example, the first sub-model 210 may be designed to perform initial “channel + equalization” jointly.
The second sub-model 220 may be used to generate respective predictions on bits for the plurality of layers based on the shared feature representations. For example, the second sub-model 220 may output a prediction on bits for each layer k (where k=1, 2, ……., N_T) , such as the prediction on bits for layer 1, the prediction on bits for layer 2, ……., predictions on bits for layer N_T, as shown in the FIG. 2. The prediction on bits for a layer may be of any suitable form, including but not limited to hard decisions or soft decisions of the bits for the layer. As an example, the prediction on bits for the layer k may comprise bit LLRs for the layer k. In the following, some example embodiments are described with respect to the bit LLRs. However, it is to be understood that this is merely as an example without any limitation to the protection scope.
The signal may be transmitted in various waveforms, for example, OFDM, DFT-s-OFDM. In some example embodiments, the signal may be transmitted in a DFT-s-OFDM waveform. Accordingly, the second sub-model 220 may comprise at least one module for IDFT.
There may be different schemes for implementing the second sub-model 220. In some example embodiments, the second sub-model may comprise a plurality of branches and each branch corresponds to a layer of the plurality of layers. The shared feature representations generated by the first sub-model may be fed into each branch to generate the prediction on bits for the layer corresponding to that branch. For example, the shared feature representations may be fed into the k-th branch to generate the prediction on bits for the k-th layer. In the following, such a structure of the ML model is also referred to as a first scheme and a branch may be also referred to as a detection branch.
In other words, the second sub-model 220 may include N_T branches. The N_T branches are employed to learn the layer-specified detection task in parallel for soft bit LLRs detection of all N_T layers. It is to be noted that the neural network structures in different branches may possess different parameter values or weights and may be trained with respective loss computation, as will be described below. For example, the neural network from the 1^st branch is different from that from the N_T-th branch.
The branches may be implemented by any suitable network structure. For example, the specific configuration of the branches may be determined based on the signal detection task at hand.
In some example embodiments, the branch may comprise one or more trainable neural networks. For example, the branch may include a first neural network used to extract, from the shared feature representations, feature representations for the layer corresponding to the branch. In other words, the first neural network in the k-th branch may perform layer feature selection for the k-th layer, and thus may be also referred to as a layer selection network.
Alternatively, or in addition, the branch may include a second neural network used to perform joint channel estimation and equalization specific to the layer corresponding to the branch. In other words, the second neural network in the k-th branch may perform channel smoothing and equalization customized for the k-th layer, and thus may be referred to as a customized smoothing and equalization network.
Alternatively, or in addition, the branch may include a third neural network used to perform demodulation specific to the layer corresponding to the branch. In other words, the second neural network in the k-th branch may perform demodulation for the k-th layer, and thus may be referred to as a customized demodulation network.
In some example embodiments, if the signal has the DFT-s-OFDM waveform, each branch may comprise a module for performing IDFT.
Reference is now made to FIG. 3 to describe an example ML model of the first scheme. FIG. 3 illustrates an example signal processing workflow 300 by using the ML model of the first scheme. As an example, but without any limitation, FIG. 3 shows the case of the DFT-s-OFDM waveform. It is to be understood that similar structures can be applied to detection of another waveform. In general, the ML model in this example may comprise a preliminary block 310 and N_T detection branches for the N_T layers, respectively.
As shown, the input data 301 may comprise the channel estimation and the received data. Hereinafter, the received data may also be referred to received symbols or signals. The channel estimation and the received data are similar as those described with reference to FIG. 2.
The two sources of input data may be constructed in any suitable way. For example, the channel estimation may be obtained from the raw channel estimate from the pilots (e.g., demodulation reference signals) . The interpolation involved in the process of calculating the channel estimation may adopt any methods including linear interpolation, etc.
As an example, the original channel estimation and received symbols may be preprocessed as below. The complexed-valued channel estimation may be originally a 4-dimensional array with the size of N_T×N_R×F×S where N_T represents the number of transmission layers as mentioned above, N_R represents the number of antennas in the receiving device (which is the first device 110 in the example) , F represents the number of subcarriers, and S represents the number of OFDM symbols in one transmission time-interval (TTI) which may be for example 14. To apply the first scheme, the channel estimation array may be reshaped to 3-dimensional array with the size of N_TN_R×F×S for complexed-value, or 2N_TN_R×F×S for real value if real part and imaginary part will be split.
The received symbols are a 3-dimensional array with the size of N_R×F×S, for complexed-value, or 2N_R×F×S for real value if real part and imaginary part will be split.
In this example of FIG. 3, the first sub-model may comprise the preliminary block 310 for generating the shared feature representations for the N_T layers. The preliminary block 310 may be designed to create shared representations and perform “channel smoothing and equalization” jointly based on the input data 301. In other words, joint channel estimation and equalization may be preliminarily performed by using neural network structures.
The formation of the input data 301 may adopt different manners depending on the structure of the preliminary block 310. In some example embodiments, a first manner is adopted. Accordingly, a concatenation of the received data and the channel estimation may be fed into the preliminary block 310 to generate the shared feature representations. In other words, the preliminary block 310 takes in the concatenated channel estimation and received symbols as its input data and outputs the shared feature representations. For instance, a concatenation of the channel estimation and the received symbols in the first dimension may result in the input data dimension of N_R (1+N_T) ×F×S for complex-valued numbers, or 2N_R (1+N_T) ×F×S for real numbers.
In such example embodiments, the preliminary block 310 may comprise trainable neural network structure s. Reference is now made to FIG. 4A. The neural network 400A of FIG. 4A can be considered as an example of the preliminary block 310.
As shown in FIG. 4A, the neural network 400A comprises a convolutional (Conv) layer 410 and a Residual Network (ResNet) block 420 which is repeated for M times. The concatenation of the channel estimation and received symbols may be processed by the Conv-layer 410 firstly. Then, the output of the Conv-layer 410 is fed into the ResNet block 420, which outputs the shared feature representations.
In some example embodiments, the formation of the input data 301 may adopt a second manner, where the channel estimation and received symbols are fed into the preliminary block 310, respectively. Specifically, the received data may be fed into a first neural network of the preliminary block 310 to generate first feature representations for the received data. The channel estimation may be fed into a second neural network of the preliminary block 310 to generate second feature representations for the channel estimation. Then, the first and second feature representations may be transformed to the shared feature representations. For example, the first and second feature representations may be transformed based on one or more untrainable operations such as but not limited to MRC.
For example, the preliminary block 310 may comprise two trainable neural networks to process the channel estimation and received symbols, respectively. The two outputs of the two trainable neural networks are then processed by the untrainable MRC process, yielding the shared feature representations. In this design, the two trainable neural networks take in the channel estimation and received symbols respectively.
Reference is now made to FIG. 4B. The module 400B can be considered as another example of the preliminary block 310. As shown in FIG. 4B, the module 400B may comprise a trainable neural network 430, a trainable neural network 440, and the untrained MRC process 450. The channel estimation is processed by the trainable neural network 430 to generate the feature representations for the channel estimation. Specifically, as shown in the FIG. 4B, the channel estimation is fed to the trainable neural network 430 and is processed by the Conv-layer 431, the ResNet block 432 which is repeated for M1 times and the Conv-layer 433 sequentially. Similarly, the received symbols are processed by the trainable neural network 440 to generate the feature representations for the received symbols. Specifically, the received symbols are fed to the trainable neural network 440 and are processed by the Conv-layer 441, the ResNet block 442 which is repeated for M2 times, and the Conv-layer 443 sequentially.
The outputs of the trainable neural network 430 and trainable neural network 440, which are the feature representations for the channel estimation and the received symbols, are further processed in the untrained MRC process 450. Specifically, the feature representations for the channel estimation are reshaped to an array H at block 451. Then, at block 452, the array H is normalized in the N_R dimension to get an array V. The, at block 453, the array V is further used to compute the array B. Besides, the array H is also used to compute to get the array C at block 454. The feature representations for the received symbols are reshaped to an array y at block 455. At block 456, the arrays B, C, y are together used to compute the array Y, which is the output of the untrained MRC process 450. In this example, the array Y represents the shared feature representations for the N_T transmission layers. It is to be noted that this is an example of untrainable operation embedded without any limitation and any suitable untrainable operation may be adopted.
Two examples for generating the shared feature representations are described above. It is to be noted that FIG. 4A and FIG. 4B are given as examples without any limitation and other structures are possible.
Reference is now made back to FIG. 3. The feeding block 320 may be used to feed the shared feature representations to each detection branch, which is an untrainable operation. For example, the shared feature representations from the preliminary block 310 may be duplicated and be fed to all the detection branches dedicated for respective layers.
In this example, the second sub-model comprises N_T branches with the k-th detection branch corresponding to the k-th layer. For example, the detection branch 321 corresponds to the layer 1 and the detection branch 322 corresponds to the layer N_T. Now take the detection branch 321 for layer 1 as an example to illustrate the example structure of the detection branch. However, it is noted that other detection branches may have the same or similar structure but with possibly different parameter values or weights.
As shown, a layer selection network 330 may be used to extract the feature representations for the corresponding layer from the shared feature representations. The layer selection network 330 may refer to a neural network structure with the layer selection functionality. The layer selection network 330 may be used to select and attenuate the features from the feeding block 320 for the dedicated detection of corresponding layer. The neural network structure for the layer selection network 330 may employ convolutional based structures as an example.
The customized smoothing and equalization network 340 may be used to perform joint channel estimation and equalization specific to the corresponding layer. For example, the customized smoothing and equalization network 340 may refer to a neural network structure, which further optimize the joint channel estimation and equalization customized for the dedicated transmission layer. The neural network structure for the customized smoothing and equalization network 340 may be a convolutional based structure as an example. In some example embodiments, a few ResNet blocks may be employed.
The IDFT block 350 may be used to perform IDFT on the output of the customized smoothing and equalization network 340. The IDFT block 350 may refer to the untrainable IDFT operation, which perform the flexible IDFT de-precoding (size of number of subcarriers for data transmission used) only along the frequency dimension, thus enabling the detection of DFT-s-OFDM waveform. It is noted that the IDFT block 350 may be omitted in the case of waveforms other than DFT-s-OFDM.
The intermediate output 360 of the IDFT block 350 may provide the non-traditional constellation like pattern for monitoring. For different layers, the intermediate output of the non-traditional constellation may have different patterns, which depend on the corresponding propagation condition. Thus, the detection branches cannot be shared between the different layers.
The customized demodulation network 370 may be used to perform demodulation specific to the layer corresponding to the branch. For example, the customized demodulation network 370 may refer to a neural network structure that performs the demodulation detection customized for the corresponding layer. The output data dimension of the customized demodulation network 370 may be N_b×F×S for one layer, where N_b represents the number of bits carried in a resource element and depends on the modulation scheme. For example, in the case of 64 Quadrature Amplitude Modulation (QAM) , N_b=6. The neural network structure for the customized demodulation network 370 may be convolutional based structure as an example. A few ResNet blocks may be employed.
The customized demodulation network 370 may output the bit LLRs 380 for layer 1, which may be an example of the prediction on bits for the corresponding layer. Similarly, the bit LLRs for other layers may be output by the customized demodulation networks of corresponding detection branches.
In some example embodiments, in the case of model inference, the bit LLRs for all the transmission layers may be further processed to recover the transmission bits at the TX device. For example, the bit LLRs for all the N_T transmission layers may be utilized to perform the layer combining, descrambling and channel decoding sequentially, thus recovering all the transmission bits.
In some example embodiments, in the case of model training, the bit LLRs for all the transmission layers may be used for model update. In such example embodiments, the received data as input to the ML model may be part of a training sample for the ML model and the training sample may further comprise respective transmitted bit sequences for the plurality of layers of the signal, which can be considered as the ground truth for training.
For example, the received data may be the uncoded bits that are transparent to the receiver for computation of loss in the model training. The total dimension of the data may be N_T×N_b×F×S. For training, the data may be divided into Nx groups according to the transmission layers with each group data size of N_b×F×S, thus for weight update of the corresponding branch.
In the example embodiments of the first scheme, the plurality of detection branches may be updated individually or separately from each other. For example, for a specific layer of the plurality of layers, a loss may be determined based on a difference between the prediction on bits for the specific layer and the transmitted bit sequence for the specific layer (that is the ground truth) . Then, one or more neural networks of the branch corresponding to the specific layer may be updated based on the loss.
For example, as shown in FIG. 3, at block 390, the loss for the layer 1 is computed. The loss for the layer 1 is then used for weight update of the layer selection network 330, the customized smoothing and equalization network 340 and the customized demodulation network 370 in the detection branch. Similarly, the losses for other layers may be computed and used to update the neural networks in corresponding detection branches.
In such example embodiments, the first sub-model may be updated based on a sum of respective losses determined for the plurality of layers. For example, as shown in FIG. 3, at block 391, the summation of the losses from all the layers is computed. The summation of the losses is then used for model weight update of the preliminary block 310.
In the first scheme, the plurality of detection branches is dedicated for respective transmission layers. In this way, a higher gain can be achieved.
Reference is now made back to FIG. 2. In some example embodiments, a scheme 2 may be used to implement the second sub-model 220. Specifically, the shared feature representations from the first sub-model 210 may be divided into a plurality of groups of feature representations, with each group corresponding to a layer of the plurality of layers. Then, each group of feature representations may be fed into the second sub-model 220 to generate the prediction on bits for the corresponding layer.
In such example embodiments, the second sub-model 220 is implemented by using a detection branch common to the plurality of layers. In the other words, the first sub-model 210 is designed to generate shared feature representations and to perform both the initial “channel smoothing and equalization” . Then, the generated feature representations may be split into N_T groups. These N_T groups are inputted into a detection model that shares the same set of parameters (referred to as a Siamese model) for the N_T layers. In contrast to the first scheme, the Siamese model allows for the common representation of propagation conditions across each detection layer to be learned. It should be noted that a common detection branch is utilized for signal detections of all the transmission layers.
In such example embodiments, the second sub-model 220 or the common detection branch may be implemented by any suitable network structure. In some example embodiments, the common detection branch may comprise one or more trainable neural networks. For example, the common detection branch may comprise a fourth neural network used to perform joint channel estimation and equalization common to the plurality of layers. In other words, the fourth neural network may perform smoothing and equalization averaged for all the layers, and thus may be also referred to as an averaged smoothing and equalization network.
Alternatively, or in addition, the common detection branch may include a fifth neural network used to perform demodulation common to the plurality of layers. In other words, the fifth neural network may perform demodulation averaged for all the transmission layers, and thus may be referred to as an averaged demodulation network.
In some example embodiments, if the signal has the DFT-s-OFDM waveform, the common detection branch may comprise a module for performing IDFT.
Reference is now made to FIG. 5 to describe an example ML model of the second scheme. FIG. 5 illustrates an example signal processing workflow 500 by using the ML model of the second scheme. As an example, but without any limitation, FIG. 5 shows the case of the DFT-s-OFDM waveform. It is to be understood that similar structures can be applied to detection of another waveform. In general, the ML model in this example may comprise a preliminary block 510 and one detection branch common for the N_T layers.
As shown, the input data 501 may comprise the channel estimation and the received symbols. The input data 501 and its construction are similar as those described with reference to the input data 301, and thus descriptions thereof are not repeated here.
In this example of FIG. 5, the first sub-model may comprise the preliminary block 510 for generating the shared feature representations for the N_T layers. In this way, joint channel estimation and equalization may be preliminarily performed by using neural network structures. The specific implementation of the preliminary block 510 is similar to that of the preliminary block 310 with the difference that the output data dimension from the preliminary block 510 scales up with the transmission layers N_T so that it could be divided to N_T groups.
The splitting block 520 may be used to perform an untrainable operation that splits the shared feature representations from the preliminary block 510 into N_T groups, each of which is to be utilized for signal detection of one of the N_T layers. For example, if it is assumed that the output data dimension from the preliminary block 510 is a×F×S, then the size of each group should be a/N_T×F×S, where a should be a real number that can be divided by N_T.
The layer selection block 530 may be used to perform the untrainable operation that selects one group of feature representations from the N_T groups to feed to the common detection branch. Such operation will be repeatedly performed on the N_t groups of feature representations using the same shared detection branch for detection of all layers.
The averaged smoothing and equalization network 540 may be used to perform joint channel estimation and equalization shared for all the transmission layers. For example, the averaged smoothing and equalization network 540 may refer to a neural network structure, which further optimizes the joint channel estimation and equalization for averaged propagation conditions across the layers. The neural network structure for the averaged smoothing and equalization network 540 can be a convolutional based structure as an example. A few ResNet blocks may be employed. It should be noted that the design of the neural network structure should comply with an input data dimension of a/N_T×F×S.
The IDFT block 550 may be used to perform IDFT on the output of the averaged smoothing and equalization network 540. The IDFT block 550 is the untrainable IDFT operation similar as the IDFT block 350 and thus descriptions thereof will not be repeated here.
The intermediate output 560 of the IDFT block 550 may be a constellation like pattern which may be the consequence of the averaged channel condition across all the layers.
The averaged demodulation network 570 may be used to perform the demodulation for all the transmission layers. For example, the customized demodulation network 570 may refer to a neural network structure that performs the demodulation detection for all the layers. The implementation of the averaged demodulation network 570 is similar as that of the customized demodulation network 370 and thus descriptions thereof are not repeated here.
The averaged demodulation network 570 may output the bit LLRs 580 for layer 1. Then, the averaged smoothing and equalization network 540, the IDFT block 550, the averaged demodulation network 570 may be repeatedly applied for all the groups and the bit LLRs are generated for the corresponding layers.
In some example embodiments, in the case of model inference, the bit LLRs for all the transmission layers may be further processed to recover the transmission bits at the TX device. For example, the bit LLRs for all the N_T transmission layers may be utilized to perform the layer combining, descrambling and channel decoding sequentially, thus recovering all the transmission bits.
In some example embodiments, in the case of model training, the bit LLRs for all the transmission layers may be used for model update. In such example embodiments, the received data as input to the ML model may be part of a training sample for the ML model and the training sample may further comprise respective transmitted bit sequences for the plurality of layers of the signal, which can be considered as the ground truth for training. In the example embodiments of the second scheme, for a layer of the plurality of layers, a loss may be determined based on a difference between the prediction on bits for the layer and the transmitted bit sequence for the layer. Then, one or more neural networks of the second sub-model may be updated based on respective losses determined for the plurality of layers.
For example, as shown in FIG. 5, at block 590, the loss for the layer 1 is computed. Similarly, the losses may be computed for corresponding layers. These losses for the transmission layers may be utilized for neural network update and impacts on the model weight update of Block550 and Block580. Losses from different layers will update the averaged smoothing and equalization network 540 and the averaged demodulation network 570 iteratively.
In such example embodiments, the first sub-model 210 may be updated based on a sum of respective losses determined for the plurality of layers. For example, as shown in FIG. 5, at block 591, the summation of the losses from all the layers is computed. The summation of the losses is then used for training the preliminary block 510 averagely.
Any type of losses may be used to compute the losses. As an example, but without any limitation, the binary cross entropy may be used as the loss function for model training as shown in the following formula (1) :
where y_i represent binary bits 0/1 and P (y_i) is the corresponding probability which is computed by feeding the output LLRs to a sigmoid function. N is the total number of bits after channel encoding that is computed by N= N_T×N_b×F×S-RE_n where RE_n represents the resource elements not occupied by data. Therefore, non-data positions are ignored during the backpropagation process.
Two schemes are described above as examples. It is to be understood that variations to these schemes are possible. For example, the layer selection network in each detection branch of the first scheme (such as the layer selection network 330) may be removed from the detection branch and a splitting block (similar as the splitting block 520) and a layer selection block (similar as the layer selection block 530) may replace the feeding block 320. For another example, the trainable neural network may include non-trainable components or layers, such as conventional equalizer blocks.
Example embodiments regarding model training and deployment are now described.
In some example embodiments, the model training may be implemented at a training entity for the ML model. The training entity may obtain the received data and the channel estimation from the first device. For example, the training entity may be Local Data Pool and Computing Unit (LDP &CU) .
Reference is now made to FIG. 6, which illustrates an example signaling chart 600 for training of the ML model according to some example embodiments of the present disclosure. The signaling chart 600 involves the first device 110, the second device 120 and the training entity 601.
At 610, the first device 110 as the RX device triggers the online training. At 620, the second device 120 responses with acknowledges of the training request.
At 630, the first device 110 sends the configurations to the second device120. The configurations may indicate or include for example carrier frequency, the number of subcarriers, Modulation and Coding Scheme (MCS) , the rank, transform precoding enabling indication (DFT precoding) , a sequence random seed, the number of TTI required for training, etc.
The DFT-precoding enabling information may indicate the utilization of DFT-s-OFDM waveform and may enable selection of an ML model including a module for performing IDFT.
The number of subcarriers may provide the size of flexible IDFT de-precoding.
The rank number may provide the number of layers for transmitted data and may impact the selection of the corresponding models. Specifically, the rank number corresponds to the number of the transmission layers, which is denoted as N_T above.
The modulation scheme in the MCS may impact the selection of neural network structure for the demodulation network, for example the customized demodulation network 370 or the averaged demodulation network 570. Specifically, the modulation scheme may correspond to the number of bits carried in the resource element, which is denoted as N_b.
Continuing with the chart 600, at 640, the second device 120 prepares to send a sequence according to the received configurations. At 650, the second device 120 sends the sequence over the air to the first device 110. The training data should be known to the first device 110 for model training and update.
At 660, the first device 110 forms the training data from the received sequence. At 670, the first device 110 sends the training data to the training entity 601 repeatedly. The first device 110 may also notify which model (for which rank, which modulation scheme) should be updated.
At 680, the corresponding model is trained in the training entity 601 with training data. For example, the training entity is the LDP &CU. Then, the corresponding model is trained in CU with training data in LDP.
At 690, the training entity 604 sends the trained model to the first device 110. At 691, the first device 110 deploys the model.
Reference is now made to FIG. 7, which illustrates an example signaling chart 600 for deployment of the ML model according to some example embodiments of the present disclosure. The signaling chart 700 involves the first device 110, and the second device 120.
At 710, the second device 120 sends to the first device 110 a configuration for the signal to be transmitted. The configuration may at least comprise: a rank for determining the number of the plurality of layers, and a modulation scheme (for example, the MCS) for determining the number of the bits for a layer of the plurality of layers. Additionally, the configuration may further comprise at least one of: the number of subcarriers for transmitting the signal, or whether transform-precoding is enabled for the signal (for example, transform-precoding enabling information) . In some example embodiments, if the first device 110 is a network device and the second device 120 is a terminal device served by the network device, sending of the configuration may be omitted.
In some example embodiments, the first device 110 may comprise a radio resource management (RRM) plane and a user plane as shown in FIG. 7. In such example embodiments, at 710, the second device 120 may send the configuration to the RRM plane of the first device 110.
At 720, the scheduling of the ML model may be performed by the RRM plane scheduler. The information used by the RRM plane scheduler may include but not limited to the DFT-precoding enabling information which indicates the utilization of DFT-s-OFDM waveform and may enable selection of the ML model in the first device 110; the number of subcarriers for providing the size of flexible IDFT de-precoding; the rank number for providing the number of layers for transmitted data and which may impacts the selection of the corresponding models; the modulation scheme in the MCS which may impact the selection of neural network structure for the demodulation network, for example the customized demodulation network 370 or the averaged demodulation network 570; the fallback information for determining if the first device 110 should fall back to conventional approach or apply the ML model.
At 730, the user plane of the first device 110 may determine if the ML model can be applied. At 740, the second device 120 sends transmission data to the first device 110.
If the ML model can be applied, the chart 700 proceeds to 750. At 750, the first device 110 deploys the ML model to process the data with the model. The first device 110 selects the ML model from a plurality of candidate ML models based on the configuration of the signal.
If the model cannot be applied, the chart 700 proceeds to 760. At 760, the first device 110 uses the conventional approach.
In some example embodiments, the first device 110 may perform model switch between OFDM and DFT-s-OFDM waveforms. For example, the scheduler may notify if the model for DFT-s-OFDM waveform should be activated. Upon the activation of the ML model, the scheduler may determine the proper neural network structure according to the rank, and modulation scheme. If DFT-precoding is enabled, the model for DFT-s-OFDM waveform is applied. If not, the model for another waveform (OFDM, etc. ) is applied.
In some example embodiments, the first device 110 may fall back to conventional approach. For example, the scheduler may determine if the first device 110 should fall back to the conventional approach or not.
If falling back to conventional approach, the first device 110 applies conventional approach to signal detection. If not falling back to conventional approach, the first device 110 deploys the ML model for signal detection.
The fallback of the ML model shall be triggered by the RX device under certain conditions. The triggering of fallback may also trigger the model updates.
To clarify the beneficial effects of the present disclosure, the following evaluation on real-world data has been made. Table 1 shows a list of configurations in the evaluation.
Table 1
The structures of machine learning model are evaluated with data pre-collected from a real-world setup for wireless transmission. The setup is transmitting 2-layer data with 2 transmitting antennas and 2 receiving antennas. The hardware nonlinearity exists, and the DFT-precoding is enabled in data collection phase in order to lower the PAPR in transmission. The detailed configuration can be found in Table 1.
The evaluation results of conventional algorithm are presented by curve 801 and curve 805 in Fig. 8. The evaluation results for a combination of the first scheme and first manner are presented by the curves 804 and 808. The evaluation results for a combination of the second scheme and first manner are presented by the curves 802 and 806. The evaluation results for a combination of the first scheme and second manner are presented by the curves 803 and 807.
The benchmark performance of conventional algorithm is provided by using practical Linear Minimum Mean Squared Error (LMMSE) receiver where Least Squares (LS) channel estimation and Minimum Mean Square Error (MMSE) equalization is employed.
Please be noted that the training and testing data are only collected at the highest Signal-to-Noise Ratio (SNR) point. For other SNR points, the Bit Error Rate (BER) is obtained based on the same testing data that Gaussian noise is added to.
It should be noted that the results are obtained on the testing data set which is not employed in training model. This indicates that the approach exhibits impressive performance under untrained conditions. The benchmark result represented by the curve801 shows that the conventional algorithm is not capable of detecting the data signal under such critical conditions. However, when the proposed approaches are applied, performance is improved, which can be clearly observed from both uncoded BER and coded BER curves. From the 1%BER perspective, it may be estimated that the best performance gain from the proposed scheme can be higher than 2 dB (the best coded BER for conventional algorithm is about 5-6%and cannot reach 1%, thus the gain is estimated) .
It can be seen that respective detection-branch structure outperforms the shared detection-branch structure since the former one may learn the specified transmitting conditions for each layer. For transmitting channel where higher crosstalk exists, MRC based approach may have the chance to provide better performance gain compared with the other scheme. However, performance gain is observed compared with conventional approach by using all the proposed scheme.
Example Methods
FIG. 9 shows a flowchart of an example method 900 implemented at a device in accordance with some example embodiments of the present disclosure. In some example embodiments, the method 900 may be implemented at the first device 110. Alternatively, the method 900 may be implemented at a training entity.
At block 910, data representing a signal received at a first device transmitted by a second device and channel estimation between the first and second devices is obtained. The signal is transmitted over a plurality of layers.
At block 920, by a first sub-model of a machine learning model, shared feature representations for the plurality of layers are generated based on the data and the channel estimation.
At block 930, by a second sub-model of the machine learning model, respective predictions on bits for the plurality of layers are generated based on the shared feature representations.
In some example embodiments, the signal is transmitted in a DFT-s-OFDM waveform and the second sub-model comprises at least one module for Inverse Discrete Fourier Transform.
In some example embodiments, the second sub-model comprises a plurality of branches and a branch of the plurality of branches corresponds to a layer of the plurality of layers. The method 900 comprises feeding the shared feature representations into each branch of the plurality of branches to generate the prediction on bits for the corresponding layer.
In some example embodiments, the branch of the plurality of branches comprises: a first neural network used to extract, from the shared feature representations, feature representations for the layer corresponding to the branch; a second neural network used to perform joint channel estimation and equalization specific to the layer corresponding to the branch; and a third neural network used to perform demodulation specific to the layer corresponding to the branch.
In some example embodiments, the data is part of a training sample for the machine learning model and the training sample further comprises respective transmitted bit sequences for the plurality of layers of the signal. The method 900 comprises for a layer of the plurality of layers, determining a loss based on a difference between the prediction on bits for the layer and the transmitted bit sequence for the layer; updating one or more neural networks of the corresponding branch based on the loss; and updating the first sub-model based on a sum of respective losses determined for the plurality of layers.
In some example embodiments, the method 900 further comprises: dividing the shared feature representations into a plurality of groups of feature representations, a group of the plurality of groups corresponding to a layer of the plurality of layers; and feeding each group of feature representations into the second sub-model to generate the prediction on bits for the corresponding layer.
In some example embodiments, the second sub-model comprises: a fourth neural network used to perform joint channel estimation and equalization common to the plurality of layers; and a fifth neural network used to perform demodulation common to the plurality of layers.
In some example embodiments, the data is part of a training sample for the machine learning model and the training sample further comprises respective transmitted bit sequences for the plurality of layers of the signal The method 900 comprises for a layer of the plurality of layers, determining a loss based on a difference between the prediction on bits for the layer and the transmitted bit sequence for the layer; updating one or more neural networks of the second sub-model based on respective losses determined for the plurality of layers; and updating the first sub-model based on a sum of the respective losses.
In some example embodiments, the method 900 further comprises: feeding a concatenation of the data and the channel estimation into a neural network of the first sub-model to generate the shared feature representations.
In some example embodiments, the method 900 further comprises: feeding to the data into a first neural network of the first sub-model to generate first feature representations for the data; feeding the channel estimation into a second neural network of the first sub-model to generate second feature representations for the channel estimation; and transforming the first and second feature representations to the shared feature representation.
In some example embodiments, the method 900 is implemented at a training entity for the machine learning model, and the data and the channel estimation are obtained from the first device for training the machine learning model.
In some example embodiments, the method 900 further comprises: transmitting the trained machine learning model to the first device.
In some example embodiments, the method 900 is implemented at the first device, and the method 900 further comprises selecting the machine learning model from a plurality of candidate machine learning models based on a configuration of the signal, wherein the configuration at least comprises: a rank for determining the number of the plurality of layers, and a modulation scheme for determining the number of the bits for a layer of the plurality of layers.
In some example embodiments, the configuration further comprises at least one of:the number of subcarriers for transmitting the signal, or whether transform-precoding is enabled for the signal.
In some example embodiments, the method 900 further comprises: replacing the machine learning model by a non-machine learning approach to decode a further signal.
In some example embodiments, the method 900 further comprises: in response to the replacement of the machine learning model by the non-machine learning approach, updating the machine learning model.
Example Apparatus, Device and Medium
In some example embodiments, an apparatus capable of performing any of the method 900 (for example, the first device 110 in FIG. 1 or the training entity 601) may comprise means for performing the respective operations of the method 900. The means may be implemented in any suitable form. For example, the means may be implemented in a circuitry or software module. The apparatus may be implemented as or included in the first device 110 in FIG. 1 or the training entity 601.
In some example embodiments, the apparatus comprises means for obtaining data representing a signal received at a first device transmitted by a second device and channel estimation between the first and second devices, wherein the signal is transmitted over a plurality of layers; means for generating, by a first sub-model of a machine learning model, shared feature representations for the plurality of layers based on the data and the channel estimation; and means for generating, by a second sub-model of the machine learning model based on the shared feature representations, respective predictions on bits for the plurality of layers.
In some example embodiments, the signal is transmitted in a Discrete Fourier Transform-Spread-Orthogonal Frequency Division Multiplexing waveform and the second sub-model comprises at least one module for Inverse Discrete Fourier Transform.
In some example embodiments, the second sub-model comprises a plurality of branches and a branch of the plurality of branches corresponds to a layer of the plurality of layers, and the apparatus further comprises means for feeding the shared feature representations into each branch of the plurality of branches to generate the prediction on bits for the corresponding layer.
In some example embodiments, the branch of the plurality of branches comprises: a first neural network used to extract, from the shared feature representations, feature representations for the layer corresponding to the branch; a second neural network used to perform joint channel estimation and equalization specific to the layer corresponding to the branch; and a third neural network used to perform demodulation specific to the layer corresponding to the branch.
In some example embodiments, the data is part of a training sample for the machine learning model and the training sample further comprises respective transmitted bit sequences for the plurality of layers of the signal, and the apparatus further comprises for a layer of the plurality of layers, means for determining a loss based on a difference between the prediction on bits for the layer and the transmitted bit sequence for the layer; means for updating one or more neural networks of the corresponding branch based on the loss; and means for updating the first sub-model based on a sum of respective losses determined for the plurality of layers.
In some example embodiments, the apparatus further comprises: means for dividing the shared feature representations into a plurality of groups of feature representations, a group of the plurality of groups corresponding to a layer of the plurality of layers; and means for feeding each group of feature representations into the second sub-model to generate the prediction on bits for the corresponding layer.
In some example embodiments, the second sub-model comprises: a fourth neural network used to perform joint channel estimation and equalization common to the plurality of layers; and a fifth neural network used to perform demodulation common to the plurality of layers.
In some example embodiments, the data is part of a training sample for the machine learning model and the training sample further comprises respective transmitted bit sequences for the plurality of layers of the signal, and the apparatus further comprises, for a layer of the plurality of layers, means for determining a loss based on a difference between the prediction on bits for the layer and the transmitted bit sequence for the layer; means for updating one or more neural networks of the second sub-model based on respective losses determined for the plurality of layers; and means for updating the first sub-model based on a sum of the respective losses.
In some example embodiments, the apparatus further comprises: means for feeding a concatenation of the data and the channel estimation into a neural network of the first sub-model to generate the shared feature representations.
In some example embodiments, the apparatus further comprises: means for feeding to the data into a first neural network of the first sub-model to generate first feature representations for the data; means for feeding the channel estimation into a second neural network of the first sub-model to generate second feature representations for the channel estimation; and means for transforming the first and second feature representations to the shared feature representation.
In some example embodiments, the apparatus comprises a training entity for the machine learning model, and the data and the channel estimation are obtained from the first device for training the machine learning model.
In some example embodiments, the apparatus further comprises: means for transmitting the trained machine learning model to the first device.
In some example embodiments, the apparatus comprises the first device, and the apparatus further comprises means for selecting the machine learning model from a plurality of candidate machine learning models based on a configuration of the signal, wherein the configuration at least comprises: a rank for determining the number of the plurality of layers, and a modulation scheme for determining the number of the bits for a layer of the plurality of layers.
In some example embodiments, the configuration further comprises at least one of: the number of subcarriers for transmitting the signal, or whether transform-precoding is enabled for the signal.
In some example embodiments, the apparatus further comprises: means for replacing the machine learning model by a non-machine learning approach to decode a further signal.
In some example embodiments, the apparatus further comprises: means for in response to the replacement of the machine learning model by the non-machine learning approach, updating the machine learning model.
In some example embodiments, the apparatus further comprises means for performing other operations in some example embodiments of the method 900 or the first device 110 or the training entity 601. In some example embodiments, the means comprises at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the performance of the apparatus.
FIG. 10 is a simplified block diagram of a device 1000 that is suitable for implementing example embodiments of the present disclosure. The device 1000 may be provided to implement a communication device, for example, the first device 110 or the second device 120 as shown in FIG. 1, or the training entity 601. As shown, the device 1000 includes one or more processors 1010, one or more memories 1020 coupled to the processor 1010, and one or more communication modules 1040 coupled to the processor 1010.
The communication module 1040 is for bidirectional communications. The communication module 1040 has one or more communication interfaces to facilitate communication with one or more other modules or devices. The communication interfaces may represent any interface that is necessary for communication with other network elements. In some example embodiments, the communication module 1040 may include at least one antenna.
The processor 1010 may be of any type suitable to the local technical network and may include one or more of the following: general purpose computers, special purpose computers, microprocessors, digital signal processors (DSPs) and processors based on multicore processor architecture, as non-limiting examples. The device 1000 may have multiple processors, such as an application specific integrated circuit chip that is slaved in time to a clock which synchronizes the main processor.
The memory 1020 may include one or more non-volatile memories and one or more volatile memories. Examples of the non-volatile memories include, but are not limited to, a Read Only Memory (ROM) 1024, an electrically programmable read only memory (EPROM) , a flash memory, a hard disk, a compact disc (CD) , a digital video disk (DVD) , an optical disk, a laser disk, and other magnetic storage and/or optical storage. Examples of the volatile memories include, but are not limited to, a random access memory (RAM) 1022 and other volatile memories that will not last in the power-down duration.
A computer program 1030 includes computer executable instructions that are executed by the associated processor 1010. The instructions of the program 1030 may include instructions for performing operations/acts of some example embodiments of the present disclosure. The program 1030 may be stored in the memory, e.g., the ROM 1024. The processor 1010 may perform any suitable actions and processing by loading the program 1030 into the RAM 1022.
The example embodiments of the present disclosure may be implemented by means of the program 1030 so that the device 1000 may perform any process of the disclosure as discussed with reference to FIG. 2 to FIG. 9. The example embodiments of the present disclosure may also be implemented by hardware or by a combination of software and hardware.
In some example embodiments, the program 1030 may be tangibly contained in a computer readable medium which may be included in the device 1000 (such as in the memory 1020) or other storage devices that are accessible by the device 1000. The device 1000 may load the program 1030 from the computer readable medium to the RAM 1022 for execution. In some example embodiments, the computer readable medium may include any types of non-transitory storage medium, such as ROM, EPROM, a flash memory, a hard disk, CD, DVD, and the like. The term “non-transitory, ” as used herein, is a limitation of the medium itself (i.e., tangible, not a signal) as opposed to a limitation on data storage persistency (e.g., RAM vs. ROM) .
FIG. 11 shows an example of the computer readable medium 1100 which may be in form of CD, DVD or other optical storage disk. The computer readable medium 1100 has the program 1030 stored thereon.
Generally, various embodiments of the present disclosure may be implemented in hardware or special purpose circuits, software, logic or any combination thereof. Some aspects may be implemented in hardware, and other aspects may be implemented in firmware or software which may be executed by a controller, microprocessor or other computing device. Although various aspects of embodiments of the present disclosure are illustrated and described as block diagrams, flowcharts, or using some other pictorial representations, it is to be understood that the block, apparatus, system, technique or method described herein may be implemented in, as non-limiting examples, hardware, software, firmware, special purpose circuits or logic, general purpose hardware or controller or other computing devices, or some combination thereof.
Some example embodiments of the present disclosure also provide at least one computer program product tangibly stored on a computer readable medium, such as a non-transitory computer readable medium. The computer program product includes computer-executable instructions, such as those included in program modules, being executed in a device on a target physical or virtual processor, to carry out any of the methods as described above. Generally, program modules include routines, programs, libraries, objects, classes, components, data structures, or the like that perform particular tasks or implement particular abstract data types. The functionality of the program modules may be combined or split between program modules as desired in various embodiments. Machine-executable instructions for program modules may be executed within a local or distributed device. In a distributed device, program modules may be located in both local and remote storage media.
Program code for carrying out methods of the present disclosure may be written in any combination of one or more programming languages. The program code may be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the program code, when executed by the processor or controller, cause the functions/operations specified in the flowcharts and/or block diagrams to be implemented. The program code may execute entirely on a machine, partly on the machine, as a stand-alone software package, partly on the machine and partly on a remote machine or entirely on the remote machine or server.
In the context of the present disclosure, the computer program code or related data may be carried by any suitable carrier to enable the device, apparatus or processor to perform various processes and operations as described above. Examples of the carrier include a signal, computer readable medium, and the like.
The computer readable medium may be a computer readable signal medium or a computer readable storage medium. A computer readable medium may include but not limited to an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the computer readable storage medium would include an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM) , a read-only memory (ROM) , an erasable programmable read-only memory (EPROM or Flash memory) , an optical fiber, a portable compact disc read-only memory (CD-ROM) , an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.
Further, although operations are depicted in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. In certain circumstances, multitasking and parallel processing may be advantageous. Likewise, although several specific implementation details are contained in the above discussions, these should not be construed as limitations on the scope of the present disclosure, but rather as descriptions of features that may be specific to particular embodiments. Unless explicitly stated, certain features that are described in the context of separate embodiments may also be implemented in combination in a single embodiment. Conversely, unless explicitly stated, various features that are described in the context of a single embodiment may also be implemented in a plurality of embodiments separately or in any suitable sub-combination.
Although the present disclosure has been described in languages specific to structural features and/or methodological acts, it is to be understood that the present disclosure defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing the claims.

Claims

An apparatus comprising:

at least one processor; and

at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to:

obtain data representing a signal received at a first device transmitted by a second device and channel estimation between the first and second devices, wherein the signal is transmitted over a plurality of layers;

generate, by a first sub-model of a machine learning model, shared feature representations for the plurality of layers based on the data and the channel estimation; and

generate, by a second sub-model of the machine learning model based on the shared feature representations, respective predictions on bits for the plurality of layers.
The apparatus of claim 1, wherein the signal is transmitted in a Discrete Fourier Transform-Spread-Orthogonal Frequency Division Multiplexing waveform and the second sub-model comprises at least one module for Inverse Discrete Fourier Transform.
The apparatus of claim 1, wherein the second sub-model comprises a plurality of branches and a branch of the plurality of branches corresponds to a layer of the plurality of layers, and

the instructions, when executed by the at least one processor, cause the apparatus to:

feed the shared feature representations into each branch of the plurality of branches to generate the prediction on bits for the corresponding layer.
The apparatus of claim 3, wherein the branch of the plurality of branches comprises:

a first neural network used to extract, from the shared feature representations, feature representations for the layer corresponding to the branch;

a second neural network used to perform joint channel estimation and equalization specific to the layer corresponding to the branch; and

a third neural network used to perform demodulation specific to the layer corresponding to the branch.
The apparatus of claim 4, wherein the data is part of a training sample for the machine learning model and the training sample further comprises respective transmitted bit sequences for the plurality of layers of the signal, and

the instructions, when executed by the at least one processor, further cause the apparatus to:

for a layer of the plurality of layers,

determine a loss based on a difference between the prediction on bits for the layer and the transmitted bit sequence for the layer;

update one or more neural networks of the corresponding branch based on the loss; and

update the first sub-model based on a sum of respective losses determined for the plurality of layers.
The apparatus of claim 1, wherein the instructions, when executed by the at least one processor, cause the apparatus to:

divide the shared feature representations into a plurality of groups of feature representations, a group of the plurality of groups corresponding to a layer of the plurality of layers; and

feed each group of feature representations into the second sub-model to generate the prediction on bits for the corresponding layer.
The apparatus of claim 6, wherein the second sub-model comprises:

a fourth neural network used to perform joint channel estimation and equalization common to the plurality of layers; and

a fifth neural network used to perform demodulation common to the plurality of layers.
The apparatus of claim 6, wherein the data is part of a training sample for the machine learning model and the training sample further comprises respective transmitted bit sequences for the plurality of layers of the signal, and

the instructions, when executed by the at least one processor, further cause the apparatus to:

for a layer of the plurality of layers, determine a loss based on a difference between the prediction on bits for the layer and the transmitted bit sequence for the layer;

update one or more neural networks of the second sub-model based on respective losses determined for the plurality of layers; and

update the first sub-model based on a sum of the respective losses.
The apparatus of claim 1, wherein the instructions, when executed by the at least one processor, cause the apparatus to:

feed a concatenation of the data and the channel estimation into a neural network of the first sub-model to generate the shared feature representations.
The apparatus of claim 1, wherein the instructions, when executed by the at least one processor, cause the apparatus to:

feed to the data into a first neural network of the first sub-model to generate first feature representations for the data;

feed the channel estimation into a second neural network of the first sub-model to generate second feature representations for the channel estimation; and

transform the first and second feature representations to the shared feature representation.
The apparatus of claim 1, wherein the apparatus comprises a training entity for the machine learning model, and the data and the channel estimation are obtained from the first device for training the machine learning model.
The apparatus of claim 11, wherein the instructions, when executed by the at least one processor, further cause the apparatus to:

transmit the trained machine learning model to the first device.
The apparatus of claim 1, wherein the apparatus comprises the first device, and the instructions, when executed by the at least one processor, further cause the apparatus to:

select the machine learning model from a plurality of candidate machine learning models based on a configuration of the signal, wherein the configuration at least comprises:

a rank for determining the number of the plurality of layers, and

a modulation scheme for determining the number of the bits for a layer of the plurality of layers.
The apparatus of claim 13, wherein the configuration further comprises at least one of:

the number of subcarriers for transmitting the signal, or

whether transform-precoding is enabled for the signal.
The apparatus of claim 13, wherein the instructions, when executed by the at least one processor, further cause the apparatus to:

replace the machine learning model by a non-machine learning approach to decode a further signal.
The apparatus of claim 15, wherein the instructions, when executed by the at least one processor, further cause the apparatus to:

in response to the replacement of the machine learning model by the non-machine learning approach, update the machine learning model.
A method comprising:

obtaining data representing a signal received at a first device transmitted by a second device and channel estimation between the first and second devices, wherein the signal is transmitted over a plurality of layers;

generating, by a first sub-model of a machine learning model, shared feature representations for the plurality of layers based on the data and the channel estimation; and

generating, by a second sub-model of the machine learning model based on the shared feature representations, respective predictions on bits for the plurality of layers.
An apparatus comprising:

means for obtaining data representing a signal received at a first device transmitted by a second device and channel estimation between the first and second devices, wherein the signal is transmitted over a plurality of layers;

means for generating, by a first sub-model of a machine learning model, shared feature representations for the plurality of layers based on the data and the channel estimation; and

means for generating, by a second sub-model of the machine learning model based on the shared feature representations, respective predictions on bits for the plurality of layers.
A computer readable medium comprising instructions stored thereon for causing an apparatus at least to perform the method of claim 17.