Receiver based on machine learning in wireless communication network

Abstract

The present disclosure relates to a machine learning based receiver in a wireless communication network that is invariant for the number of multiple-input multiple-output (MIMO) layers processed. For each MIMO layer, the ML-based receiver is configured to obtain an intermediate estimate set by performing an equalization operation based on channel information and a symbol array received on the MIMO layer; and obtaining a final estimation set for the symbol array by using a layer association block of the ML model. Each layer association block of the ML model is configured to receive the intermediate estimate set as input data and output a final estimate set. The layer association blocks are configured to operate in parallel and exchange weights with each other during the training phase. In some embodiments, the layer-associated blocks may also be configured to exchange input data with each other during the training phase and the reasoning phase.

Description

Technical Field

[0001] The present disclosure generally relates to the field of wireless communications. In particular, the present disclosure relates to a machine learning (ML) based receiver and a method of operating the same in a wireless communication network. Background Art

[0002] Recently, machine learning (ML) based receivers have been developed, where part of the ML based receiver is learned by a neural network (NN). This facilitates improved performance and higher flexibility, since everything is learned directly from the input data. A specific implementation of such a receiver is the DeepRx receiver (for example, see Mikko Honekala et al., "DeepRx: Fully Convolutional Deep Learning Receiver", IEEE Transactions on Wireless Communications, vol. 20, no. 6, pp. 3925-3940, June 2021, doi: 10.1109 / TWC.2021.3054520 and D. Korpi et al., "DeepRx MIMO: Convolutional MIMO Detection with Learned Multiplicative Transformation", IEEE International Conference on Communications (ICC), 14-23 June, 2021, DOI: 10.1109 / ICC42927.2021.9500518). The DeepRx receiver is based on a deep convolutional NN (CNN), and it achieves high performance in various fifth generation (5G) multiple-input multiple-output (MIMO) scenarios.

[0003] More specifically, simulation results show that DeepRx receivers achieve high performance with sparser demodulation reference signal (DMRS) patterns than required by conventional (non-ML-based) receivers. This is because the DeepRx receiver essentially learns to perform data-assisted detection, where it exploits unknown received data symbols to make detection more accurate. During data-assisted detection, the DeepRx receiver knows the underlying constellation shape and can effectively use this knowledge to properly equalize the received signal.

[0004] While the original MIMO DeepRx architecture can in principle be trained to operate with any number of MIMO layers, it can be quite inefficient in terms of model size, training length, and radio performance because the original DeepRx architecture is monolithic and was originally trained based on a single maximum number of MIMO layers. Therefore, more efficient solutions are needed, especially when going from today's 2-4 to 8-16 MIMO layers, which are planned for next-generation (e.g., 6G) communication systems. Summary of the invention

[0005] This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features of the disclosure, nor is it intended to be used to limit the scope of the disclosure.

[0006] An object of the present disclosure is to provide a technical solution that allows a single ML-based receiver to be used effectively for any number of MIMO layers.

[0007] The above objects are achieved by the features of the independent claims in the attached claims. Further embodiments and examples are evident from the dependent claims, the detailed description and the drawings.

[0008] According to a first aspect, a ML-based receiver in a wireless communication network is provided. The ML-based receiver includes at least one processor and at least one memory storing instructions, which, when executed by the at least one processor, cause the ML-based receiver to perform at least the following operations. First, the ML-based receiver receives a symbol array and a reference signal set on a plurality of MIMO layers associated with a MIMO channel. Then, the ML-based receiver uses the reference signal set and the symbol array to obtain channel information indicating the state of the MIMO channel. Thereafter, for each of the plurality of MIMO layers, the ML-based receiver obtains an intermediate estimate set for the symbol array by performing an equalization operation based on the channel information and the symbol array; and obtains a final estimate set for the symbol array by using a layer association block of an ML model. Each layer association block of the ML model has a weight set and is configured to receive the intermediate estimate set as input data and output a final estimate set as a response. The layer association blocks of the ML model are further configured to operate in parallel and exchange weight sets with each other during a training phase. Once trained, the ML-based receiver according to the first aspect can perform parallel data-assisted processing of MIMO layers using shared training weights for each MIMO layer. Thus, the ML-based receiver according to the first aspect can be first trained for a number of MIMO layers and then used directly for a smaller number of MIMO layers or even for a larger number of MIMO layers after simple and efficient fine-tuning. In other words, no training is required every time a new number of MIMO layers is to be processed. Thus, the ML-based receiver according to the first aspect can be considered as a receiver independent of the number of MIMO layers to be processed.

[0009] In an exemplary embodiment of the first aspect, each of the layer association blocks of the ML model is further configured to receive input data (i.e., intermediate estimation sets) from other layer association blocks of the ML model as a combined value independent of the number of MIMO layers in the plurality of MIMO layers. Consistent with the shared or exchanged trained weights, the exchange of input data between the layer association blocks of the ML model can contribute to better independence of the ML-based receiver from the number of MIMO layers, thereby resulting in its appropriate detection accuracy for any number of MIMO layers.

[0010] In an exemplary embodiment of the first aspect, the combined value includes at least one of: an arithmetic mean of input data of other layer-associated blocks of the ML model; a sum of input data of other layer-associated blocks of the ML model; a difference of input data of other layer-associated blocks of the ML model; a product of input data of other layer-associated blocks of the ML model; and a value obtained by concatenating input data of other layer-associated blocks of the ML model. By using the combined value defined in this way, input data from layer-associated blocks of the ML model can be easily and simply “mixed”, so that the operation of the ML-based receiver is invariant to the number of MIMO layers.

[0011] In an exemplary embodiment of the first aspect, each of the layer association blocks of the ML model is further configured to iteratively process input data until a user-defined stop condition is met. If the stop condition is met for one of the MIMO layers, the corresponding layer association block of the ML model is further configured to output a processing result obtained in the last iteration as a final set of estimated values. Therefore, if the stop condition is met, the iterative scheme may provide an early exit for a single MIMO layer processing while continuing to process other MIMO layers. By doing so, average power consumption can be reduced because the trained model complexity is linearly related to the number of MIMO layers.

[0012] In an exemplary embodiment of the first aspect, each of the layer association blocks of the ML model is further configured to use a different set of intermediate weights at each iteration. In other words, the weights can be trained separately for each iteration. This embodiment can be advantageously used depending on the hardware and radio performance requirements applied to the wireless communication network.

[0013] In an exemplary embodiment of the first aspect, each of the layer association blocks of the ML model includes at least one of: a convolutional neural network (CNN), a transformer neural network (TNN), and a self-attention based neural network. By using these types of neural networks, the final estimation set can be predicted more efficiently, thereby improving the operation of the ML-based receiver.

[0014] In an exemplary embodiment of the first aspect, the equalization operation is a linear minimum mean square error (LMMSE) based equalization or a maximum ratio combining (MRC) based equalization. By using these types of equalization techniques, the intermediate estimation set can be predicted more efficiently, thereby improving the operation of the ML-based receiver. In addition, these types of equalization techniques can additionally "mix" information about the MIMO layers, so that the dependence on the number of MIMO layers has been excluded from further consideration. In view of this, it may not be necessary to mix the input data of the ML model. Therefore, by using these types of equalization techniques, the ML-based receiver can have a more simplified architecture.

[0015] According to a second aspect, a method for operating an ML-based receiver in a wireless communication network is provided. The method starts with a step of receiving a symbol array and a reference signal set on a plurality of MIMO layers associated with a MIMO channel. Then, the method proceeds to a step of obtaining channel information indicating the state of the MIMO channel using the reference signal set and the symbol array. Next, for each of the plurality of MIMO layers, the following steps are performed: obtaining an intermediate estimation set for the symbol array by performing an equalization operation based on the channel information and the symbol array; and obtaining a final estimation set for the symbol array by using a layer association block of an ML model. Each layer association block of the ML model has a weight set and is configured to receive the intermediate estimation set as input data and output a final estimation set. The layer association block of the ML model is further configured to operate in parallel during a training phase and to exchange weight sets with each other. By doing so, the ML-based receiver can perform parallel data-assisted processing of the MIMO layers using shared training weights for each MIMO layer. Therefore, the ML-based receiver can be trained for a plurality of MIMO layers and then directly used for a smaller number of MIMO layers or even for a greater number of MIMO layers after simple and efficient fine-tuning. In other words, there is no need to train the ML-based receiver every time a new number of MIMO layers is to be processed.

[0016] In an exemplary embodiment of the second aspect, each of the layer association blocks of the ML model is further configured to receive input data from other layer association blocks of the ML model during a training phase as a combined value independent of the number of MIMO layers in the plurality of MIMO layers. Consistent with the shared or exchanged trained weights, the exchange of input data between the layer association blocks of the ML model can contribute to better independence of the ML-based receiver from the number of MIMO layers, thereby resulting in its appropriate detection accuracy for any number of MIMO layers.

[0017] In an exemplary embodiment of the second aspect, the combined value includes at least one of: an arithmetic mean of input data of other layer-associated blocks of the ML model; a sum of input data of other layer-associated blocks of the ML model; a difference of input data of other layer-associated blocks of the ML model; a product of input data of other layer-associated blocks of the ML model; and a value obtained by concatenating input data of other layer-associated blocks of the ML model. By using the combined value configured in this way, input data from layer-associated blocks of the ML model can be easily and simply “mixed”, so that the operation of the ML-based receiver is invariant to the number of MIMO layers.

[0018] In an exemplary embodiment of the second aspect, each of the layer association blocks of the ML model is further configured to iteratively process input data until a user-defined stop condition is met. If the stop condition is met for one of the MIMO layers, the corresponding layer association block of the ML model is further configured to output a processing result obtained in the last iteration as a final set of estimated values. Therefore, if the stop condition is met, the iterative scheme may provide an early exit for a single MIMO layer processing while continuing to process other MIMO layers. By doing so, average power consumption can be reduced because the trained model complexity is linearly related to the number of MIMO layers.

[0019] In an exemplary embodiment of the second aspect, each of the layer association blocks of the ML model is further configured to use a different set of intermediate weights at each iteration. In other words, the weights can be trained separately for each iteration. This embodiment can be advantageously used depending on the hardware and radio performance requirements applied to the wireless communication network.

[0020] In an exemplary embodiment of the second aspect, each of the layer association blocks of the ML model includes at least one of: a CNN, a TNN, and a self-attention based neural network. By using these types of neural networks, the final estimation set can be predicted more efficiently, thereby improving the operation of the ML-based receiver.

[0021] In an example embodiment of the second aspect, the equalization operation is LMMSE-based equalization or MRC-based equalization. By using these types of equalization techniques, the intermediate estimation set can be predicted more efficiently, thereby improving the operation of the ML-based receiver. In addition, these types of equalization techniques can additionally "mix" information about the MIMO layers, so that the dependence on the number of MIMO layers has been excluded from further consideration. In view of this, it may not be necessary to mix the input data of the layer association block of the ML model. Therefore, by using these types of equalization techniques, the ML-based receiver can have a more simplified architecture.

[0022] According to a third aspect, a computer program product is provided. The computer program product comprises a computer readable storage medium storing computer code. When the computer code is executed by at least one processor, the at least one processor is caused to perform the method according to the second aspect. By using such a computer program product, the implementation of the method according to the second aspect can be simplified in any ML-based receiver, such as the ML-based receiver according to the first aspect.

[0023] According to a fourth aspect, a ML-based receiver in a wireless communication network is provided. The ML-based receiver includes: a component for receiving a symbol array on multiple MIMO layers associated with a MIMO channel; and a component for receiving a reference signal set on multiple MIMO layers. The ML-based receiver also includes: a component for obtaining channel information indicating the state of the MIMO channel using the reference signal set and the symbol array. The ML-based receiver also includes: a component for performing the following operations for each MIMO layer in the multiple MIMO layers: obtaining an intermediate estimation set for the symbol array by performing an equalization operation based on the channel information and the symbol array; and obtaining a final estimation set for the symbol array by using a layer association block of an ML model. Each layer association block of the ML model has a weight set and is configured to receive the intermediate estimation set as input data and output a final estimation set. The layer association block of the ML model is further configured to operate in parallel during the training phase and exchange weight sets with each other. Once trained, the ML-based receiver according to the fourth aspect can perform parallel data-assisted processing of the MIMO layers using shared training weights for each MIMO layer. Thus, the ML-based receiver according to the fourth aspect can be first trained for a number of MIMO layers and then used directly for a smaller number of MIMO layers or even for a larger number of MIMO layers after simple and efficient fine-tuning. In other words, no training is required every time a new number of MIMO layers is to be processed. Thus, the ML-based receiver according to the fourth aspect can be considered as a receiver independent of the number of MIMO layers to be processed.

[0024] Other features and advantages of the present disclosure will be apparent by reading the following detailed description and reviewing the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The present disclosure is explained below in conjunction with the accompanying drawings:

[0026] Figure 1 A block diagram of a machine learning (ML) based receiver (i.e., a DeepRx receiver) according to the prior art is shown;

[0027] Figure 2 shows a block diagram of an ML-based receiver according to an example embodiment;

[0028] Figure 3 A method for operating according to an example embodiment is shown. Figure 2 A flowchart of a method for an ML-based receiver;

[0029] Figure 4 The first exemplary embodiment includes Figure 2 A block diagram of a processor in a receiver;

[0030] Figure 5 The second exemplary embodiment includes Figure 2 A block diagram of a processor in a receiver;

[0031] Figure 6 It is shown that according to an example embodiment, Figure 5 A block diagram of a layer mixer used in a processor of;

[0032] Figure 7 The third exemplary embodiment includes Figure 2 A block diagram of a processor in a receiver;

[0033] Fig. 8A and Figure 8B Two different methods for training each layer association block of an ML model according to an example embodiment are schematically shown, namely: Fig. 8A A first (iterative) approach is shown, where each next training iteration uses the intermediate weights obtained on the previous training iteration, while Figure 8B A second (expanded) approach is shown, where each training iteration uses a different set of intermediate weights;

[0034] Fig.9A and Fig. 9B shows the result of performing the operation without a layer mixer ( Fig.9A ) and with a layer mixer ( Fig. 9B ) using MRC-based equalization for the two cases of uncoded bit error rate (BER) and signal to interference and noise ratio (SINR) in dB;

[0035] Fig. 10A and Fig. 10B shows the result of performing the operation without a layer mixer ( Fig. 10A ) and with a layer mixer ( Fig. 10B ) using LMMSE-based equalization for the two cases of uncoded BER and SINR in dB; and

[0036] Fig.11A and Fig. 11B The uncoded BER and SINR in dB are shown for two different MIMO settings, namely: 16x4 MIMO setting ( Fig.11A ) and 16x2 MIMO settings ( Fig. 11B ). DETAILED DESCRIPTION

[0037] Various embodiments of the present disclosure are described in more detail with reference to the accompanying drawings. However, the present disclosure may be embodied in many other forms and should not be construed as limited to any specific structure or function discussed in the following description. Instead, these embodiments are provided to make the description of the present disclosure detailed and complete.

[0038] According to the detailed description, it is obvious to those skilled in the art that the scope of the present disclosure covers any embodiment thereof disclosed herein, whether the embodiment is implemented independently or in accordance with any other embodiment of the present disclosure. For example, the apparatus and method disclosed herein can be implemented in practice by using any number of embodiments provided herein. In addition, it should be understood that any embodiment of the present disclosure can be implemented using one or more elements presented in the attached claims.

[0039] Any embodiment described herein as an "exemplary embodiment" is not necessarily to be construed as preferred or having advantages over other embodiments, unless otherwise stated.

[0040] According to the example embodiments disclosed herein, a user equipment (UE) may refer to an electronic computing device configured to perform wireless communication. UE may be implemented as a mobile station, a mobile terminal, a mobile subscriber unit, a mobile phone, a cellular phone, a smart phone, a cordless phone, a personal digital assistant (PDA), a wireless communication device, a desktop computer, a laptop computer, a tablet computer, a gaming device, a netbook, a smartbook, an ultrabook, a medical mobile device or equipment, a biometric sensor, a wearable device (e.g., a smart watch, smart glasses, a smart wristband, etc.), an entertainment device (e.g., an audio player, a video player, etc.), a vehicle component or sensor (e.g., a driver assistance system), a smart meter / sensor, an unmanned vehicle (e.g., an industrial robot, a quadcopter, etc.) and its components (e.g., an autonomous vehicle computer), industrial manufacturing equipment, a global positioning system (GPS) device, an Internet of Things (IoT) device, an industrial IoT (IIoT) device, a machine type communication (MTC) device, a group of large-scale IoT (MIoT) or large-scale MTC (mMTC) devices / sensors, or any other suitable mobile device configured to support wireless communication. In some embodiments, a UE may refer to at least two juxtaposed and interconnected UEs defined in this way.

[0041] As used in the example embodiments disclosed herein, a network node may refer to a communication fixed point / communication node for a UE in a specific wireless communication network. More specifically, a network node may be used to connect a UE to a data network (DN) through a core network (CN), and may be referred to as a base transceiver station (BTS) according to 2G communication technology, a NodeB according to 3G communication technology, an evolved NodeB (eNodeB or eNB) according to 4G communication technology, and a gNB according to 5G new radio (NR) communication technology. A network node may serve different cells, such as macro cells, micro cells, micro cells, femto cells, and / or other types of cells. A macro cell may cover a relatively large geographic area (e.g., a radius of at least several kilometers). For example, a micro cell may cover a geographic area with a radius of less than two kilometers. A micro cell may cover a relatively smaller geographic area, such as an office, a shopping mall, a train station, a stock exchange, etc. A femto cell may cover an even smaller geographic area (e.g., a home). Accordingly, a network node serving a macro cell may be referred to as a macro node, a network node serving a micro cell may be referred to as a micro node, and so on.

[0042] According to the example embodiments disclosed herein, the wireless communication network in which the UE and the network node communicate with each other may refer to a cellular or mobile network, a wireless local area network (WLAN), a wireless personal area network (WPAN), a wireless wide area network (WWAN), a satellite communication (SATCOM) system, or any other type of wireless communication network. Each of these types of wireless communication networks supports wireless communication according to one or more communication protocol standards. For example, a cellular network may operate according to a global system for mobile communications (GSM) standard, a code division multiple access (CDMA) standard, a wideband code division multiple access (WCDM) standard, a time division multiple access (TDMA) standard, or any other communication protocol standard, a WLAN may operate according to one or more versions of the IEEE 802.11 standard, a WPAN may operate according to an infrared data association (IrDA), wireless USB, Bluetooth, or ZigBee standard, and a WWAN may operate according to a global microwave access interoperability (WiMAX) standard.

[0043] MIMO technology can be used to perform data transmission between UEs, between network nodes, or between UEs and network nodes. MIMO technology involves using multiple transmit antennas at a transmitting entity (e.g., UE or network node) and multiple receive antennas at a receiving entity (e.g., another UE or network node) for data transmission. The MIMO channel formed by the transmit antenna and the receive antenna can be decomposed into spatial layers (also referred to as MIMO layers). MIMO layers can be used to transmit data in parallel to achieve higher throughput and / or redundantly transmit data to achieve higher reliability. MIMO layers may experience various harmful channel conditions (e.g., fading, multipath, interference effects, etc.), for which they can achieve different signal-to-noise-interference ratios (SNRs). The SNR of each MIMO layer determines its transmission capacity, which is usually quantified by a specific data rate that is reliably transmitted on the MIMO layer. For time-varying wireless channels, channel conditions change over time, and the SNR of each MIMO layer also changes over time. The different SNRs of the MIMO layers plus the time-varying nature of the SNR for each MIMO layer make it challenging to efficiently transmit data in a MIMO system.

[0044] ML-based receivers have recently been developed that allow data transmitted in a MIMO system to be efficiently and reliably recovered or decoded at a receiving entity. More specifically, the ML-based receiver is configured to predict an estimate (e.g., a log-likelihood ratio (LLR)) of the transmitted data by using an ML model (e.g., a neural network (NN)).

[0045] Figure 1 A block diagram of an ML-based receiver 100 according to the prior art is shown. In particular, the ML-based receiver 100 corresponds to the original architecture of the DeepRx receiver. The ML-based receiver 100 comprises a pre-processing block 102 for receiving an array of data or symbols on a set of MIMO layers which may express: y=Hx+n, where y is the received data array, F is the number of subcarriers, S is the number of symbols carrying pilots (typically 14 in 5G systems), 3M is the dimension of the input / output data (e.g., if block 102 is implemented as a convolutional NN (CNN), it may correspond to the number of convolutional channels or the number of neurons in some other ML-based architecture), H is the CSI matrix, x is the transmitted data array, and n is the noise plus interference signal. Thus, y may be considered a distorted version of x (e.g., due to noise). It is also assumed that block 102 is responsible for receiving reference signals or pilot signals and using them together with y to obtain the CSI matrix H. Block 102 may be implemented as a trained (i.e., NN-based) block.

[0046] The ML-based receiver 100 also includes a set of equalization blocks 104 (or in other words, equalizers (EQ)), each of which is configured to perform an equalization operation (e.g., LMMSE) based on y and H for each MIMO layer to obtain an intermediate estimate set associated with the transmitted data array (i.e., x) for that MIMO layer. After obtaining the set of intermediate estimates, they are scaled to a certain size in block 106. Block 106 can also be implemented as a training block, and its output is a set of learned projections, each of which corresponds to one MIMO layer and has a size of F×S×3N, where N is the dimension of the input / output data (e.g., if block 106 is implemented as a CNN, it can correspond to the number of convolutional channels or the number of neurons in some other ML-based architecture). The learned projections are next fed to the DeepRx part 108, where they are refined to obtain the final set of estimates (as regular, probabilistic estimates) for each MIMO layer. The DeepRx part 108 can be based on the training set by using the mean squared error (MSE) loss function (i.e., ) is implemented by a CNN trained with . The final estimation set is further fed to a demapper 110, which is configured to obtain bit LLRs based on the final estimation set for each MIMO layer. More specifically, the demapper 110 outputs an array of bit LLRs Where N t is the number of MIMO layers, and N b is the number of bits. The demapper 110 may also be implemented as a training block, and the ML-based receiver 100 may be configured to compute a binary cross entropy (CE) loss function at its output.

[0047] It should be noted that the training blocks of the ML-based receiver 100 are monolithic in the sense that they are always trained to some exact number of MIMO layers, and there is no flexibility in using the training architecture of the ML-based receiver 100 to some other number of MIMO layers.

[0048] The example embodiments disclosed herein provide technical solutions that allow mitigation or even elimination of the above-mentioned disadvantages that are disadvantageous to the prior art. In particular, the technical solution disclosed herein relates to an ML-based receiver that is invariant to the number of MIMO layers processed. More specifically, for each MIMO layer, the ML-based receiver is configured to obtain an intermediate estimation set by performing an equalization operation based on channel information (e.g., a CSI matrix) and a symbol array received on a MIMO layer constituting a MIMO channel, and then obtain a final estimation set of the symbol array by using a layer association block of an ML model. Each layer association block of the ML model is configured to receive an intermediate estimation set as input data and output a final estimation set. The layer association blocks of the ML model are also configured to operate in parallel during the training phase and exchange weight sets with each other. In some embodiments, the layer association blocks of the ML model may also be configured to exchange input data with each other during the training phase and the inference phase.

[0049] Figure 2 1 shows a block diagram of an ML-based receiver 200 according to an example embodiment. The ML-based receiver 200 is intended to be part of a UE or a network node in a wireless communication network. Figure 2 As shown, the ML-based receiver 200 includes a processor 202 and a memory 204. The memory 204 stores processor-executable instructions 206, which, when executed by the processor 202, cause the processor 202 to perform various aspects of the present disclosure, as will be described in more detail below. Figure 2 The number, arrangement, and interconnection of the constructive elements of the ML-based receiver 200 shown in the figure are not intended to be any limitation of the present disclosure, but are only used to provide a general concept of how the constructive elements can be implemented within the ML-based receiver 200. For example, the processor 202 can be replaced by several processors, and the memory 204 can be replaced by several removable and / or fixed storage devices, depending on the specific application. In addition, it is assumed that the processor 202 is capable of performing different operations required to implement data reception and transmission, such as signal modulation / demodulation, encoding / decoding, etc.

[0050] The processor 202 may be implemented as a CPU, a general purpose processor, a special purpose processor, a microcontroller, a microprocessor, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), a complex programmable logic device, etc. It should also be noted that the processor 202 may be implemented as any combination of one or more of the above. As an example, the processor 202 may be a combination of two or more microprocessors.

[0051] The memory 204 may be implemented as a classic non-volatile or volatile memory used in modern electronic computing machines. As examples, non-volatile memory may include read-only memory (ROM), ferroelectric random access memory (RAM), programmable ROM (PROM), electrically erasable PROM (EEPROM), solid-state drive (SSD), flash memory, magnetic disk storage devices (such as hard disk drives and tapes), optical disk storage devices (such as CDs, DVDs, and Blu-ray discs), etc. For volatile memory, examples include dynamic RAM, synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), static RAM, etc.

[0052] The processor executable instructions 206 stored in the memory 204 may be configured to cause the processor 202 to execute computer executable program codes of various aspects of the present disclosure. The computer executable program codes for performing the operations or steps of various aspects of the present disclosure may be written in any combination of one or more programming languages ​​(such as Java, C++, etc.). In some examples, the computer executable program codes may be in the form of a high-level language or in a precompiled form and generated by an interpreter (also pre-stored in the memory 204) during operation.

[0053] Figure 3A flow chart of a method 300 for operating an ML-based receiver 200 according to an example embodiment is shown. The method 300 begins at step S302, where the processor 202 (e.g., via an antenna that is part of or coupled to the ML-based receiver 200) receives a symbol array on multiple MIMO layers associated with a MIMO channel. The method then proceeds to step S304, where the processor 202 receives (e.g., via the same antenna) a reference signal set (e.g., pilot signal) on multiple MIMO layers. It should be noted that steps S302 and S304 can be performed in parallel if necessary and depending on the capabilities of the processor 202. Next, the method 300 proceeds to step S306, where the processor 202 obtains channel information based on the reference signal set and the symbol array. The channel information indicates the state of the MIMO channel and may include, for example, a CSI matrix. Thereafter, steps S308 and S310 are performed in parallel for each of the MIMO layers, where the processor 202 first obtains an intermediate estimate set for the symbol array and then obtains a final estimate set for the symbol array. More specifically, an intermediate estimation set is obtained by performing an equalization operation based on channel information and a symbol array. Some examples of equalization operations may include, but are not limited to, linear minimum mean square error (LMMSE)-based equalization or maximum ratio combining (MRC)-based equalization. A final estimation set is obtained for each MIMO layer using a layer association block of an ML model (e.g., CNN), the layer association block being configured to receive an intermediate estimation set as input data and, in response, predict a final estimation set. The layer association blocks of the ML model are also configured to operate in parallel and exchange their weight sets with each other during the training phase of the ML model. Each layer association block of the ML model can be implemented as a separate block of the same NN (e.g., CNN, TNN, or self-attention-based NN).

[0054] In one embodiment, each layer-associated block of the ML model may also be configured to receive input data (i.e., intermediate estimation sets) from other layer-associated blocks of the ML model as a combined value independent of the number of MIMO layers during the training phase and the inference phase of the ML model. The combined value may be expressed as an arithmetic mean of the input data of the other layer-associated blocks of the ML model. However, the present disclosure is not limited to such a combined value, and in some embodiments, other mathematical operations may be used to combine the input data from other layer-associated blocks of the ML model so that the combined value is independent of the number of MIMO layers. Such other mathematical operations may include, but are not limited to, addition, subtraction, multiplication, concatenation (e.g., concatenation of arithmetic means and their variances), and different moving average techniques.

[0055] In one embodiment, each layer-associated block of the ML model can be configured to iteratively refine the input data (i.e., the intermediate estimation set) until a user-defined stopping condition is met. In other words, if the stopping condition is met, the iterative refinement scheme can provide an early exit for the processing of a separate MIMO layer while continuing to process other MIMO layers. The stopping condition can be related to the accuracy that the user expects to obtain the final estimation set. If the stopping condition is met for one of the layer-associated blocks of the ML model, the layer-associated block can stop the iterative processing and output the processing result obtained in the last iteration as the final estimation value set. It should also be noted that if necessary, each layer-associated block of the ML model can use a different set of intermediate weights at each iteration, i.e., the weights of different iterations may not be bound to each other (hereinafter referred to as the unfolded case). The opposite case (hereinafter referred to as the iterative case) is also possible, i.e., the weights of the previous iteration can be used in the next iteration. The selection between the iterative case and the unfolded case can be made based on the hardware and radio performance requirements applicable to the wireless communication network.

[0056] Figure 4 A block diagram of a processor 202 included in a receiver 200 according to a first example embodiment is shown. In the first example embodiment, the processor 202 includes a set of pre-processing blocks 400, a set of equalization blocks 402, and a set of layer association blocks 404-1, 404-2, and 404-3 of an ML model. Each pre-processing block in the set of pre-processing blocks 400 may be configured as a pre-processing block 102 of an ML-based receiver 100. In other words, each pre-processing block in the set of pre-processing blocks 400 may receive a symbol array (in the frequency domain) and a reference signal set on a certain MIMO layer, and use them to obtain channel information (e.g., a CSI matrix). The set of equalization blocks 402 may be configured as the set of equalization blocks 104, i.e., each equalization block in the set of equalization blocks 402 may perform an equalization operation for one of the MIMO layers. The equalization operation may be represented by LMMSE, written as: in is a set of intermediate estimates of the symbol array to be recovered in the ML receiver 200, is the noise variance, and I is the identity matrix. Note that the inverse here uses information from all MIMO layers for all of its outputs. An approximation of the LMMSE can also be used (e.g., using an approximate matrix inverse), however, it sacrifices accuracy for a lower computational cost.

[0057] The equalization operation can also be represented by MRC, which involves processing each MIMO layer independently of each other and can be written as: where hi represents the i-th column of H.

[0058] The layer association block 406-1 of the ML model is configured to refine the first MIMO layer (L 1 ) obtained by the intermediate estimation set (i.e., ), the layer association block 406-2 of the ML model is configured to refine the second MIMO layer (L 2 ) obtained, and the layer association block 406-3 of the ML model is configured to refine the intermediate estimation set for the third MIMO layer (L 3 ) The intermediate estimation set obtained by the layer association blocks 406-1, 406-2 and 406-3 of the ML model can each output a final estimation set as a set of LLRs. Figure 4 As shown, the layer association blocks 406-1, 406-2, and 406-3 of the ML model are related in the sense that their weights are exchanged between each other during the training phase of the ML model. Those skilled in the art will recognize that the number and arrangement of the constructive elements or blocks of the processor 202 are for illustrative purposes only. Figure 4 As shown in FIG. 1 and should not be considered as any limitation of the present disclosure, in other embodiments, the processor 202 may include more than three layer association blocks of the ML model, each of which is related to one of the MIMO layers to be processed.

[0059] Figure 5 2 shows a block diagram of a processor 202 included in a receiver 200 according to a second exemplary embodiment. In the second exemplary embodiment, the processor 202 includes a set of pre-processing blocks 500 and a set of equalization blocks 502, which may be implemented in the same or similar manner as the set of pre-processing blocks 400 and the set of equalization blocks 402, respectively. Figure 5 As shown, the processor 202 also includes learning projection blocks 504-1, 504-2 and 504-3, each of which can be configured as block 106 of the ML-based receiver 100. In other words, each of the learning projection blocks 504-1, 504-2 and 504-3 can scale the intermediate estimation set of one of the three MIMO layers so that it has an appropriate (for further processing) size (i.e., F×S×N). For example, this can be done by using a linear 1x1 convolution. Thereafter, the learning projections (i.e., the intermediate estimation set of appropriate size) obtained by blocks 504-1, 504-2 and 504-3 are fed to layer mixers 506-1, 506-2 and 506-3, respectively. The layer mixer 506-1 is configured to scale the intermediate estimation set for one of the three MIMO layers L 1 The learned projections obtained are the same as for the remaining (two) MIMO layers L 2 and L 3The obtained learning projections are mixed so that the obtained learning projections are independent of the number of MIMO layers. Each of the layer mixers 506-2 and 506-3 is configured to perform similar mixing operations, but for the second and third MIMO layers, respectively.

[0060] Figure 6 1 shows a block diagram of a layer mixer 600 that can be used in the processor 202 according to an example embodiment. In other words, the layer mixer 600 can be used as any one of the layer mixers 506-1, 506-2, and 506-3 in the architecture of the processor 202 according to the second example embodiment. Figure 5 The layer mixer 600 uses the first MIMO layer L 1 The linear point-by-point convolution is used to select the other MIMO layers L 2 and L 3 Which information will be sent, and on the other hand from other MIMO layers L 2 and L 3 It should be noted that this implementation is independent of the number of MIMO layers, since the information from other MIMO layers L is used. 2 and L 3 The average function of the incoming information (i.e., the incoming learning projection) can be calculated for the MIMO layer L 2 and L 3 Each layer in performs similar operations.

[0061] Back to Figure 5 According to the second exemplary embodiment, the processor 202 further includes CNN 508-1, CNN 508-2, and CNN 508-3, which are configured to refine the output data of the layer mixers 506-1, 506-2, and 506-3, respectively. More specifically, each of CNN 508-1, CNN 508-2, and CNN 508-3 can iteratively refine the corresponding input data until a certain user-defined stop condition is met. If met, the individual MIMO layer processing can be stopped; otherwise, the next iteration can be initiated. At the end of each individual MIMO layer processing, each of CNN 508-1, CNN 508-2, and CNN 508-3 outputs a final estimation set for one of the MIMO layers. Then, the estimation set for the first MIMO layer L can be generated. 1 The final estimated set of is fed to the demapper 510 to be converted into a set of bit LLRs; it can be used for the second MIMO layer L 2 and the third MIMO layer L 3 The same operation is performed on the final estimated set (in order not to Figure 5 Overload, not shown for the second MIMO layer L2 and the third MIMO layer L 3 204). A demapper for each MIMO layer in the processor 200 may be provided, but according to a second exemplary embodiment, such a demapper is intended to be included in the processor 202).

[0062] It should be noted that the combination of the learning projection block 504-1, the layer mixer 506-1, the CNN 508-1 and the demapper 510 forms the first MIMO layer L 1 Layer association block for the second MIMO layer L 2 and the third MIMO layer L 3 A layer of associated blocks is formed by similar combinations of blocks.

[0063] In addition, the preprocessing block, equalization block, layer mixer, CNN, and demapper are all trained blocks that run separately for each MIMO layer while sharing their trained weights across MIMO layers.

[0064] As mentioned above, MIMO layers are processed separately independent of their total number. In view of this, one can implement the above-mentioned early exit by changing the number of iterations for each MIMO layer according to the stopping condition.

[0065] Now let us consider a non-constrained approach to implement the early exit model. The CE loss function for training the ML-based receiver 200 can be written as follows: in is the predicted bit probability after the kth iteration, and b is the ground truth (sent) bit. For the remainder, the CE loss function can be written as where D is the set of indices corresponding to resource elements carrying data, #D is the number of such indices, and B is the number of samples in the sample batch. yes Where L ijl is the LLR (output of each layered block). A maximum number of iterations N is trained so that this number of iterations is always performed during training.

[0066] As mentioned before, the iterative processing of each MIMO layer can be stopped individually when the stopping condition is met. For multi-user MIMO scenarios, the ideal stopping condition for a specific MIMO layer can be given by a low-density parity check (LDPC) decoder: LDPC can be performed after each iteration, and the iteration for the MIMO layer can be stopped after the first successful decoding. However, this approach is not possible in practice because LDPC is too expensive in terms of latency. Therefore, a proxy model of the remaining iterations required for the prediction given by the NN is used to approximate the stopping condition. In this case, a NN is applied for each MIMO layer. For example, such a NN can be as follows: Given: P is the probability of LDPC success required by the system (eg, 0.95). enter: X stop is the concatenation of the outputs of the layer-association blocks of the ML model for the ML layer after the last iteration (LLR estimates). Output: i^ is the estimated number of iterations required until the P success rate of LDPC coding for the current MIMO layer is achieved. Architecture: k is the number of ResNet CNN layers, followed by t fully connected layers.

[0067] The training data can be generated by simulating data. The ground truth / label for each sample can be obtained by running the ML-based receiver 200 for a full number of iterations and performing LDPC after each iteration. The label is the first iteration index where LDPC succeeds.

[0068] Note that for single-user MIMO scenarios, LDPC needs to succeed jointly for all MIMO layers, but it is still possible to stop processing some MIMO layers earlier and only utilize LLRs from all MIMO layers in the decoding stage.

[0069] It is also possible to apply a stopping condition during the inference phase. In this case, the NN proposed above evaluates the stopping condition for each MIMO layer after each iteration and outputs the estimated remaining iterations before decoding is successful. The trained variable depth model is used as follows: Given: P is the threshold probability (e.g., 95%), indicating what percentage of success is expected; n is the maximum number of iterations; ML model: training NN including layer association blocks; NN stop is the trained stop NN that outputs the remaining iterations; conduct: 1. Run the trained NN; 2. For each MIMO layer at each iteration k in range(n): a. Run the layer-associated block once using the demapper and output LLR estimates; b. Let M = M + 1; c. If M>N, stop the iteration process; d. Otherwise, concatenate the output of the layer association block and the LLR estimate to X stop and through NN stop To feed, the NN stop Output the predictions for the remaining iterations i^; e. If i^==0, stop the iteration process; f. If (k+i^)>n, stop the iteration process because there is no hope.

[0070] Once the iterative process is stopped, the estimated LLRs are fed to the LDPC decoder to find the final result.

[0071] Figure 7 2 shows a block diagram of a processor 202 included in a receiver 200 according to a third example embodiment. In the third example embodiment, the processor 202 includes a set of pre-processing blocks 700 and a set of equalization blocks 702, which may again be implemented in the same or similar manner as the set of pre-processing blocks 400 and the set of equalization blocks 402, respectively. Figure 7 As shown, the processor 202 also includes learning projection blocks 704-1, 704-2, and 704-3, each of which can again be configured as a block 106 of the ML-based receiver 100. Unlike the second exemplary embodiment, in the third exemplary embodiment, the processor 202 does not include any layer mixer. This is because the set 702 of equalization blocks includes an LMMSE-based equalization block, and it is found that its mixing operation is sufficient to exclude any dependence on the number of MIMO layers from the processing of additional MIMO layers. Therefore, when using an LMMSE-based equalization block, the use of a layer mixer does not necessarily bring benefits. The learning projections obtained by blocks 704-1, 704-2, and 704-3 are directly fed to CNN 706-1, CNN 706-2, and CNN 706-3, respectively. CNN 706-1, CNN 706-2, and CNN 706-3 can be implemented in the same or similar manner as CNN 508-1, CNN 508-2, and CNN 508-3, respectively. In other words, CNN 706-1, CNN 706-2, and CNN 706-3 are configured to receive the learning projections (i.e., the intermediate estimation sets of appropriate sizes) and, in response, output the learning projections for the MIMO layer L 1 , L2 and L 3 The final estimate set output by CNN 706-1 is converted into bit LLRs by demapper 708-1, the final estimate set output by CNN 706-2 is converted into bit LLRs by demapper 708-2, and the final estimate set output by CNN 706-3 is converted into bit LLRs by demapper 708-3. Each of demapper 708-1, demapper 708-2, and demapper 708-3 may be implemented in the same or similar manner as demapper 510. In the third exemplary embodiment, the first MIMO layer L 1 The layer association block is formed by a combination of a learning projection block 704-1, a CNN 706-1 and a demapper 708-1 for the second MIMO layer L 2 The layer association block is formed by a combination of a learning projection block 704-2, a CNN 706-2 and a demapper 708-2 and is used for the third MIMO layer L 3 The layer association block is formed by a combination of a learning projection block 704-3, a CNN 706-3 and a demapper 708-3.

[0072] Fig. 8A and Figure 8B Two different methods for training each layer association block of an ML model according to an example embodiment are schematically shown. More specifically, Fig. 8A A first (iterative) approach is shown, where each next training iteration uses the intermediate weights obtained on the previous training iteration, and Figure 8B A second (expanded) approach is shown, where each training iteration uses a different set of intermediate weights. For simplicity, by using a layer association block formed by a combination of a learning projection block 704-1, a CNN 706-1, and a demapper 708-1, Fig. 8A and Figure 8B The first MIMO layer L is shown in 1 Regarding the remaining (second and third) MIMO layers L 2 and L 3 The iteration and expansion of the processing can be timed in the same way, but with the help of other layer association blocks (see Figure 7 ).

[0073] Fig.9A and Fig. 9B shows the result of performing the operation without a layer mixer ( Fig.9A ) and with a layer mixer ( Fig. 9B) versus the signal to interference and noise ratio (SINR) in dB using MRC-based equalization for two cases of . The dashed curve with diamond points corresponds to a conventional non-ML-based receiver (hereinafter referred to as the baseline), and the solid curve with circles corresponds to the ML-based receiver 200 using MRC-based equalization with and without a layer mixer (similar to the layer mixer 600). All curves are obtained using a single pilot signal. It should be noted that these and other (discussed below) curves have been obtained using a time domain channel model using data generated by 5GMax or Sionna. If not otherwise stated, the MIMO setting is 16x4 and the modulation scheme used is 256QAM. Since the MRC-based equalization only utilizes information from the current MIMO layer (i.e., no mixing operation is provided by using MRC-based equalization), a layer mixer is required to improve the BER.

[0074] Fig. 10A and Fig. 10B shows the result of performing the operation without a layer mixer ( Fig. 10A ) and with a layer mixer ( Fig. 10B ). Again, the dashed curve with diamond points corresponds to the baseline, and the solid curve with circles corresponds to the ML-based receiver 200 using LMMSE-based equalization with and without a layer mixer (similar to layer mixer 600). As previously mentioned, LMMSE-based equalization already involves some mixing of all MIMO layers in the computation of the inverse. Therefore, the layer mixer may not be used in this case. This is achieved by Fig. 10A and Fig. 10B The curves in Figure 2 show that layer mixers do not bring additional performance because they Fig.9A and Fig. 9B It should be noted that the results are different for, e.g., different interference configurations, where the layer mixer may provide further gain.

[0075] Fig.11A and Fig. 11B The uncoded BER and SINR in dB are shown for two different MIMO settings, namely: 16x4 MIMO setting ( Fig.11A ) and 16x2 MIMO settings ( Fig. 11B). In both MIMO settings, 5GMax generated data and 256QAM have been used. The dashed curve with diamond points corresponds to a baseline using a single pilot signal, the solid curve with square points corresponds to a baseline using a known MIMO channel (i.e., known channel characteristics), and the solid curve with circle points corresponds to a receiver 200 and a layer mixer (similar to layer mixer 600) using a single pilot signal. The ML-based receiver 200 (i.e., the ML model used therein) was initially trained with respect to a 16x4 MIMO setting and has then been applied (without training) for a setting with smaller MIMO layers (i.e., a 16x2 MIMO setting). It can be seen that the ML-based receiver 200 works well for both MIMO settings, confirming its independence from the number of MIMO layers.

[0076] It should be noted that each step or operation of method 300 or any combination of steps or operations can be implemented by various means, such as hardware, firmware and / or software. As an example, one or more of the above steps or operations can be embodied by processor executable instructions, data structures, program modules and other suitable data representations. In addition, the processor executable instructions embodying the above steps or operations can be stored on a corresponding data carrier and executed by processor 202. The data carrier can be implemented as any computer-readable storage medium, which is configured to be readable by the at least one processor to execute the processor executable instructions. Such computer-readable storage media may include volatile and non-volatile media, removable and non-removable media. As an example and not a limitation, computer-readable media include media implemented in any method or technology suitable for storing information. In more detail, actual examples of computer-readable media include, but are not limited to, information transfer media, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD), holographic media or other optical disk storage, magnetic tape, cassette, disk storage and other magnetic storage devices.

[0077] Although example embodiments of the present disclosure are described herein, it should be noted that any of the various changes and modifications may be made in the embodiments of the present disclosure without departing from the scope of legal protection defined by the appended claims. In the appended claims, the word "comprising" does not exclude other elements or operations, and the indefinite article "a" or "an" does not exclude a plurality. The fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

Claims

1. A machine learning (ML) based receiver in a wireless communication network, comprising: at least one processor; as well as at least one memory storing instructions that, when executed by the at least one processor, cause the ML-based receiver to at least: receiving an array of symbols at a plurality of MIMO layers associated with a multiple-input multiple-output MIMO channel; receiving a set of reference signals on the plurality of MIMO layers; Based on the reference signal set and the symbol array, obtaining channel information indicating a state of the MIMO channel; as well as For each MIMO layer in the plurality of MIMO layers: Obtaining an intermediate estimation set for the symbol array by performing an equalization operation based on the channel information and the symbol array; as well as Obtaining a final set of estimates for the symbol array by using a layer association block of an ML model, the layer association block having a set of weights and configured to receive the intermediate set of estimates as input data and output the final set of estimates; The layer association blocks of the ML model are further configured to operate in parallel and exchange the weight sets with each other during a training phase of the ML model.

2. The ML-based receiver of claim 1 , wherein each of the layer association blocks of the ML model is further configured to receive the input data from other layer association blocks of the ML model as a combined value independent of the number of MIMO layers in the plurality of MIMO layers.

3. The ML-based receiver of claim 2, wherein the combined value comprises at least one of: an arithmetic mean of the input data of the other layer associated blocks of the ML model; the sum of the input data of the other layer associated blocks of the ML model; The difference of the input data of the other layer associated blocks of the ML model; The product of the input data of the other layer association block of the ML model; as well as A value obtained by concatenating the input data of the associated blocks of the other layers of the ML model.

4. The ML-based receiver according to any one of claims 1 to 3, wherein each of the layer association blocks of the ML model is further configured to: iteratively processing the input data until a user-defined stopping condition is met; and If the user-defined stopping condition is met, the processing result obtained in the last iteration is output as the final estimated value set. 5 . The ML-based receiver of claim 4 , wherein each of the layer association blocks of the ML model is further configured to use a different set of intermediate weights at each iteration.

6. The ML-based receiver according to any one of claims 1 to 3, wherein each of the layer-associated blocks of the ML model comprises at least one of: a convolutional neural network (CNN), a transformer neural network (TNN), and a self-attention-based neural network.

7. The ML-based receiver according to any one of claims 1 to 3, wherein the equalization operation is a Linear Minimum Mean Square Error (LMMSE) based equalization or a Maximum Ratio Combining (MRC) based equalization.

8. A method for operating a machine learning (ML) based receiver in a wireless communication network, comprising: receiving an array of symbols at a plurality of MIMO layers associated with a multiple-input multiple-output MIMO channel; receiving a set of reference signals on the plurality of MIMO layers; Based on the reference signal set and the symbol array, obtaining channel information indicating a state of the MIMO channel; as well as For each MIMO layer in the plurality of MIMO layers: Obtaining an intermediate estimation set for the symbol array by performing an equalization operation based on the channel information and the symbol array; as well as Obtaining a final set of estimates for the symbol array by using a layer association block of an ML model, the layer association block having a set of weights and configured to receive the intermediate set of estimates as input data and output the final set of estimates; The layer association blocks of the ML model are further configured to operate in parallel and exchange the weight sets with each other during a training phase of the ML model.

9. The method of claim 8, wherein each of the layer association blocks of the ML model is further configured to receive the input data from other layer association blocks of the ML model as a combined value independent of the number of MIMO layers in the plurality of MIMO layers.

10. The method of claim 9, wherein the combined value comprises at least one of the following: an arithmetic mean of the input data of the other layer associated blocks of the ML model; the sum of the input data of the other layer associated blocks of the ML model; The difference of the input data of the other layer associated blocks of the ML model; The product of the input data of the other layer association block of the ML model; as well as A value obtained by concatenating the input data of the associated blocks of the other layers of the ML model.

11. The method according to any one of claims 8 to 10, wherein each of the layer association blocks of the ML model is further configured to: iteratively processing the input data until a user-defined stopping condition is met; and If the user-defined stopping condition is met, the processing result obtained in the last iteration is output as the final estimated value set.

12. The method of claim 11, wherein each of the layer association blocks of the ML model is further configured to use a different set of intermediate weights at each iteration.

13. The method according to any one of claims 8 to 10, wherein each of the layer-associated blocks of the ML model comprises at least one of the following: a convolutional neural network (CNN), a transformer neural network (TNN), and a self-attention-based neural network.

14. The method according to any one of claims 8 to 10, wherein the equalization operation is a linear minimum mean square error (LMMSE) based equalization or a maximum ratio combining (MRC) based equalization.

15. A computer program product comprising a computer readable storage medium, wherein the computer readable storage medium stores computer code which, when executed by at least one processor, causes the at least one processor to perform the method according to any one of claims 8 to 14.

Images