Devices and methods for channel estimation using a transformer ML model

Abstract

A wireless communication apparatus (10) is disclosed for communication with a further wireless communication apparatus (20) via a multi-dimensional channel (15), wherein each channel dimension is associated with a physical characteristic of the multi-dimensional channel (15). The wireless communication apparatus (10) comprises processing circuitry (11) configured to implement a transformer machine learning, ML, module for channel estimation of the multi-dimensional channel (15), wherein the transformer ML module comprises a plurality of processor ML modules. Each processor ML module is associated with a different channel dimension and configured to determine channel information associated with a respective channel dimension based on input data. The plurality of processer ML modules comprise one or more attention ML modules and / or one or more ID convolutional ML modules. Moreover, the transformer ML module comprises an aggregation module configured to determine an output based on the channel information provided by the plurality of processor ML modules.

Description

[0001] DEVICES AND METHODS FOR CHANNEL ESTIMATION USING A TRANSFORMER ML MODEL

[0002] TECHNICAL FIELD

[0003] The present invention relates to telecommunication and data processing. More specifically, the present invention relates to devices and methods for channel estimation using a transformer machine learning, ML, model, in particular a transformer ML model with multi-domain self-attention.

[0004] BACKGROUND

[0005] Channel estimation is a critical process, for instance, in Multiple-Input Multiple-Output (MIMO) systems, which are widely used in modem wireless communications to enhance data throughput, reliability, and spectral efficiency. MIMO systems employ multiple antennas at both the transmitter and receiver ends, allowing the system to exploit spatial diversity and spatial multiplexing. Channel estimation typically uses pilot-based techniques, where known symbols, referred to as pilots, are transmitted from the sender to the receiver. The receiver then compares the received pilots with the expected values to estimate the channel matrix. Common methods for this estimation include Least Squares (LS) and Minimum Mean Square Error (MMSE). LS estimation is straightforward and computationally efficient but may be less accurate in noisy environments. In contrast, MMSE estimation provides better accuracy by considering noise statistics and prior channel knowledge, though it involves more complex calculations. These channel estimation methods rely on certain assumptions, such as Gaussian channel characteristics and stationarity. However, MIMO channels are multidimensional and subject to variations over time, frequency, space, and the like as well as issues like noise and interference. Consequently, the accuracy of channel estimates using these mathematical methods depends on how closely the actual channel aligns with the assumptions of the mathematical model.

[0006] To tackle this problem, deep learning models have been incorporated into the channel estimation process for MIMO systems. Deep neural networks have proven to be exceptionally adept at identifying complex patterns and relationships in data. This makes them particularly effective for estimating the detailed channel characteristics in MIMO systems. Unlike traditional approaches, deep learning-based channel estimation can utilize extensive datasets to understand the channel's statistical properties, which helps reduce dependence on pilot symbols and enhances estimation accuracy, especially in difficult conditions.

[0007] In machine learning, ML, a transformer ML model (or short “transformer”) is a deep learning architecture that is based on an attention mechanism and has been originally developed for natural language processing such as natural language translation by a machine. The functional principle of a transformer is to learn relationships between terms of the sequential input data consisting of tokens which are subjected to a given learning task, for instance language translation, text classification, or text generation. Specifically, the said “relationship” is expressed by the transformer in terms of what is called “attention score”. This score is the essence of information routed and processed inside the transformer. The attention score is computed by a component called self-attention layer.

[0008] SUMMARY

[0009] It is an object of the invention to provide improved devices and methods for channel estimation using a transformer machine learning, ML, model, in particular a transformer ML model with multi-domain self-attention. The foregoing and other objects are achieved by the subject matter of the independent claims. Further implementation forms are apparent from the dependent claims, the description and the figures.

[0010] According to a first aspect a wireless communication apparatus is provided for communication with a further wireless communication apparatus via a multi-dimensional communication channel, wherein each channel dimension is associated with a physical characteristic of the multi-dimensional channel. The wireless communication apparatus according to the first aspect, which may be a MIMO wireless communication apparatus, comprises processing circuitry configured to implement a transformer machine learning, ML, module for channel estimation of the multi-dimensional channel. The transformer ML module comprises a plurality of processor ML modules, wherein each processor ML module is associated with a different channel dimension of the multi-dimensional channel and is configured to determine channel information associated with a respective channel dimension based on input data and wherein the plurality of processer ML modules comprise one or more attention ML modules and / or one or more ID convolutional ML modules. Furthermore, the transformer ML module of the wireless communication apparatus according to the first aspect comprises an aggregation module configured to determine an output based on the channel information provided by the plurality of processor ML modules.

[0011] Thus, a wireless communication apparatus is provided allowing for improved channel estimation using a transformer ML model, in particular a transformer ML model with multi-domain self-attention. By incorporating the multiple different domains of the multi-dimensional channel into the transformer ML module and the self-attention mechanism thereof and processing these in parallel, the wireless communication apparatus according to the first aspect allows to more accurately capture the intricate and multi-dimensional characteristics of the channel, which are often highly interrelated and complex. This multi-domain approach allows the transformer ML module of the wireless communication apparatus according to the first aspect to process and integrate information across different domains simultaneously, leading to a more comprehensive extraction of the channel information. As a result, the ability of the wireless communication apparatus according to the first aspect to correct estimated channels is significantly enhanced, leading to improved accuracy and reliability in channel estimation. This advancement not only boosts the performance of MIMO systems but also sets the stage for more robust and adaptable communication systems capable of operating effectively in diverse and challenging environments.

[0012] In a further possible implementation form, each of the one or more attention ML modules is configured to determine the channel information associated with the respective channel dimension by determining one or more attention scores for one or more pairs of a plurality of input data tokens of the input data.

[0013] In a further possible implementation form, at last one of the one or more attention ML modules is associated with a time dimension of the multi-dimensional channel, wherein each input data token of the plurality of input data tokens is associated with a different time step.

[0014] In a further possible implementation form, at least one of the one or more attention ML modules is associated with a frequency dimension of the multi-dimensional channel, wherein each input data token of the plurality of input data tokens is associated with a different frequency band.

[0015] In a further possible implementation form, at least one of the one or more attention ML modules is associated with a communication beam dimension of the multi-dimensional channel, wherein each input data token of the plurality of input data tokens is associated with a different communication beam. In a further possible implementation form, each of the one or more ID convolutional ML modules is configured to determine the channel information associated with the respective channel dimension by applying a plurality of ID convolutional layers to the input data.

[0016] In a further possible implementation form, at least one of the one or more ID convolutional ML modules is associated with a time dimension of the multi-dimensional channel, wherein a stride of the plurality ID convolutional layers of the ID convolutional ML module corresponds to the time dimension of the multi-dimensional channel.

[0017] In a further possible implementation form, at least one of the one or more ID convolutional ML modules is associated with a frequency dimension of the multi-dimensional channel, wherein a stride of the plurality ID convolutional layers of the ID convolutional ML module corresponds to the frequency dimension of the multi-dimensional channel.

[0018] In a further possible implementation form, at least one of the one or more ID convolutional ML modules is associated with a communication beam dimension of the multi-dimensional channel, wherein a stride of the plurality ID convolutional layers of the ID convolutional ML module corresponds to the communication beam dimension of the multi-dimensional channel.

[0019] In a further possible implementation form, the aggregation module is configured to determine the output based on the channel information provided by the plurality of processor ML modules using a summation, a concatenation or an aggregation operation of the channel information provided by the plurality of processor ML modules.

[0020] In a further possible implementation form, the plurality of processor ML modules comprises one or more full-precision processor ML modules and / or one or more Boolean processor ML modules.

[0021] In a further possible implementation form, the wireless communication apparatus according to the first aspect is a base station, in particular a MIMO base station, and the further wireless communication apparatus is a user equipment, UE, in particular a MIMO UE.

[0022] According to a second aspect a method is provided for operating a wireless communication apparatus for communication with a further wireless communication apparatus via a multi-dimensional communication channel, wherein each channel dimension is associated with a physical characteristic of the multi-dimensional channel. The wireless communication apparatus comprises processing circuitry configured to implement a transformer machine learning, ML, module for channel estimation of the multi-dimensional channel. The method according to the second aspect comprises the following steps implemented by the transformer ML module:

[0023] implementing a plurality of processor ML modules, wherein each processor ML module is associated with a different channel dimension and configured to determine channel information associated with a respective channel dimension based on input data, wherein the plurality of processer ML modules comprise one or more attention ML modules and / or one or more ID convolutional ML modules; and

[0024] determining an output based on the channel information provided by the plurality of processor ML modules.

[0025] The method according to the second aspect can be performed by the wireless communication apparatus according to the first aspect. Thus, further features of the method according to the second aspect result directly from the functionality of the wireless communication apparatus according to the first aspect and its different implementation forms described above and below. According to a third aspect a computer program or a computer program product is provided, comprising a computer-readable storage medium carrying program code which causes a computer or a processor to perform the method according to the second aspect when the program code is executed by the computer or the processor.

[0026] The different aspects of the invention can be implemented in software and / or hardware.

[0027] Details of one or more embodiments are set forth in the accompanying drawings and the description below. Other features, objects, and advantages will be apparent from the description, drawings, and claims.

[0028] BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In the following embodiments of the invention are described in more detail with reference to the attached figures and drawings, in which:

[0030] Fig. 1 is a schematic diagram illustrating a wireless communication apparatus and a further wireless communication apparatus for multi-domain channel estimation using a transformer ML model according to an embodiment;

[0031] Fig. 2 is a schematic diagram illustrating a transformer module of a wireless communication apparatus according to an embodiment with a plurality of processor ML modules for multi-domain channel estimation;

[0032] Fig. 3 is a schematic diagram illustrating processing steps implemented by a wireless communication apparatus according to an embodiment for multi-domain channel estimation;

[0033] Fig. 4a is a schematic diagram illustrating a transformer module of a wireless communication apparatus according to an embodiment with a plurality of attention ML modules for multi-domain channel estimation;

[0034] Fig. 4b is a schematic diagram illustrating the architecture of an attention ML module of a transformer module of a wireless communication apparatus according to an embodiment;

[0035] Fig. 5 is a schematic diagram illustrating a transformer module of a wireless communication apparatus according to an embodiment with a plurality of ID convolutional ML modules for multi-domain channel estimation;

[0036] Fig. 6 is a schematic diagram illustrating a transformer module of a wireless communication apparatus according to an embodiment with a plurality of attention ML modules and a plurality of ID convolutional ML modules for multi-domain channel estimation;

[0037] Fig. 7a is a schematic diagram illustrating a transformer module of a wireless communication apparatus according to an embodiment with a plurality of Boolean processor ML modules for multi-domain channel estimation;

[0038] Fig. 7b is a schematic diagram illustrating the architecture of an Boolean attention ML module of a transformer module of a wireless communication apparatus according to an embodiment;

[0039] Fig. 7c is a schematic diagram illustrating the architecture of an Boolean attention ML module of a transformer module of a wireless communication apparatus according to a further embodiment; and Fig. 8 is a flow diagram illustrating a method according to an embodiment for operating a wireless communication apparatus for multi-domain channel estimation using a transformer ML model.

[0040] In the following identical reference signs refer to identical or at least functionally equivalent features.

[0041] DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] In the following description, reference is made to the accompanying figures, which form part of the disclosure, and which show, by way of illustration, specific aspects of embodiments of the invention or specific aspects in which embodiments of the present invention may be used. It is understood that embodiments of the invention may be used in other aspects and comprise structural or logical changes not depicted in the figures. The following detailed description, therefore, is not to be taken in a limiting sense, and the scope of the present invention is defined by the appended claims.

[0043] For instance, it is to be understood that a disclosure in connection with a described method may also hold true for a corresponding device or system configured to perform the method and vice versa. For example, if one or a plurality of specific method steps are described, a corresponding device may include one or a plurality of units, e.g. functional units, to perform the described one or plurality of method steps (e.g. one unit performing the one or plurality of steps, or a plurality of units each performing one or more of the plurality of steps), even if such one or more units are not explicitly described or illustrated in the figures. On the other hand, for example, if a specific apparatus is described based on one or a plurality of units, e.g. functional units, a corresponding method may include one step to perform the functionality of the one or plurality of units (e.g. one step performing the functionality of the one or plurality of units, or a plurality of steps each performing the functionality of one or more of the plurality of units), even if such one or plurality of steps are not explicitly described or illustrated in the figures. Further, it is understood that the features of the various exemplary embodiments and / or aspects described herein may be combined with each other, unless specifically noted otherwise.

[0044] Figure 1 is a schematic diagram illustrating a wireless communication apparatus 10 according to an embodiment and a further wireless communication apparatus 20 configured for communication via a multi-dimensional communication channel 15 and for multi-domain channel estimation using a transformer ML module 100, as will be described in more detail in the following under further reference to figure 2. The communication channel 15 is multi-dimensional in that each channel dimension may be associated with a different physical characteristic of the multi-dimensional channel 15, such as time, frequency, spatial beam, and the like.

[0045] In an embodiment, the wireless communication apparatus 10 may be implemented as a MIMO wireless communication apparatus 10, for instance, in the form of a mobile network base station, a user equipment, UE, a Wi-Fi access point, a Wi-Fi station, and the like. Likewise, the further wireless communication apparatus 20 may be implemented as a MIMO wireless communication apparatus 20, for instance, in the form of a mobile network base station, a user equipment, UE, a Wi-Fi access point, a Wi-Fi station, and the like.

[0046] As indicated in figure 1, the wireless communication apparatus 10 may comprise processing circuitry 11 and a communication interface 13, in particular a wireless communication interface 13 enabling communication in accordance, for instance, with the 3GPP framework of standards or the IEEE 802.11 framework of standards over the channel 15. The processing circuitry 11 may be implemented in hardware and / or software and may comprise digital circuitry, or both analog and digital circuitry. Digital circuitry may comprise components such as application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), or general-purpose processors. The wireless communication apparatus 10 may further comprise a memory or storage 15 configured to store executable program code which, when executed by the processing circuitry 11, causes the wireless communication apparatus 10 to perform the functions and methods described herein.

[0047] Likewise, the further wireless communication apparatus 20 may comprise processing circuitry 21 and a communication interface 23, in particular a wireless communication interface 23 enabling communication in accordance, for instance, with the 3GPP framework of standards or the IEEE 802.11 framework of standards over the channel 15. The processing circuitry 21 may be implemented in hardware and / or software and may comprise digital circuitry, or both analog and digital circuitry. Digital circuitry may comprise components such as application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), or general-purpose processors. The further wireless communication apparatus 20 may further comprise a memory or storage 25 configured to store executable program code which, when executed by the processing circuitry 21, causes the further wireless communication apparatus 20 to perform the functions and methods described herein.

[0048] As illustrated in figure 2, the wireless communication apparatus 10 comprises a transformer machine learning, ML, module 100 for channel estimation of the multi-dimensional channel 15. In an embodiment, the transformer ML module 100 may be implemented, i.e. operated by the processing circuitry 11 of the wireless communication apparatus 10. The transformer ML module 100 comprises a plurality of processor ML modules 1 lOa-m, wherein each processor ML module 1 lOa-m is associated with a different channel dimension (such as time, frequency, spatial stream, and the like) and configured to determine channel information associated with a respective channel dimension based on input data. As will be described in more detail below, the plurality of processer ML modules 1 lOa-m comprise one or more attention ML modules (as illustrated, for instance, in figure 4a) and / or one or more ID convolutional ML modules (s illustrated, for instance, in figure 5).

[0049] Moreover, the transformer ML module 100 of the wireless communication apparatus 10 comprises an aggregation module 120 configured to determine an output based on the channel information provided by the plurality of processor ML modules 1 lOa-m. In an embodiment, the aggregation module 120 may be configured to determine the output based on the channel information provided by the plurality of processor ML modules 1 lOa-m using a summation, a concatenation or an aggregation operation of the channel information provided by the plurality of processor ML modules 1 lOa-m.

[0050] Figure 3 is a schematic diagram illustrating processing steps implemented by the wireless communication apparatus 10 according to an embodiment for multi-domain channel estimation. In a step 301 one or more pilot signals are received, which are used in a step 303 by the wireless communication apparatus 10 to determine first channel estimates using, for instance, a predefined mathematical model for the channel 15. In an embodiment, the wireless communication apparatus 10 is configured in a step 305 to refine these first estimates using the transformer ML module 100 of figure 2.

[0051] Figure 4a is a schematic diagram illustrating an embodiment of the transformer ML module 100 of the wireless communication apparatus 10, wherein the plurality of processor ML modules 1 lOa-m are implemented as a plurality of attention ML modules 1 lOa-m for multi-domain channel estimation. In an embodiment, each of the plurality of attention ML modules 1 lOa-m is configured to determine the channel information associated with the respective channel dimension by determining one or more attention scores for one or more pairs of a plurality of input data tokens of the input data.

[0052] Figure 4b is a schematic diagram illustrating the architecture of the exemplary attention ML module 110a of the transformer ML module 100 according to an embodiment. As can be taken from figure 4b, the exemplary attention ML module 110a comprises a projector layer 111, 112, 113 in the form of a learnable deep learning layer that computes information, such as query information Q, key information K, value information V and O from the input data. In an embodiment, the projector layer 111, 112, 113 may be implemented by a fully connected layer (also referred to as linear layer). The output Q, K from the projector layers 111, 112 is used by a multiplication block 114 and a subsequent softmax function block 115. The output of the softmax function block 115 is combined with the output V from the projector layer 113 by another multiplication block 116. The output from the multiplication block 116 is processed by the projector layer 117 to provide the final output.

[0053] In an embodiment, at least one of the plurality of attention ML modules 1 lOa-m illustrated in figure 4a may be associated with and trained for a time dimension of the multi-dimensional channel 15, wherein each input data token of the plurality of input data tokens is associated with a different time step. In a further embodiment, at least one of the plurality of attention ML modules 1 lOa-m illustrated in figure 4a may be associated with and trained for a frequency dimension of the multidimensional channel 15, wherein each input data token of the plurality of input data tokens is associated with a different frequency band. In a further embodiment, at least one of the plurality of attention ML modules 1 lOa-m illustrated in figure 4a may be associated with and trained for a communication beam dimension (also referred to as spatial beam dimension) of the multi-dimensional channel 15, wherein each input data token of the plurality of input data tokens is associated with a different communication beam.

[0054] Figure 5 is a schematic diagram illustrating a further embodiment of the transformer ML module 100 of the wireless communication apparatus 10, wherein the plurality of processor ML modules 1 lOa-m are implemented as a plurality of ID convolutional ML modules 1 lOa-m for multi-domain channel estimation and wherein each ID convolutional ML module 1 lOa-m is configured to determine the channel information associated with the respective channel dimension by applying a plurality of ID convolutional layers to the input data.

[0055] In an embodiment, at least one of the plurality of ID convolutional ML modules 1 lOa-m illustrated in figure 5 may be associated with and trained for a time dimension of the multi-dimensional channel 15, wherein a stride of the plurality ID convolutional layers of the ID convolutional ML module corresponds to the time dimension of the multi-dimensional channel 15. In a further embodiment, at least one of the plurality of ID convolutional ML modules 1 lOa-m illustrated in figure 5 may be associated with and trained for a frequency dimension of the multi-dimensional channel, wherein a stride of the plurality ID convolutional layers of the ID convolutional ML module corresponds to the frequency dimension of the multi-dimensional channel 15. In a further embodiment, at least one of the plurality of 1 D convolutional ML modules 11 Oa-m illustrated in figure 5 may be associated with and trained for a communication beam (i.e. spatial beam) dimension of the multi-dimensional channel 15, wherein a stride of the plurality ID convolutional layers of the ID convolutional ML module corresponds to the communication beam dimension of the multi-dimensional channel 15.

[0056] Figure 6 is a schematic diagram illustrating a further embodiment of the transformer ML module 100 of the wireless communication apparatus 10, wherein the plurality of processor ML modules 1 lOa-m are implemented as a mixture of a plurality of attention ML modules and a plurality of ID convolutional ML modules for multi-domain channel estimation.

[0057] Figure 7a is a schematic diagram illustrating a further embodiment of the transformer ML module 100 of the wireless communication apparatus 10, wherein the plurality of processor ML modules 11 Oa-m are implemented as Boolean processor ML modules for multi-domain channel estimation. Thus, according to embodiments disclosed herein, the transformer ML module 100 of the wireless communication apparatus 10 may comprises one or more full-precision processor ML modules and / or one or more Boolean processor ML modules. In an embodiment, each processor ML module 1 lOa-m may be defined by a plurality of activations and / or a plurality of parameters, wherein the plurality of activations and / or the plurality of parameters of a full-precision processor ML module are full precision (i.e. real) values and wherein the plurality of activations and / or the plurality of parameters of a Boolean processor ML module are Boolean values, i.e. binary values. Figures 7b and 7c is a schematic diagram illustrating different architectures of a processor ML module 1 lOa-m implemented as a Boolean attention ML module of the transformer ML module 100 of the wireless communication apparatus 10 according to an embodiment. As already described above, the Boolean attention ML module of the transformer ML module 100 has a binary or Boolean architecture of the self-attention module in which at least some data and / or parameters are binary / Boolean numbers consuming 1 bit instead of real- valued, i.e. full precision numbers requiring a multiple-bit format such as 16 bits.

[0058] In the embodiment shown in figure 7b, a binarization scheme is used in which the projectors ofQ 111, K 112, V 113 and O 117 are fully connected layers of binarized weights. Learning binarized weights Wblnis based on the use of continuous latent weights Wrealthat are learnt by the standard gradient descent method, i.e.,

[0059] Wbln= sign(W

[0060]

[0061] real— a), where a = — W / Tal.

[0062] The activations 118 for Q, K, V, and A are all also binarized, resulting in 1 -bit by 1 -bit computation of the dot product blocks. Further details of the possible embodiment of the Boolean attention module 110a illustrated in figure 7b are described in Liu et al. BiT: Robustly Binarized Multi-distilled Transformer. NeurlPS 2022. https: / / arxiv.org / abs / 2205.13016, which is fully incorporated herein by reference. The processing blocks 114, 115 and 116 of the embodiment shown in figure 7b corresponds to the blocks 114, 115 and 116 of the embodiment shown in figure 4b, which has been described above.

[0063] The embodiment of the Boolean attention module 110a shown in figure 7c differs from the embodiment of figure 7a mainly in that (a) a layer normalization 119 is used before binarizing Q, K, and V and (b) the continuous weights and activations are binarized as follows:

[0064] / Wreal

[0065] Wbin= CLIP — — ,-l, 1

[0066] \

[0067] .. / I / X \| \

[0068] Xbln= CLIP — ,-l, 11 + 0.5 x max(|X|)

[0069]

[0070] \L \BX7J )

[0071] Further details of the possible embodiment of the Boolean attention module 110a illustrated in figure 7c are described in Zhang et al. Binarized Neural Machine Translation. NeurlPS 2023. https: / / arxiv.org / abs / 2302.04907, which is fully incorporated herein by reference.

[0072] Figure 8 is a flow diagram illustrating a method 800 for operating the wireless communication apparatus 10 for communication with the further wireless communication apparatus 20 via the multi-dimensional communication channel 15. As alreadv described above, the wireless communication apparatus 10 comprises a transformer ML module 100 for channel estimation of the multi-dimensional channel 15. The method 800 comprises a step 801 by the transformer ML module 100 of implementing, i.e. operating a plurality of processor ML modules 1 lOa-m, wherein, as already described above, each processor ML module is associated with a different channel dimension and configured to determine channel information associated with a respective channel dimension based on input data and wherein the plurality of processer ML modules 1 Warn comprise one or more attention ML modules and / or one or more ID convolutional ML modules. Moreover, the method 800 comprises a step 803 by the transformer ML module 100 of determining an output based on the channel information provided by the plurality of processor ML modules 1 lOa-m, for instance, by means of the aggregation module 120 described above.

[0073] The person skilled in the art will understand that the "blocks" ("units") of the various figures (method and apparatus) represent or describe functionalities of embodiments of the invention (rather than necessarily individual "units" in hardware or software) and thus describe equally functions or features of apparatus embodiments as well as method embodiments (unit = step).

[0074] In the several embodiments provided in the present application, it should be understood that the disclosed system, apparatus, and method may be implemented in other manners. For example, the described apparatus embodiment is merely exemplary. For example, the unit division is merely logical function division and may be other division in actual implementation. For example, a plurality of units or components may be combined or integrated into another system, or some features may be ignored or not performed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections may be implemented by using some interfaces. The indirect couplings or communication connections between the apparatuses or units may be implemented in electronic, mechanical, or other forms.

[0075] The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one position, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.

[0076] In addition, functional units in the embodiments of the invention may be integrated into one processing unit, or each of the units may exist alone physically, or two or more units are integrated into one unit.

Claims

CLAIMS1. A wireless communication apparatus (10) for communication with a further wireless communication apparatus (20) via a multi-dimensional channel (15), each channel dimension associated with a physical characteristic of the multidimensional channel, wherein the wireless communication apparatus (10) comprises a transformer machine learning, ML, module (100) for channel estimation of the multi-dimensional channel (15), wherein the transformer ML module (100) comprises:a plurality of processor ML modules (1 lOa-m), wherein each processor ML module is associated with a different channel dimension and configured to determine channel information associated with a respective channel dimension based on input data, wherein the plurality of processer ML modules (1 lOa-m) comprise one or more attention ML modules and / or one or more ID convolutional ML modules; andan aggregation module (120) configured to determine an output based on the channel information provided by the plurality of processor ML modules (1 lOa-m).

2. The wireless communication apparatus (10) of claim 1, wherein each of the one or more attention ML modules (1 lOa-m) is configured to determine the channel information associated with the respective channel dimension by determining one or more attention scores for one or more pairs of a plurality of input data tokens of the input data.

3. The wireless communication apparatus (10) of claim 2, wherein one of the one or more attention ML modules (11 Oa-m) is associated with a time dimension of the multi-dimensional channel (15) and wherein each input data token of the plurality of input data tokens is associated with a different time step.

4. The wireless communication apparatus (10) of claim 2 or 3, wherein one of the one or more attention ML modules (1 lOa-m) is associated with a frequency dimension of the multi-dimensional channel (15) and wherein each input data token of the plurality of input data tokens is associated with a different frequency band.

5. The wireless communication apparatus (10) of any one of claims 2 to 4, wherein one of the one or more attention ML modules (1 lOa-m) is associated with a communication beam dimension of the multi-dimensional channel (15) and wherein each input data token of the plurality of input data tokens is associated with a different communication beam.

6. The wireless communication apparatus (10) of any one of the preceding claims, wherein each of the one or more ID convolutional ML modules (11 Oa-m) is configured to determine the channel information associated with the respective channel dimension by applying a plurality of ID convolutional layers to the input data.

7. The wireless communication apparatus (10) of claim 6, wherein one of the one or more ID convolutional ML modules (11 Oa-m) is associated with a time dimension of the multi-dimensional channel (15) and wherein a stride of the plurality ID convolutional layers of the ID convolutional ML module corresponds to the time dimension of the multidimensional channel (15).

8. The wireless communication apparatus (10) of claim 6 or 7, wherein one of the one or more ID convolutional ML modules (11 Oa-m) is associated with a frequency dimension of the multi-dimensional channel (15) and wherein a stride of theplurality ID convolutional layers of the ID convolutional ML module corresponds to the frequency dimension of the multidimensional channel (15).

9. The wireless communication apparatus (10) of any one of claims 6 to 8, wherein one of the one or more ID convolutional ML modules (1 lOa-m) is associated with a communication beam dimension of the multi-dimensional channel (15) and wherein a stride of the plurality ID convolutional layers of the ID convolutional ML module corresponds to the communication beam dimension of the multi-dimensional channel (15).

10. The wireless communication apparatus (10) of any one of the preceding claims, wherein the aggregation module (120) is configured to determine the output based on the channel information provided by the plurality of processor ML modules (1 lOa-m) using a summation, a concatenation or an aggregation of the channel information provided by the plurality of processor ML modules (1 lOa-m).

11. The wireless communication apparatus (10) of any one of the preceding claims, wherein the plurality of processor ML modules (1 lOa-m) comprises one or more full-precision processor ML modules and / or one or more Boolean processor ML modules.

12. The wireless communication apparatus (10) of any one of the preceding claims, wherein the wireless communication apparatus (10) is a base station and wherein the further wireless communication apparatus (20) is a user equipment, UE.

13. A method (800) for operating a wireless communication apparatus (10) for communication with a further wireless communication apparatus (20) via a multi-dimensional channel (15), each channel dimension associated with a physical characteristic of the multi-dimensional channel (15), wherein the wireless communication apparatus (10) comprises a transformer machine learning, ML, module (100) for channel estimation of the multi-dimensional channel (15), wherein the method (800) comprises the following steps implemented by the transformer ML module (100):implementing (801 ) a plurality of processor ML modules (11 Oa-m), wherein each processor ML module is associated with a different channel dimension and configured to determine channel information associated with a respective channel dimension based on input data, wherein the plurality of processer ML modules (11 Oa-m) comprise one or more attention ML modules and / or one or more ID convolutional ML modules; anddetermining (803) an output based on the channel information provided by the plurality of processor ML modules (1 lOa-m).

14. A computer program product comprising a computer-readable storage medium for storing program code which causes a computer or a processor to perform the method (800) of claim 13 when the program code is executed by the computer or the processor.

Images