Transformer-based encoding and decoding of wireless channel measurements

Abstract

There is provided techniques for encoding and decoding wireless channel measurements. An encoding method is performed by an information encoder device. The method comprises obtaining wireless channel measurements of a UE as available at Physical Layer. The method further comprises generating a refined representation of the wireless channel measurements using a transformer architecture having an attention component and according to which each available basic element of the wireless channel measurements is transformed into a respective embedding vector. A decoding method is performed by an information decoder device. According to this method, a refined representation of the wireless channel measurements for a UE is obtained from an information encoder device. A further refined representation of the wireless channel measurements for the UE is generated using a transformer architecture having an attention component and according to which the embedding vectors and the refined representation are provided as input.

Description

[0001] P110510W001 1

[0002] ENCODING AND DECODING OF WIRELESS CHANNEL MEASUREMENTS

[0003] TECHNICAL FIELD

[0004] Embodiments presented herein relate to a method, an information encoder device, a computer program, and a computer program product for encoding wireless channel measurements. Embodiments presented herein further relate to a method, an information decoder device, a computer program, and a computer program product for decoding wireless channel measurements.

[0005] BACKGROUND

[0006] In general terms, the quality of a wireless channel can be assessed by processing large number of measurements collected by large number of antenna elements at the network side (such as at a transmission and reception point; TRP) as well as at user equipment (UE). Since processing of this amount of data takes time and requires computational resources, typically some techniques are used to reduce dimensions of the data. This reduction can be implemented through operations such as averaging, data interleaving, quantization, etc. These operations commonly take place in hardware dedicated to physical layer (i.e., network Layer 1) processing and are typically used to estimate the quality of the wireless channel in form of signal to interference plus noise (SINR) values. Processing at other network layers (such as network Layer 2 or Layer 3) can use the SINR values as input to their decision making, for example for scheduling purposes, load balancing, carrier selection, etc.

[0007] As an example, the scheduling for a TRP is implemented in network Layer 2 makes use of SINR values provided for each UE served by the TRP. Other parameters used by the scheduling can be channel state information (CSI) reports, reporting of acknowledgements (ACKs) and negative acknowledgements (NACKs), etc.

[0008] As another example, secondary carrier selection is implemented in network Layer 3 and could benefit from channel quality indicator (CQI) reporting or SINR values. For example, the secondary carrier selection could e.g., benefit from capturing interdependencies between UEs for example in frequency and time.

[0009] In general terms, increasing system performance, such as capacity of network, better quality of service, energy efficient operation, etc. comes at the cost of increasing P110510W001 2 complexity of the (radio) access network system. This also implies the use of more advanced algorithms. These algorithms are often data driven, and can be based on artificial intelligence (Al) and / or machine learning (ML) techniques. In many cases, a large amount of input data describing as precisely as possibly the underlaying behavior of the system in different scenarios is required. The more detailed data is available, the bigger the chance is to capture patterns, such as traffic patterns, mobility patterns, etc.

[0010] For example, AI / ML technologies could potentially assist, or entirely replace, some algorithms that typically used SINR values and UE reported values. But providing existing data sets to ML / AI based algorithms could be a limiting factor. This is because techniques are used to reduce dimensions of the data before the data reaches the ML / AI based algorithms. That is, the data that reaches the ML / AI based algorithms has already been reduced in size and accuracy. This limits the ability of the AI / ML algorithms to discover underlying dependencies and patterns.

[0011] On the other hand, unrestricted transmission of all the data collected at network Layer 1 might be either expensive or even technically impossible due to its volume. Thus, sending all the data measured and being available at network Layer 1 is challenging, since it requires advanced hardware to support it.

[0012] Hence, there is a need for improved processing of data collected at network Layer 1 such that it can be used by data-driven algorithms, such as ML / AI based algorithms.

[0013] SUMMARY

[0014] An object of embodiments herein is to address the above issues.

[0015] A particular object is to find a way to pre-process the data before it is transmitted from network Layer 1.

[0016] According to a first aspect there is presented a method for encoding wireless channel measurements. The method is performed by an information encoder device. The method comprises obtaining wireless channel measurements of a UE as available at Physical Layer. The method comprises generating a refined representation of the wireless channel measurements using a transformer architecture having an attention P110510W001 3 component and according to which each available basic element of the wireless channel measurements is transformed into a respective embedding vector.

[0017] According to a second aspect there is presented an information encoder device for encoding wireless channel measurements. The information encoder device comprises processing circuitry. The processing circuitry is configured to cause the information encoder device to obtain wireless channel measurements of a UE as available at Physical Layer. The processing circuitry is configured to cause the information encoder device to generate a refined representation of the wireless channel measurements using a transformer architecture having an attention component and according to which each available basic element of the wireless channel measurements is transformed into a respective embedding vector.

[0018] According to a third aspect there is presented a computer program for encoding wireless channel measurements. The computer program comprises computer code which, when run on processing circuitry of an information encoder device, causes the information encoder device to perform actions. One action comprises the information encoder device to obtain wireless channel measurements of a UE as available at Physical Layer. One action comprises the information encoder device to generate a refined representation of the wireless channel measurements using a transformer architecture having an attention component and according to which each available basic element of the wireless channel measurements is transformed into a respective embedding vector.

[0019] According to a fourth aspect there is presented a method for decoding wireless channel measurements. The method is performed by an information decoder device. The method comprises obtaining a refined representation of the wireless channel measurements for a UE from an information encoder device. The refined representation of the wireless channel measurements comprises embedding vectors. The method comprises generating a further refined representation of the wireless channel measurements for the UE using a transformer architecture having an attention component and according to which the embedding vectors and the refined representation are provided as input.

[0020] According to a fifth aspect there is presented an information decoder device for decoding wireless channel measurements, the information decoder device comprises P110510W001 4 processing circuitry. The processing circuitry is configured to cause the information decoder device to obtain a refined representation of the wireless channel measurements for a UE from an information encoder device. The refined representation of the wireless channel measurements comprises embedding vectors. The processing circuitry is configured to cause the information decoder device to generate a further refined representation of the wireless channel measurements for the UE using a transformer architecture having an attention component and according to which the embedding vectors and the refined representation are provided as input.

[0021] According to a sixth aspect there is presented a computer program for decoding wireless channel measurements. The computer program comprises computer code which, when run on processing circuitry of an information decoder device, causes the information decoder device to perform actions. One action comprises the information decoder device to obtain a refined representation of the wireless channel measurements for a UE from an information encoder device. The refined representation of the wireless channel measurements comprises embedding vectors. One action comprises the information decoder device to generate a further refined representation of the wireless channel measurements for the UE using a transformer architecture having an attention component and according to which the embedding vectors and the refined representation are provided as input.

[0022] According to a seventh aspect there is presented a computer program product comprising a computer program according to at least one of the third aspect and the sixth aspect and a computer readable storage medium on which the computer program is stored. The computer readable storage medium could be a non-transitory computer readable storage medium.

[0023] Advantageously, these aspects enable algorithms implemented in Physical Layer, Layer 2 and Layer 3 to get access to more data. In turn, this enables these algorithms to benefit from data containing more information.

[0024] Advantageously, these aspects enable data-driven algorithms, such as ML / AI based algorithms, to discover underlying dependencies and patterns. In turn, this enables the network to be controlled based on better observability, leading to increased network performance, in terms of throughput, etc. P110510W001 5

[0025] Advantageously, these aspects enable different types of auxiliary information to be blended with the wireless channel measurements, thereby providing a multidimensional view of the network behaviour.

[0026] Other objectives, features and advantages of the enclosed embodiments will be apparent from the following detailed disclosure, from the attached dependent claims as well as from the drawings.

[0027] Generally, all terms used in the claims are to be interpreted according to their ordinary meaning in the technical field, unless explicitly defined otherwise herein. All references to "a / an / the element, apparatus, component, means, module, step, etc." are to be interpreted openly as referring to at least one instance of the element, apparatus, component, means, module, step, etc., unless explicitly stated otherwise. The steps of any method disclosed herein do not have to be performed in the exact order disclosed, unless explicitly stated.

[0028] BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The inventive concept is now described, by way of example, with reference to the accompanying drawings, in which:

[0030] Fig. 1 is a schematic diagram illustrating a communication network according to embodiments;

[0031] Fig. 2 is a schematic illustration of radio and UE state representation in a MU MIMO scenario according to embodiments;

[0032] Fig. 3 is a schematic illustration of obtaining combined CSR and transforming combined CSR into embedding vectors according to embodiments;

[0033] Fig. 4 is a schematic illustration of a process for concatenating wireless channel measurements from different subcarriers according to embodiments;

[0034] Fig. 5 is a schematic illustration of preparing embedding vectors for a single UE according to embodiments;

[0035] Fig. 6 is a block diagram of a distributed transformer according to embodiments;

[0036] Fig. 7 is a flowchart of methods according to embodiments; P110510W001 6

[0037] Fig. 8 is a block diagram of an information encoder device according to embodiments;

[0038] Fig. 9 is a schematic illustration of a process for updating an embedding vector according to embodiments;

[0039] Fig. 10 is a flowchart of methods according to embodiments;

[0040] Fig. 11 is a block diagram of an information decoder device according to embodiments;

[0041] Fig. 12 is a schematic diagram showing structural units of an information encoder device according to an embodiment;

[0042] Fig. 13 is a schematic diagram showing structural units of an information decoder device according to an embodiment; and

[0043] Fig. 14 shows one example of a computer program product comprising computer readable means according to an embodiment.

[0044] DETAILED DESCRIPTION

[0045] The inventive concept will now be described more fully hereinafter with reference to the accompanying drawings, in which certain embodiments of the inventive concept are shown. This inventive concept may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided by way of example so that this disclosure will be thorough and complete, and will fully convey the scope of the inventive concept to those skilled in the art. Like numbers refer to like elements throughout the description. Any step or feature illustrated by dashed lines should be regarded as optional.

[0046] Fig. 1 is a schematic diagram illustrating a communication network 100 where embodiments presented herein can be applied. The communication network 100 comprises a TRP 120, controlled by a network node 110, that is configured to provide network access to UEs 1300:1300. The network node and / or the TRP could be, be part of, or be integrated with, a (radio) access network node, radio base station, base transceiver station, node B, evolved node B, gNB, access point, access node, P110510W001 7 integrated access and backhaul node, etc. Each of the UEs 1300:1300 could be a portable wireless device, mobile station, mobile phone, handset, wireless local loop phone, smartphone, laptop computer, tablet computer, wireless modem, wireless sensor device, network equipped vehicle, Internet-of-Things device, etc.

[0047] As disclosed above, the quality of the wireless channel between the TRP 120 and the UEs 130a: 130c can be assessed by processing large number of measurements collected by large number of antenna elements at the TRP 120 as well as at the UE 130a: 130c. However, as further disclosed above, sending all the data measured and being available at Layer 1 is challenging, since it requires advanced hardware to support it. Hence, there is a need for improved processing of data collected at Layer 1 such that it can be used by data-driven algorithms, such as ML / AI based algorithms.

[0048] An increasing number of antenna elements used in the TRP, e.g., as a means to increase capacity, reliability and efficiency of transmission channels between the TRP and the UEs, comes with challenges, such as growing complexity of processing taking place in hardware. An increasing number of antenna elements, bandwidth size, number of UE antennas also contribute to increasing the amount of data being available for processing at the baseband part of the hardware. However, this data is often only available in reduced form to other network layers.

[0049] As an example, every Transmission Time Interval (TTI), e.g., every 0.5 ms, thousands of data points describing the channel quality can be collected for a single UE. The quality of the estimation of the wireless channel is one of the fundamental constraints of the system performance. For simplicity and due to hardware capability, the amount of the data is reduced before being sent from Physical layer to other network layers. Simple techniques of dimensionality reduction can be applied in frequency domain and antenna domain. This may cause dependencies and relations as present in the original data to be lost.

[0050] In further detail, the aforementioned data reduction performed before the wireless channel measurements are provided from Physical layer to Layer 2 and Layer 3 causes applications implemented at Layer 2 and Layer 3 to have very limited access to the wireless channel measurements available at Physical Layer. This limited access can be a bottleneck e.g., for radio resource management (RRM) algorithms implemented in Layer 2 and Layer 3. This lack of measurement data could prevent P110510W001 8 data-driven algorithms from having access to more accurate channel and channel- related information, thereby leading to inefficient of resource utilization.

[0051] On the other hand, having unconstrained ability to process all available data collected at Physical layer could potentially enable algorithms implemented at other network layers to provide better performance or enable new classes of algorithms to be used. In this respect, one way to address this challenge is to redefine the type of data that is delivered from / to Physical Layer to / from other layers of the network. In further detail, at least some of the herein disclosed embodiments are based on creating a representation of the wireless channel measurements in a compressed, or at least encoded, way. In addition, the representation can, optionally, incorporate UE reported data, e.g., UE feedback in any form. Such a combination will hereinafter be referred to as a channel and UE state representation (CSR). Thus, the CSR comprises information related to the wireless channel as well as UE specific details. The CSR can be generated either at the UE or the TRP. The size of the CSR can be dynamically adjusted such that it well describes wireless channel measurements and UE feedback data received at the TRP. In any case, the CSR is generated per UE. Fig. 2 illustrates three different examples of Channel and UE State Representation in a MU MIMO scenario. A first representation 200a illustrates channels 210a, 210b, 210c of different UEs being transformed into CSR 220a, 220b, 220c. A second representation 200b illustrates that the dimensionality of channel representation can be changed based on requirements, or preferences, of other algorithms using the CSR. A third representation 200c illustrates that further changes can be imposed so to meet requirements, or preferences, of other algorithms implemented at network Layer 1, Layer 2, and / or Layer 3.

[0052] The CSR can be generated based on large amount of data available in Physical Layer and can be utilized by algorithms across different network layers. The CSR is thus generated at Physical Layer and made available for other network layers. In this way, CSR can be used at different entities and network layers in the network. Further, this enables the network to be tested on the fly, using testing procedures that require, or at least can handle, large amounts of data. In turn, this also enables in-depth analysis of the network and thus provides network observability. Such in-depth analysis of the network could enable self-healing automation, if needed. Examples of how CSR can be generated and used will be disclosed below. P110510W001 9

[0053] Generally, the CSR can be of different types. For example, CSR can be generated based on wireless channel measurements from a subset or all available antenna elements, referred to as antenna domain CSR (ACSR). For example, CSR can be generated based on a channel representation created for part of the bandwidth the UE can be allocated (for example 20 MHz out of 200 MHz), referred to as frequency domain CSR (FCSR). For example, CSR can be generated based on a channel representation created for part of the time domain, referred to as time domain CSR (TCSR). For example, the CSR could be a combination of any of the above types of CSR. An example of a process 300 for obtaining combined CSR and transform the combined CSR into embedding vectors directly or into embedding vectors indirectly (thought additional processing) is illustrated in Fig. 3. As can be seen, wireless channel measurements 320 are extracted from four antenna ports 310, combined 340 with UE data 330 (in terms of CQI, RI, and buffer size) to combined data 350 and transformed into embedding vectors 360.

[0054] Further, the CSR can be generated based on wireless channel measurements in the actual network in combination with data from simulations (e.g., using Digital Twins technologies). In this way, the CSR can be used to provide more information to simulation environments and thereby improve the modeling of the network. Additionally, more than one CSR can be generated per UE, depending on needs, e.g., network deployment, scenario, traffic, service, etc. Still further, the CSR could be valid for different periods of time. For example, the CSR could be updated as soon as new information is made available to the UE. One advantage of using CSR is that it can be represented using less data compared to the amount of data available at Physical Layer. But due to its construction, it still preserves information in an efficient way, e.g., through encoding. By sending CSR over interfaces to other entities in the network, either within one and the same device or between different devices, the amount of data to be sent can be reduced, as well as reducing the delay within which this information is received. Further, algorithms implemented at different entities in the network can benefit from more detailed information being available for actions to be taken and / or decision to be taken. In this way, using CSR, it is possible for algorithms implemented at Layer 2 and / or Layer 3 to discover complex dependencies more efficiently between UEs in different scenarios, thus enabling P110510W001 10 better scheduling decisions in single-user (SU) and multi-user (MU) multiple-input multiple-output (MIMO) scenarios.

[0055] In general terms, the herein disclosed embodiments are based on an encoder and a decoder with a transformer architecture having an attention component. One example of a transformer architecture having an attention component is disclosed in the document “Attention is All You Need” by Ashish Vaswani et al, last revised 2 August 2023, as available at https: / / arxiv.0rg / pdf / 1706.03762. This type of architecture accepts a variable number of inputs and can generate a variable number of outputs. The attention component enables complex patterns to be captured. Furthermore, this type of architecture is highly scalable and can be adapted based on hardware needs as well as the amount of data to be processed. As will be demonstrated below, the herein disclosed embodiments therefore make it possible to capture complex dependencies and relations between various elements, building up information about the transmission channel and the UE’s perception of the transmission channel. Further, by means of the aforementioned CSR, it is possible for reported CSI feedback, and other types of reported information, to be included.

[0056] Transformer architectures generally accept sequence of elements as input and produce (or predict) another sequence as output. In this respect, the elements generally refer to the input of the transformer. These elements can be of any unit. This unit is most likely a set of numbers arranged in systematic way, such as a vector, a matrix, or even single number. A token is the smallest unit which is transformed into an embedding vector. This transformation may be one-to-one mapping which basically does not introduce any change except that an element becomes a token. However, the transformation could also be more sophisticated. For example, each element could be divided further into smaller parts, which then becomes tokens. This process is called tokenization. Each token is transformed into an embedding vector. This transformation may be, but does not have to be, a one-to-one mapping. An encoder is configured to capture dependencies between elements of the input sequence. These dependencies can be described by embedding vectors and their relations in embedding space in which these vectors were initially created (out of the input sequence). Converting the elements of the input sequence into vectors in embedding space determines how relations between the elements are captured. According to the herein disclosed embodiments, and as will be further disclosed P110510W001 11 below, the input to the encoder (and thus to the transformer architecture in the encoder) are wireless channel measurements of a UE, and the output from the encoder (and thus from the transformer architecture) is a refined representation of the wireless channel measurements. Furthermore, according to the herein disclosed embodiments, and as will be further disclosed below, the input to the decoder (and thus to the transformer architecture in the decoder) is the refined representation of the wireless channel measurements of the UE as well as further refined representation of the wireless channel measurements of the UE as received via feedback from the output from the decoder. The output from the decoder, and thus from the transformer architecture, is thus a further refined representation of these wireless channel measurements. Generally, the elements of the input sequence can be either of single value type or of a more complex structure, such as a combination of more than one value, categorical data, etc. The same applies to the output sequence. The length of the input and output can be of variable size and can differ from each other, independently.

[0057] A transformer that spans over more than one network layer of the architecture is referred to as a distributed transformer. Hence, a distributed transformer can span over Physical Layer and Layer 2 and / or Layer 3, where the Physical layer could be either in the TRP or the UE and where Layer 2 and Layer 3 are in the TRP. For example, in case the UE sends embedding vectors to the TRP, the embedding vectors can be used at any network layer in the TRP.

[0058] In case the decoder is located at Layer 2, before the final output sequence is generated, the decoder can accept as input addition UE or cell related information that helps to provide more accurate predictions, and thus improve the quality of the further refined representation of the wireless channel measurements of the UE.

[0059] Depending on the implementation of the decoder, the element, or elements, of the output sequence could be of different types. For example, the output sequence could be a tuple of UE indexes, SINR values and their corresponding transmission rank, e.g., for SU / MU MIMO transmission. For example, the output sequence could be a tuple of UE indexes, modulation and coding scheme (MCS) and transmission rank. For example, the output sequence could be a tuple of UE indexes of UEs grouped together in a MU MIMO scenario, as given for the whole bandwidth, per resource P110510W001 12 block, or for multiple resource blocks. For example, the output sequence could be a resource block level resource allocation description comprising indexes of UEs transmitting in a given resource block, as well as other parameters of the transmission for UEs in that particular resource blocks, for example, MCSs and transmission ranks. Further, at the decoder, the attention component is used to capture complex patterns between elements of the input sequence as well as the iteratively generated output sequence.

[0060] As an introductory and non-limiting example, assume that the wireless channel measurements are based on uplink reference signals sent from each antenna of a UE and that the TRP measures the received power at each element of its own antennas. Assume further that the uplink reference signal is sent such that it is sparse in time and frequency. Due to this sparse nature of the wireless channel measurement, a sparse channel matrix H (t) can be constructed at a given time t. At this given time t, the input to the transformer at the encoder could comprise P number of sparse channel matrices carrying the wireless channel measurements represented by H(t), H(t — 1), H (t — 2), ..., H (t — P — 1). Assuming further that some processing, such as eigenvector decomposition, averaging, etc., is applied to the wireless channel measurements such that a single matrix representing the channel estimation at time t is obtained. This a single matrix can be constructed based on wireless channel measurements captured over time, frequency and antenna dimensions. For example, the content of the matrix can be gradually built over time as new measurements become available.

[0061] In a cell there are typically a number of UEs that can simultaneously provide channel state information. Therefore, matrices representing the wireless channel measurements of all (or subset) UEs served by the TRP and having channel state information reporting available, such as periodic, semipersistent or aperiodic channel state information reporting, can be provided as input of the transformer architecture.

[0062] Further, stacks of multiple such matrices can define channel related data that can be provided as input to the transformer architecture in the encoder. In some examples, this input data is of dimensions N-by-M-by-L-by-Q. Here, N is number of UE antennas per UE, M is the number of antennas at the TRP, L is the number of UEs stacked together, and Q is the number of wireless channel measurements used in P110510W001 13 frequency domain for a single UE, e.g., the number of subcarriers used for providing channel state information. Scalability can take place along the Q-dimension. It is possible to rely on only subset of available subcarriers if there is a need to downsize the channel state information reporting due to limitations. These limitations might be related to processing hardware capabilities, delay constraints or other. Further scalability is also possible along antenna dimensions, in particular along the M- dimension. In Fig. 4 is illustrated a process 400 where wireless channel measurements 420 from different subcarriers 410 are concatenated 430, such that they represent the channel toward the same UE antenna. There will thus be set of N vectors 440, each of size M-by-Q. In the illustrated example there is one vector per each of the N=4 UE antennas of size 64x2 corresponding to the number of antennas, M, at the TRP and the number of channel measurements Q.

[0063] In Fig. 5 is illustrated an example of a process 500 where four embedding vectors 510 are prepared for a single UE equipped with four antennas. Each embedding vector 540 consists of wireless channel measurements q and q2for two subcarriers, CQI, transmission rank indicator (RI) reported by the UE, possible together with other metrics, such as MCS. Channel data from different subcarriers can be combined such that the channel data represents the channel toward the same UE antenna. Further, combining 530 channel data with part of the UE specific data 520 may be performed before the data is provided as input to the transformer architecture in the encoder. Each of the embedding vectors 540 in the example of Fig. 5 comprise the following type of data: channel data between the TRP antennas and one of UE antennas, channel data collected from a number of subcarriers, UE reported CQI, UE reported RI, other metrics, etc. There could be other types of data combined to form an embedding vector or to form data transformed into embedding vectors.

[0064] Due to the nature of the transformer architecture, the embedding vectors may be subject to further processing in order for the transformer architecture to capture and use complex patterns in different domains. These complex pattern may pertain to any, or any combination, of: dependencies between the quality of special channels for a single UE or for multiple UEs, time domain pattern that captures the behavior of the wireless channel over time (which can be used for channel estimation SRS resources allocation), frequency domain patterns (which can be used for frequency selective radio resources scheduling), UE hardware information (which translates P110510W001 14 into UE channel state information reporting values; how the UE estimates channel quality may affect the reported CQI and RI), network patterns (e.g., relating to dependencies between cells arising, for example, from SRS resources configuration or cell traffic loads).

[0065] One way to achieve a computation cost reduction is to replace raw wireless channel measurements with SINR values. This provides a reduction of model dimension. Such possibilities for model dimension reduction provides flexibility of the model to adjust to hardware requirements if necessary. Hence, at its simplest form, wireless channel measurements represented by SINR are provided as tokens.

[0066] Reference is next made to the block diagram in Fig. 6 a distributed transformer 600 implemented in an information encoder device 630 (denoted “Encoder”) and an information decoder device 640 (denoted “Decoder”). As illustrated in Fig. 6, the information encoder device is implemented at network Layer 1 (or Physical Layer) whereas the information decoder device is implemented at network Layer 2 or Layer 3. The information encoder device and the information decoder device are operatively connected over an interface. Either the constructions of embedding vectors are performed in the information encoder device, or embedding vectors as constructed by an embedding vector construction block 610 (denoted “Emb. Vect. Contsr.”) are provided to the information encoder device. In any case, the embedding vectors can be constructed based on wireless channel measurements and auxiliary information such as UE and cell level related data available at Physical Layer, as provided by block 620, at Layer 2 and / or Layer 3, as provided by block 650,. The output from the information decoder device can be provided to an application 660 implemented at network Layer 2 or Layer 3.

[0067] Reference is now made to Fig. 7 illustrating a method for encoding wireless channel measurements as performed by the information encoder device according to an embodiment. In some embodiments, the information encoder device is provided in either a network node or a UE.

[0068] S102: The information encoder device obtains wireless channel measurements of a UE as available at Physical Layer. P110510W001 15

[0069] Sio6: The information encoder device generates a refined representation of the wireless channel measurements using a transformer architecture. As previously disclosed, the transformer architecture has an attention component. According to the transformer architecture, each available basic element of the wireless channel measurements is transformed into a respective embedding vector.

[0070] Embodiments relating to further details of encoding wireless channel measurements as performed by the information encoder device will now be disclosed with continued reference to Fig. 7

[0071] As disclosed above, at its simplest form, wireless channel measurements represented by SINR are provided as tokens. Therefore, in some embodiments, the basic elements of the wireless channel measurements are any of: SINR values, CQI, RI, channel estimates, UE scheduling metric, UE reported noise estimation, UE reported interference level estimation, channel characteristic (e.g., indication if the channel is of type Line of Sight (LOS) or Non Line of Sight (NLOS), Doppler shift, UE speed, UE buffer size, UE hardware description (e.g., number of antennas, type of receiver, etc.).

[0072] In some aspects, and as disclosed above, the CSR includes wireless channel measurements and UE reported data, such as UE feedback in any form, for example CQI, ACK / NACK, etc, if not already included as a basic element. Hence, in some embodiments, each of the embedding vectors is based on the wireless channel measurements and auxiliary information as reported by the UE.

[0073] As disclosed above, channel data from different subcarriers can be combined such that the channel data represents the channel toward the same UE antenna. Hence, in some embodiments, each of the embedding vectors is based on a concatenation of wireless channel measurements for at least two subcarriers.

[0074] As further disclosed above, CSR can be generated based on wireless channel measurements from a subset or all available antenna elements. Hence, in some embodiments, the embedding vectors are created from wireless channel measurements of a subset of antenna elements of the UE and / or of a transmission and reception point communicating with the UE. P110510W001 16

[0075] As further disclosed above, CSR can be generated based on a channel representation created for part of the bandwidth the UE can be allocated. Hence, in some embodiments, the embedding vectors are created from wireless channel measurements of a subset of bandwidth allocated to the UE.

[0076] The information encoder device might further obtain information about preferences, or requirements, e.g., based on the intended use of the wireless channel measurements, e.g., prioritize energy efficiency, prioritize performance (in terms of throughput, latency, energy consumption, etc.) for a particular group of UEs, etc. In this way there can be more than one CSR per UE. Hence, in some embodiments, the information encoder device is configured to perform (optional) step S104.

[0077] S104: The information encoder device obtains information regarding which type of auxiliary information the embedding vectors are to be based on.

[0078] In some aspects, the attention mechanism is used to make some embedding vectors more important than others. That is, in some embodiments, each embedding vector has a respective priority as given by the attention component.

[0079] Further aspects of how the refined representation of the wireless channel measurements can be generated using the transformer architecture. In some aspects, the basic elements are transformed into respective embedding vectors. In particular, in some embodiments, the information encoder device is configured to perform (optional) step S106-2 as part of generating the refined representation.

[0080] S106-2: The information encoder device transforms each basic element of the wireless channel measurements into a respective embedding vector.

[0081] In some aspects, the embedding vectors are updated based on a similarity score.

[0082] In particular, in some embodiments, the information encoder device is configured to perform (optional) steps S106-4, S106-6, and S106-8 as part of generating the refined representation.

[0083] S106-4: The information encoder device divides each embedding vector into parts for which a parallel process of obtaining an attention matrix is performed for determining similarity scores. P110510W001 17

[0084] S1O6-6: The information encoder device obtains similarity scores for the embedding vectors by performing a parallel process on the embedding vectors as divided into parts.

[0085] S106-8: The information encoder device updates the embedding vectors based on the similarity scores.

[0086] Steps S106-4, S106-6, and S106-8 may be repeatedly performed. In this respect, the dividing of each embedding vector in step S106-4 may be repeatedly performed based on the embedding vectors as updated.

[0087] A Feed Forward Network may be used after having updated the embedding vectors. Hence, in some embodiments, the information encoder device is configured to perform (optional) step S106-10 as part of generating the refined representation.

[0088] S106-10: The information encoder device applies a Feed Forward Network based non-linear transform to the embedding vectors.

[0089] In some aspects, the refined representation of the wireless channel measurements are provided to an application at network Layer 3, Layer 2, or Physical Layer. Hence, in some embodiments, the information encoder device is configured to perform (optional) step S108.

[0090] S108: The information encoder device provides the refined representation of the wireless channel measurements to a data link layer application, a network layer application, or a physical layer application.

[0091] Reference is next made to Fig. 8 in which is illustrated a block diagram 800 of the information encoder device according to an embodiment. In an input embedding block 802, preparation, such as scaling of dimension, using for example interleaving and / or averaging, can be performed for the input data set, i.e., the wireless channel measurements. In a positional encoding block 804, UE order related information is added. A multi-head attention block 806 is configured to capture patterns and interdependencies between elements of the input sequence, where each basic element of the input sequence, so called tokens, is transformed into an embedding vector. The multi-head attention block allows the information encoder device to relate input embedding vectors to each other, enabling identification of complex patterns and P110510W001 18 dependencies among input elements (embedding vectors). This feature is of particular use in communication networks, since the context can be rapidly changing. For example, in a MU MIMO scenario, the context can be defined by different UEs which are about to be scheduled. Alternatively, the context can be defined by interference from other cells, which can be also a subject of frequent changes. Alternatively, the context can be defined by traffic requirements; e.g., that certain UEs should be a subject of prioritization. As illustrated, each embedding vector is divided into parts, called heads (denoted WQ, WX, and WV), for which parallel process for obtaining an attention matrix is performed. Calculations are performed using similarity score calculation as shown in Fig. 9. In this case the similarity score calculation is the dot product of two vectors. In more detail, in Fig. 9 is schematically illustrated a process 900 for updating one embedding vector EV1 through similarity scores calculation between EVi and all input embedding vectors EVi, EV2, EV3, EV4. With reference again to Fig. 8, processing can be implemented separately, and thus in parallel, for each head. Outputs from different heads are concatenated in a concatenation block 808 and further transformed via a linear layer in a linear block 810. After all calculations are done an attention matrix is obtained in which all the similarity scores used for updating all embedding vectors can be stored. The process of calculating the attention matrix is performed many times in the information encoder device for a given set of input embedding vectors and can be perceived as a refinement process. Residual connection and layer normalization is performed in a first Add & Norm block 812. This block is configured to handle vanishing gradient that may appear during the training phase and thereby enhances stability of the training and provides faster convergence. A Feed Forward (Network) block 814 is configured to implement a non-linear transformation that is applied to the embedding vectors and allows the transformer architecture to learn complex patterns. The Feed Forward (Network) block is applied to each input position separately and identically. Residual connection and layer normalization is performed in a second Add & Norm block 816. The output can then be provided to the information decoder device.

[0092] In the processing, through the (self-)attention mechanism, tokens update their previous representations of themselves. It is the context that affects the meaning, or significance, of a given embedding vector. In this way interdependencies and P110510W001 19 relationships between different tokes of sequence are learnt. Throughout the attention mechanism, a similarity score calculation takes place. The similarity score can be implemented as a simple dot product or in more advanced way. In the information encoder device, the similarity score can be calculated between each element of the input sequence. The similarity score can then be used to update previous representations of an embedding vector.

[0093] Embedding vectors processed through the information encoder device capture interdependencies and patterns present in the input sequence. The information encoder device can therefore be regarded as implementing a process of refinement which increase significance of certain embedding vector and decrease significance of others. Each embedding vector in this process can be updated several times, e.g., by means of iterations implemented in the information encoder device.

[0094] The embedding vectors provided at the input of the information decoder device are by the information decoder device further refined to produce a final prediction of Transformer model.

[0095] Reference is now made to Fig. 10 illustrating a method for decoding wireless channel measurements as performed by the information decoder device according to an embodiment. In some embodiments, the information decoder device is provided in a network node.

[0096] S204: The information decoder device obtains a refined representation of the wireless channel measurements for a UE from an information encoder device. The refined representation of the wireless channel measurements comprises embedding vectors.

[0097] S206: The information decoder device generates a further refined representation of the wireless channel measurements for the UE using a transformer architecture. The transformer architecture has an attention component. The embedding vectors and the refined representation are provided as input to the transformer architecture.

[0098] Hence, the information decoder device takes as input the embedding vectors from the information encoder device. The information decoder device also takes as input as previously generated by the information decoder device in some previous iteration. P110510W001 20

[0099] This is the further refined representation of wireless channel measurements that is provided as output from the information decoder device and fed back as input to the information decoder device. If there is no output generated yet (e.g., for the first iteration), optional initialization information can be provided as input to the information decoder device. For example, the input to the information decoder device could be UE SINR, which could be based on CQI and assuming SU MIMO type of transmission. Then the output from the information decoder device could be a predicted UE SINR for MU MIMO type of transmission.

[0100] Embodiments relating to further details of decoding wireless channel measurements as performed by the information decoder device will now be disclosed with continued reference to Fig. 10.

[0101] In some embodiments, the further refined representation of the wireless channel measurements is iteratively determined. Accordingly, a further refined representation of the wireless channel measurements as generated at one iteration stage is provided as input to the next iteration stage.

[0102] As disclosed above, the CSR may include wireless channel measurements and UE reported data, such as UE feedback in any form, for example CQI, ACK / NACK, etc. Hence, in some embodiments, each of the embedding vectors is based on the wireless channel measurements and auxiliary information as reported by the UE.

[0103] As furth disclosed above, the information encoder device might further obtain information about preferences, or requirements, e.g., based on the intended use of the wireless channel measurements. This information might be sent from the information decoder. Therefore, in some embodiments, the information decoder device is configured to perform (optional) step S202.

[0104] S202: The information decoder device provides information to the information encoder device regarding the type of auxiliary information the embedding vectors are to be based on.

[0105] The information decoder device may provide its output to some application implemented at network Layer 2 or Layer 3, or even Physical Layer. Hence, in some P110510W001 21 embodiments, the information decoder device is configured to perform (optional) step S208.

[0106] S208: The information decoder device provides the further refined representation of the wireless channel measurements to a data link layer application, a network layer application, or a physical layer application.

[0107] In this respect, the further refined representation of the wireless channel measurements may represent any of: an SINR value, a UE transmission RI value, an MCS value. In line with what was disclosed above, the further refined representation of the wireless channel measurements are generally provided as predictions, and hence in the present context the SINR value, the UE transmission RI value, and the MCS value are all predictions.

[0108] Further, a respective refined representation of the wireless channel measurements is obtained for a set of UEs, where the further refined representation of the wireless channel measurements may represent any of: a set of indexes of UEs grouped together per bandwidth segment and / or per resource block.

[0109] Reference is next made to Fig. 11 in which is illustrated a block diagram 1100 of the information decoder device according to an embodiment. In an output embedding block 1102, the information decoder device converts target tokens into embedding vectors. In a positional encoding block 1104, UE order related information is added. A masked multi-head attention block 1106 is configured to compute the self-attention of the target sequence. A masking mechanism is used so that each position can only attend to earlier position of the target sequence. The main approach is the same as in the multi-head attention block of the information encoder device with the difference coming from the fact the target sequence is used in the calculations. Residual connection and layer normalization is performed in a first Add & Norm block 1108. A multi-head attention block 1110 has a connection to processing taking place in the information encoder device and processing already performed in the information decoder device. This could be a connection between functionalities of different network layers, for example between functionalities of network Layer 1 (at which the information encoder device is implemented) and network Layer 2 or Layer 3 (at which the information decoder device is implemented). Residual connection and layer normalization is performed in a second Add & Norm block 1112. A Feed P110510W001 22

[0110] Forward (Network) block 1114 is, as in the information encoder device, configured to implement a non-linear transformation that is applied to the embedding vectors and allows the transformer architecture to learn complex patterns. Residual connection and layer normalization is performed in a third Add & Norm block 1116. The output can be further transformed via a linear layer in a linear block 1118. A softmax block 1120 then applies a softmax function to provide the final output 1122, in terms of probabilities, or predicted values.

[0111] An illustrative example of use of the herein disclosed embodiments in an open radio access network (ORAN) architecture will be disclosed next. Assume that a MU MIMO scenario is considered in Reciprocity-Assisted Transmission (RAT) or Reciprocity- Assisted Interference-Aware Transmission (RAIT) and that downlink transmission is considered. The transformer model is provided with input data regarding the number of UEs, and delivers, as output data, predictions about the UEs.

[0112] The input to the information encoder device contains data related to the number of UEs scheduled at a given TTI and data related to the SU SINR values for each UE.

[0113] The task of the transformer model is to provide prediction about SINR values for each UE once all UEs are considered to be part of MU type of transmission.

[0114] However, the efficiency of the transformer model could be improved if more types of input data are provided. This information can be provided in terms of channel estimation values acquired through measurements of uplink reference signals sent by the UE. Each UE can for this purpose be represented by a channel estimation matrix. For example, the channel estimation matrix may represent a single frequency carrier (<71), 4 antenna ports of UE, and 64 antenna elements at the network node. For UEx this information can be represented as Hqxx, for % = 1, . . . , N, where N is the number of UEs scheduled at the given TTI.

[0115] Further, for a large bandwidth, in which the UEs can be allocated downlink resources, data related to only one carrier might be not enough as input. A channel estimation matrix for another frequency carrier may therefore be added as input data for each of the UEs. Even more types of input data can be used. However, for the sake of simplicity in the description of the present example, only two frequency carriers are considered. For UEx the corresponding data at the input of the transformer P110510W001 23 model can then be represented jointly as Hqxx- Further, it could be beneficial to consider any UE reported data, such as CSI feedback, to further increase the efficiency of the transformer model.

[0116] The way the information is processed to create input data for the transformer model is part of the embedding vector construction functionality.

[0117] The input to the transformer model is provided on layer Li. The information encoder device is also located at layer Li. This enables the input to the transformer model to be compressed. The compression means that for each UE the context of other UEs, collected measurements and reported data, is considered. The compression further means that the amount of data representing the input to the transformer model can be reduced before it is sent to the information decoder device, which is located at layer L2 in the present example. In this way, applications at layer L2 can gain access to observability of data available at layer Li. The same approach can be applied when observability of data at layer Li is necessary at layer L3 layer, e.g. for applications that relate to energy efficiency in the network.

[0118] The output sequence is generated in an iterative manner in the information decoder device, one element of the output sequence at each iteration step. Before the first element of the output sequence is generated, a context to the information decoder device is provided that determines the way the elements of the output sequence are generated. In this respect, the context could comprise various information regarding the UEs, one or more cells, and / or other properties of the network. This context provides the information decoder device with additional information which is not available in the information received from the information encoder device.

[0119] Fig. 12 schematically illustrates, in terms of a number of structural units, the components of an information encoder device 1200 according to an embodiment. Processing circuitry 1210 is provided using any combination of one or more of a suitable central processing unit (CPU), multiprocessor, microcontroller, digital signal processor (DSP), etc., capable of executing software instructions stored in a computer program product 1410a (as in Fig. 14), e.g. in the form of a storage medium 1230. The processing circuitry 1210 may further be provided as at least one application specific integrated circuit (ASIC), or field programmable gate array (FPGA). P110510W001 24

[0120] Particularly, the processing circuitry 1210 is configured to cause the information encoder device 1200 to perform a set of operations, or steps, as disclosed above. For example, the storage medium 1230 may store the set of operations, and the processing circuitry 1210 may be configured to retrieve the set of operations from the storage medium 1230 to cause the information encoder device 1200 to perform the set of operations. The set of operations may be provided as a set of executable instructions. Thus the processing circuitry 1210 is thereby arranged to execute methods as herein disclosed.

[0121] The storage medium 1230 may also comprise persistent storage, which, for example, can be any single one or combination of magnetic memory, optical memory, solid state memory or even remotely mounted memory.

[0122] The information encoder device 1200 may further comprise a communications (comm.) interface 1220 for communications with other entities, functions, units, and devices. As such the communications interface 1220 may comprise one or more transmitters and receivers, comprising analogue and digital components.

[0123] The processing circuitry 1210 controls the general operation of the information encoder device 1200 e.g. by sending data and control signals to the communications interface 1220 and the storage medium 1230, by receiving data and reports from the communications interface 1220, and by retrieving data and instructions from the storage medium 1230. Other components, as well as the related functionality, of the information encoder device 1200 are omitted in order not to obscure the concepts presented herein.

[0124] Fig. 13 schematically illustrates, in terms of a number of structural units, the components of an information decoder device 1300 according to an embodiment. Processing circuitry 1310 is provided using any combination of one or more of a suitable central processing unit (CPU), multiprocessor, microcontroller, digital signal processor (DSP), etc., capable of executing software instructions stored in a computer program product 1410b (as in Fig. 14), e.g. in the form of a storage medium 1330. The processing circuitry 1310 may further be provided as at least one application specific integrated circuit (ASIC), or field programmable gate array (FPGA). P110510W001 25

[0125] Particularly, the processing circuitry 1310 is configured to cause the information decoder device 1300 to perform a set of operations, or steps, as disclosed above. For example, the storage medium 1330 may store the set of operations, and the processing circuitry 1310 may be configured to retrieve the set of operations from the storage medium 1330 to cause the information decoder device 1300 to perform the set of operations. The set of operations may be provided as a set of executable instructions. Thus the processing circuitry 1310 is thereby arranged to execute methods as herein disclosed.

[0126] The storage medium 1330 may also comprise persistent storage, which, for example, can be any single one or combination of magnetic memory, optical memory, solid state memory or even remotely mounted memory.

[0127] The information decoder device 1300 may further comprise a communications interface 1320 for communications with other entities, functions, units, and devices. As such the communications interface 1320 may comprise one or more transmitters and receivers, comprising analogue and digital components.

[0128] The processing circuitry 1310 controls the general operation of the information decoder device 1300 e.g. by sending data and control signals to the communications interface 1320 and the storage medium 1330, by receiving data and reports from the communications interface 1320, and by retrieving data and instructions from the storage medium 1330. Other components, as well as the related functionality, of the information decoder device 1300 are omitted in order not to obscure the concepts presented herein.

[0129] The information encoder device 1200 and / or the information decoder device 1300 may be provided as a standalone device or as a part of at least one further device. For example, the information encoder device 1200 and / or the information decoder device 1300 may be provided in a node of the radio access network or in a node of the core network. Alternatively, the information encoder device may wholly or partly reside in the UE. Alternatively, functionality of the information encoder device 1200 and / or the information decoder device 1300 may be distributed between at least two devices, or nodes. These at least two nodes, or devices, may either be part of the same network part (such as the radio access network or the core network) or may be spread between at least two such network parts. Thus, a first portion of the instructions performed by P110510W001 26 the information encoder device 1200 and / or the information decoder device 1300 may be executed in a first device, and a second portion of the instructions performed by the information encoder device 1200 and / or the information decoder device 1300 may be executed in a second device; the herein disclosed embodiments are not limited to any particular number of devices on which the instructions performed by the information encoder device 1200 and / or the information decoder device 1300 may be executed. Hence, the methods according to the herein disclosed embodiments are suitable to be performed by an information encoder device 1200 and / or an information decoder device 1300 residing in a cloud computational environment. Therefore, although a single processing circuitry 1210, 1310 is illustrated in Figs. 12 and 13 the processing circuitry 1210, 1310 may be distributed among a plurality of devices, or nodes. The same applies to the computer programs 1420a, 1420b of Fig.

[0130] 14-

[0131] Fig. 14 shows one example of a computer program product 1410a, 1410b comprising computer readable means 1430. On this computer readable means 1430, a computer program 1420a can be stored, which computer program 1420a can cause the processing circuitry 1210 and thereto operatively coupled entities and devices, such as the communications interface 1220 and the storage medium 1230, to execute methods according to embodiments described herein. The computer program 1420a and / or computer program product 1410a may thus provide means for performing any steps of the information encoder device 1200 as herein disclosed. On this computer readable means 1430, a computer program 1420b can be stored, which computer program 1420b can cause the processing circuitry 1310 and thereto operatively coupled entities and devices, such as the communications interface 1320 and the storage medium 1330, to execute methods according to embodiments described herein. The computer program 1420b and / or computer program product 1410b may thus provide means for performing any steps of the information decoder device 1300 as herein disclosed.

[0132] In the example of Fig. 14, the computer program product 1410a, 1410b is illustrated as an optical disc, such as a CD (compact disc) or a DVD (digital versatile disc) or a Blu-Ray disc. The computer program product 1410a, 1410b could also be embodied as a memory, such as a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), or an electrically erasable P110510W001 27 programmable read-only memory (EEPROM) and more particularly as a non-volatile storage medium of a device in an external memory such as a USB (Universal Serial Bus) memory or a Flash memory, such as a compact Flash memory. Thus, while the computer program 1420a, 1420b is here schematically shown as a track on the depicted optical disk, the computer program 1420a, 1420b can be stored in any way which is suitable for the computer program product 1410a, 1410b.

[0133] The inventive concept has mainly been described above with reference to a few embodiments. However, as is readily appreciated by a person skilled in the art, other embodiments than the ones disclosed above are equally possible within the scope of the inventive concept, as defined by the appended patent claims.

Claims

P110510W001 28CLAIMS1. A method for encoding wireless channel measurements, wherein the method is performed by an information encoder device, and wherein the method comprises: obtaining (S102) wireless channel measurements of a user equipment, UE, as available at Physical Layer; and generating (S106) a refined representation of the wireless channel measurements using a transformer architecture having an attention component and according to which each available basic element of the wireless channel measurements is transformed into a respective embedding vector.

2. The method according to claim 1, wherein each embedding vector has a respective priority as given by the attention component.

3. The method according to claim 1, wherein the method further comprises: providing (S108) the refined representation of the wireless channel measurements to a data link layer application, a network layer application, or a physical layer application.

4. The method according to claim 1, wherein generating the refined representation comprises: transforming (S106-2) each basic element of the wireless channel measurements into said respective embedding vector.

5. The method according to claim 1, wherein generating the refined representation comprises: dividing (S106-4) each embedding vector into parts for which a parallel process of obtaining an attention matrix is performed for determining similarity scores; obtaining (S106-6) similarity scores for the embedding vectors by performing a parallel process on the embedding vectors as divided into parts; and updating (S106-8) the embedding vectors based on the similarity scores.P110510W001 296. The method according to claim 5, wherein said dividing each embedding vector is repeatedly performed based on the embedding vectors as updated.

7. The method according to claim 1, wherein generating the refined representation comprises: applying (S106-10) a Feed Forward Network based non-linear transform to the embedding vectors.

8. The method according to claim 1, wherein the basic elements of the wireless channel measurements are signal to interference plus noise ratio, SINR, values.

9. The method according to claim 1, wherein each of the embedding vectors is based on a concatenation of wireless channel measurements for at least two subcarriers.

10. The method according to claim 1, wherein the embedding vectors are created from wireless channel measurements of a subset of antenna elements of the UE and / or of a transmission and reception point communicating with the UE.

11. The method according to claim 1, wherein the embedding vectors are created from wireless channel measurements of a subset of bandwidth allocated to the UE.

12. The method according to claim 1, wherein each of the embedding vectors is based on the wireless channel measurements and auxiliary information as reported by the UE.

13. The method according to claim 12, wherein the method further comprises: obtaining (S104) information regarding type of auxiliary information the embedding vectors are to be based on.

14. The method according to claim 1, wherein the information encoder device is provided in either a network node or the UE.

15. A method for decoding wireless channel measurements, wherein the method is performed by an information decoder device, and wherein the method comprises:P110510W001 30 obtaining (S204) a refined representation of the wireless channel measurements for a user equipment, UE, from an information encoder device, wherein the refined representation of the wireless channel measurements comprises embedding vectors; and generating (S206) a further refined representation of the wireless channel measurements for the UE using a transformer architecture having an attention component and according to which the embedding vectors and the refined representation are provided as input.

16. The method according to claim 15, wherein the further refined representation of the wireless channel measurements is iteratively determined, wherein a further refined representation of the wireless channel measurements as generated at one iteration stage is provided as input to a next iteration stage.

17. The method according to claim 15, wherein the method further comprises: providing (S208) the further refined representation of the wireless channel measurements to a data link layer application, a network layer application, or a physical layer application.

18. The method according to claim 15, wherein each of the embedding vectors is based on the wireless channel measurements and auxiliary information as reported by the UE.

19. The method according to claim 18, wherein the method further comprises: providing (S202) information to the information encoder device regarding type of auxiliary information the embedding vectors are to be based on.

20. The method according to claim 15, wherein the further refined representation of the wireless channel measurements represents any of: a signal to interference plus noise ratio, SINR, value, a UE transmission rank indicator value, a modulation and coding scheme value.

21. The method according to claim 15, wherein a respective refined representation of the wireless channel measurements is obtained for a set of UEs, and wherein the further refined representation of the wireless channel measurements represents anyP110510W001 31 of: a set of indexes of UEs grouped together per bandwidth segment and / or per resource block.

22. The method according to claim 15, wherein the information decoder device is provided in a network node.

23. An information encoder device for encoding wireless channel measurements, the information encoder device comprising processing circuitry (1210), the processing circuitry being configured to cause the information encoder device to: obtain wireless channel measurements of a user equipment, UE, as available at Physical Layer; and generate a refined representation of the wireless channel measurements using a transformer architecture having an attention component and according to which each available basic element of the wireless channel measurements is transformed into a respective embedding vector.

24. The information encoder device according to claim 23, further being configured to perform the method according to any of claims 2 to 14.

25. An information decoder device for decoding wireless channel measurements, the information decoder device comprising processing circuitry (1310), the processing circuitry being configured to cause the information decoder device to: obtain a refined representation of the wireless channel measurements for a user equipment, UE, from an information encoder device, wherein the refined representation of the wireless channel measurements comprises embedding vectors; and generate a further refined representation of the wireless channel measurements for the UE using a transformer architecture having an attention component and according to which the embedding vectors and the refined representation are provided as input.

26. The information decoder device according to claim 25, further being configured to perform the method according to any of claims 16 to 22.P110510W001 3227. A computer program (1420a) for encoding wireless channel measurements, the computer program comprising computer code which, when run on processing circuitry (1210) of an information encoder device, causes the information encoder device to: obtain (S102) wireless channel measurements of a user equipment, UE, as available at Physical Layer; and generate (S106) a refined representation of the wireless channel measurements using a transformer architecture having an attention component and according to which each available basic element of the wireless channel measurements is transformed into a respective embedding vector.

28. A computer program (1420b) for decoding wireless channel measurements, the computer program comprising computer code which, when run on processing circuitry (1310) of an information decoder device, causes the information decoder device to: obtain (S204) a refined representation of the wireless channel measurements for a user equipment, UE, from an information encoder device, wherein the refined representation of the wireless channel measurements comprises embedding vectors; and generate (S206) a further refined representation of the wireless channel measurements for the UE using a transformer architecture having an attention component and according to which the embedding vectors and the refined representation are provided as input.

29. A computer program product (1410a, 1410b) comprising a computer program (1420a, 1420b) according to at least one of claims 27 and 28, and a computer readable storage medium (1430) on which the computer program is stored.

Images