Radio frequency positioning of a user equipment

Abstract

There is provided techniques for radio signal based positioning of a UE. The method is performed by a radio transceiver device. A method comprises obtaining a multitude of radio channel measurements for the UE. The method comprises providing the radio channel measurements as input to a neural network architecture. The neural network architecture comprises a transformer encoder layer and a global average pooling layer for combining outputs from the transformer encoder layer. The method comprises obtaining an output from the neural network architecture. The output is a final decision variable of the neural network architecture that indicates a position in space of the UE.

Description

[0001] RADIO FREQUENCY POSITIONING OF A USER EQUIPMENT

[0002] TECHNICAL FIELD

[0003] Embodiments presented herein relate to a method, a radio transceiver device, a computer program, and a computer program product for radio frequency positioning of a user equipment.

[0004] BACKGROUND

[0005] Many conventional radio signal based positioning techniques rely on the existence of enough line-of-sight (LoS) paths in the radio environment in order for trilateration algorithms to work correctly. An example is provided in Fig. 1. Fig. 1 is a schematic diagram illustrating a multipath radio environment 100 where embodiments presented herein can be applied. A network node 120 configured to control transmission and reception points (TRPs) i30a:i30N. Communication between the network node no and one of the TRPs is illustrated by a dotted line, although the network node no is enabled to communicate with all TRPs and all TRPs enabled to communicate with the network node 120. In turn, the TRPs 130a: 130N are configured to wirelessly communicate with user equipment (UEs) served by the network node 120. LoS paths between the TRPs 130a: 130N and UE 120 are illustrated by solid lines, and non-LoS paths (e.g., caused by reflections in one or more physical object 140a: 140c located in the environment) between the TRPs 130a: 130N and UE 120 are illustrated by dash-dotted lines. More specifically, using the illustration in Fig. 1 as example, for TRP 130a there is one LoS path and one non- LoS path for radio signals to travel between TRP 130a and the UE 120, for TRP 130b there is one LoS path for radio signals to travel between TRP 130b and the UE 120, for TRP 130c there is one LoS path for radio signals to travel between TRP 130c and the UE 120, and for TRP I3od there is one non-LoS path but no LoS path for radio signals to travel between TRP i3od and the UE 120. Hence, for TRP i3od, all radio signals will need to be reflected off at least one surfaces to reach the UE 120.

[0006] Generally, in a cluttered radio environment, there is often a low probability of LoS paths being available for a radio link between a UE and a TRP. For example, for InF- DH (Indoor Factory with Dense clutter and High base station height (Tx or Rx elevated above the clutter)) environment, as for example specified in the third- generation partnership project technical report 3GPP TR 38.901 “Study on channel model for frequencies from 0.5 to 100 GHz”, version 18.0.0, the LoS probability ranges from 44.9% in a mildly cluttered environment to only 0.8% in a heavily cluttered environment. Conventional positioning techniques struggle to locate a target UE in such heavily cluttered radio environments. Evaluations show that the 90%-tile positioning accuracy of conventional positioning methods is more than 15 meters in an InF-DH environment with clutter parameter {60%, 6m, 2m}, due to the unavailability of sufficient LoS links.

[0007] This, for example, motivates the application of artificial intelligence (Al) or machine learning (ML) based positioning techniques in such challenging radio environments. AI / ML deep learning models have been introduced to improve positioning accuracy by utilizing richer radio channel conditions than conventional radio signal based positioning techniques. However, many AI / ML deep learning models suffer from having a high computational complexity, which can be an issue in radio equipment with limited processing capabilities and / or power availability. Efforts have therefore been made to reduce the computational complexity but this often comes with performance of the AI / ML deep learning model being sacrificed, in turn leading to less accuracy of the position of the UE.

[0008] Hence, there is still a need for improved radio signal based positioning of UEs based on AI / ML deep learning models.

[0009] SUMMARY

[0010] An object of embodiments herein is to address the above issues with respect to using AI / ML deep learning model based positioning techniques.

[0011] A particular object is therefore to reduce the computational complexity of AI / ML deep learning models without any performance degradations with respect to the accuracy of the position of the UE.

[0012] According to a first aspect there is presented a method for radio signal based positioning of a UE. The method is performed by a radio transceiver device. The method comprises obtaining a multitude of radio channel measurements for the UE. The method comprises providing the radio channel measurements as input to a neural network architecture. The neural network architecture comprises a transformer encoder layer and a global average pooling layer for combining outputs from the transformer encoder layer. The method comprises obtaining an output from the neural network architecture. The output is a final decision variable of the neural network architecture that indicates a position in space of the UE.

[0013] According to a second aspect there is presented a radio transceiver device for radio signal based positioning of a UE. The radio transceiver device comprises processing circuitry. The processing circuitry is configured to cause the radio transceiver device to obtain a multitude of radio channel measurements for the UE. The processing circuitry is configured to cause the radio transceiver device to provide the radio channel measurements as input to a neural network architecture. The neural network architecture comprises a transformer encoder layer and a global average pooling layer for combining outputs from the transformer encoder layer. The processing circuitry is configured to cause the radio transceiver device to obtain an output from the neural network architecture. The output is a final decision variable of the neural network architecture that indicates a position in space of the UE.

[0014] According to a third aspect there is presented a computer program for radio signal based positioning of a UE. The computer program comprises computer code which, when run on processing circuitry of a radio transceiver device, causes the radio transceiver device to perform actions. One action comprises the radio transceiver device to obtain a multitude of radio channel measurements for the UE. One action comprises the radio transceiver device to provide the radio channel measurements as input to a neural network architecture. The neural network architecture comprises a transformer encoder layer and a global average pooling layer for combining outputs from the transformer encoder layer. One action comprises the radio transceiver device to obtain an output from the neural network architecture. The output is a final decision variable of the neural network architecture that indicates a position in space of the UE.

[0015] According to a fourth aspect there is presented a computer program product comprising a computer program according to the third 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. Advantageously, these aspects enable AI / ML deep learning models to be used for accurate positioning of a UE.

[0016] Advantageously, these aspects provide low computational complexity AI / ML model based radio signal based positioning of a UE that does not suffer from any performance degradations with respect to the accuracy of the position of the UE.

[0017] 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.

[0018] 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.

[0019] BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The inventive concept is now described, by way of example, with reference to the accompanying drawings, in which:

[0021] Fig. 1 is a schematic diagram illustrating a multipath radio environment according to embodiments;

[0022] Fig. 2 shows examples of CIR samples according to an embodiment;

[0023] Fig. 3 shows examples of sub-sampled CIR after down-sampling according to an embodiment;

[0024] Fig. 4(a) shows examples of truncated PDP according to an embodiment;

[0025] Fig. 4(b) shows examples of sub-sampled PDP after down-sampling according to an embodiment;

[0026] Fig. 5 schematically illustrates a neural network architecture according to an embodiment; Fig. 6 is a flowchart of methods according to embodiments;

[0027] Fig. 7 is a schematic diagram showing structural units of a radio transceiver device according to an embodiment;

[0028] Fig. 8(a) schematically illustrates a radio transceiver device being part of a network node according to an embodiment;

[0029] Fig. 8(a) schematically illustrates a radio transceiver device being part of a UE according to an embodiment;

[0030] Fig. 9 shows one example of a computer program product comprising computer readable storage medium according to an embodiment.

[0031] DETAILED DESCRIPTION

[0032] 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.

[0033] In general terms, positioning techniques for UEs (i.e., techniques for determining the location of a UE) are often based on reference signals being exchanged between the TRPs and the UE. For example, either downlink reference signals are transmitted from the TRPs and received by the UE, or uplink reference signals are transmitted by the UE and received by the TRPs. In this respect, assume that is the received reference signal symbol at subcarrier k of a receive antenna port a. The measured frequency domain channel response (FD CR) samples are obtained as where skis the complex conjugate of the known reference symbol at subcarrier k. Taking the inverse fast Fourier Transform (IFFT) of the frequency domain channel response samples gives the measured time domain channel impulse response (TD CIR) samples: ha[d] IFFT({tfa[fc]}k) where d = 0, 1, ... , / VFFT— 1 and NFFTis the size of the IFFT.

[0034] These time domain or frequency domain channel measurement samples are directly observable at the receiver. Further processing on these measurement samples can be applied as follows.

[0035] A truncated TD CIR can be obtained from the TD CIR by keeping only the first Ntsamples and discarding the last AFFT- Ntsamples.

[0036] A TD power delay profile (TD PDP) can be obtained from the (truncated) TD CIR by keeping only the power information across the antenna ports at each sampling grid points, while discarding the phase info of each sample:

[0037] A sub-sampled TD CIR, or PDP, can be obtained from a (truncated) TD CIR / PDP by keeping the values at the At' samples by selecting the At' samples which satisfy a certain criteria (for example, a typical criteria is to select the At' samples with the largest powers), and setting the other Nt- samples to zeros.

[0038] A time domain delay profile (TD DP) can be obtained from a sub-sampled TD PDP by setting the At' samples with the largest powers to a specific value. The specific value could e.g., be a constant such as i or the reference signal received power (RSRP) of the wireless link in question:

[0039] The CIR samples are complex-valued. An example is provided in Fig. 2 with two receive antenna ports; port o and port i. It can be observed that each CIR sample at an antenna port consists of a real part CIR.real and an imaginary part CIR.imag. In this figure, the CIR is truncated to the first Nt= 128 samples. In Fig. 3 is provided an example of the sub-sampled CIR after down-sampling to the = 9 strongest samples. It can be observed that nonzero values are present in only 9 of the sampling points with the rest set to zero.

[0040] The PDP samples are real-valued since the samples are represented by the received power at the sample points. An examples of the truncated PDP computed from the example two-port CIR in Fig. 2 is illustrated in Fig. 4(a), and the sub-sampled PDP after down-sampling to the = 9 strongest samples is illustrated in Fig. 4(b). In both Figs. 4(a) and 4(b) the square root of the PDP is plotted for easier inspection. In Fig. 4(b) it can be observed that nonzero values are present in only 9 of the sampling points with the rest set to zero.

[0041] From the above follows that the components of different types of positioning related measurement reports can be decomposed as: CIR (with timing information of the nonzero samples, power information of the nonzero samples, and phase information of the nonzero samples), PDP (with timing information of the nonzero samples, and power information of the nonzero samples), and DP (with timing information of the nonzero samples).

[0042] Any, or any combination, of theses positioning related measurement reports can be used for direct AI / ML positioning or AI / ML assisted positioning. In this respect, for direct AI / ML positioning, the AI / ML model output is the UE location (as a two- dimensional (2D) or three-dimensional (3D) coordinate). Direct AI / ML positioning typically refers to radio fingerprinting, where radio channel observations are used as the input of the AI / ML model. On the other hand, for AI / ML assisted positioning, the AI / ML model output is a new measurement and / or enhancement of existing measurements. The AI / ML model output can be, for example, LoS / non-LoS identification, timing and / or angle measurements, likelihood or reliability of the measurements, etc.

[0043] As noted above there is still a need for improved radio signal based positioning of UEs based on AI / ML deep learning models.

[0044] In further detail, one aspect of the present disclosure is the realization that the vision transformer architecture as presented in Dosovitskiy, A., et al: “An image is worth 16x16 words: Transformers for image recognition at scale”, published as a conference paper at the 2021 International Conference on Learning Representations, and as available at http: / / arxiv.0rg / abs / 2010.11929 per 13 September 2024, can be applied for radio signal based positioning of a UE. In further detail, the inventors of the present disclosure have discovered that the received signal sequence for each antenna, or TRP, as discussed in the above, can be treated as a patch input to the vision transformer. However, due to stringent computational limitations for radio units, it is undesirable to discard most of the computed tensor output from the transformer encoder as done in the vision transformer, since this will have a negative impact on the accuracy of the positioning of the UE. That is, in the referenced document, only a vector of length nembis retained out of the nembx (np+ 1) tensor. However, a more computationally efficient approach is needed due to the limited computational resources available in some radio equipment.

[0045] The embodiments disclosed herein therefore relate to techniques for ML / AI based radio signal based positioning of a UE no, see Fig. 1. In order to obtain such techniques, there is provided a radio transceiver device, a method performed by the radio transceiver device, a computer program product comprising code, for example in the form of a computer program, that when run on a radio transceiver device, causes the radio transceiver device to perform the method. The radio transceiver device could either be part of the network node 120, the UE no, or be provided in some other device being in communication with the UE 110.

[0046] At least some of the embodiments are based on enhancing the aforementioned vision transformer model by Dosovitskiy, A., et al such that the use of all computed tensor outputs is optimized in a computationally efficient manner. In some aspects, this can be achieved by eliminating the class token and using global average pooling (GAP) in order to reduce the number of parameters and computations required, whilst still leveraging the entire output from the transformer encoder. Hence, one purpose of the disclosed embodiments is to utilize all computed tensor output in a computationally and parameter storage efficient approach. This approach ensures efficient use of resources and improves model training and performance. This approach also allows for further improvement when combined with optional normalization and further processing through a multi-layer perceptron. Reference is here made to Fig. 5 in which is illustrated a neural network architecture 500 according to an embodiment. Radio channel measurements are provided as input 510, and the output 590 is a final decision variable of the neural network architecture 500 that indicates a position in space of the UE. A tokenization layer 520 is configured to convert the radio channel measurements input into multiple vectors. As an example, for the multi-ant enna / multi-TRP scenario as in Fig. 1, the vector of received signals for each antenna / TRP can be taken as an input vector. The flattened vectors are processed by a linear projection layer 530 into vectors of length nemb. For each flattened vector, a respective positioning embedding, which is also a vector of length nemb, can be added to the flattened vectors to represent the positioning information of which TRP 130a: 130N is associated with which measurement. The npvectors, each of length nemb, form a nembx npembedded tensor input to the transformer encoder layer 550. Here, the special class token approach as proposed in the aforementioned document by Dosovitskiy, A., et al is not used. The nembx npembedded tensor is thus processed by the transformer encoder layer 550 which can be implemented by multiple transformer blocks. The output of the multiple transformer blocks is an nembx nptensor. A global average pooling layer 560 is configured to combine the nembx nptensor into a length nembvector. In this neural network architecture 500 all computed outputs from the transformer encoder layer 550 are utilized at a low computational costs. The output of the global average pooling layer 560 can optionally be processed by a normalization layer 570. Using a normalization layer 570 can improve training convergence speed and performance of trained models. Further, the output of the global average pooling layer 560 (or the combination of the global average pooling layer 570 and the optional normalization layer 570) can be further processed by a multi-layer perceptron or a swish gated linear unit 580, or other type of feed forward layer, to produce the final decision variables. Here, the input / output tensor sizes may comprise an additional batch dimension. A referred vector of length n should therefore be understood as a tensor of size B x n when implemented with batch size of B. Similarly, a referred K- dimensional tensor should be understood as a K + 1) -dimensional tensor when implemented in batch mode.

[0047] Further details of the operation of the neural network architecture 500 for radio signal based positioning of a UE no will be disclosed next. Fig. 6 is a flowchart illustrating embodiments of methods for radio signal based positioning of a UE no. The methods are performed by the radio transceiver device. The radio transceiver device thus implements the neural network architecture 500. The methods are advantageously provided as computer programs.

[0048] S102: The radio transceiver device obtains a multitude of radio channel measurements for the UE no.

[0049] S102: The radio transceiver device provides the radio channel measurements as input 510 to a neural network architecture 500. As in Fig. 5, the neural network architecture 500 comprises a transformer encoder layer 550 and a global average pooling layer 560 for combining outputs from the transformer encoder layer 550.

[0050] S102: The radio transceiver device obtains an output 590 from the neural network architecture 500. The output 590 is a final decision variable of the neural network architecture 500 that indicates a position in space of the UE no.

[0051] Embodiments relating to further details of radio signal based positioning of a UE no as performed by the radio transceiver device will now be disclosed with continued reference to the neural network architecture 500 in Fig. 5.

[0052] The radio transceiver device can be part of either the network node 120 or the UE 110. In case the radio transceiver device is part of the network node 120, the radio channel measurements can be obtained from measurements on uplink reference signals as transmitted from the UE no and received at the plurality of TRPs i30a:i30N. Likewise, in case the radio transceiver device is part of the UE no, the radio channel measurements can be obtained from measurements on downlink reference signals as transmitted by the plurality of TRPs 130a: 130N and received by the UE 110. There maybe different types of radio channel measurements. In some non-limiting examples, the radio channel measurements pertain to any, or any combination, of: CIR measurements, PDP measurements, and DP measurements.

[0053] As disclosed above, for each transformed vector, a different positional embedding, which is also a vector of length nemb, can be combined with the transformed vector to represent the positioning information. Hence, in some embodiments, the transformer encoder layer 550 is configured to take its inputs from outputs from the linear projection layer 530 and positional embeddings 540. Either additive positional embedding or multiplicative rotary positioning embedding, or other types of positioning embedding, can be used for the positional information. In some examples, the multitude of radio channel measurements are associated with TRPs i30a:i30N, and the positional embeddings 540 represent information of which of the TRPs i30a:i30N are associated with which of the radio channel measurements.

[0054] As disclosed above, both the input to, and the output from, the transformer encoder layer 550 can be a tensor of dimension nembx np. Hence, in some embodiments, the outputs from the linear projection layer 530 are represented by a first tensor, the outputs from the transformer encoder layer 550 are represented by a second tensor, and the first tensor and the second tensor have the same dimensions. Further, in some aspects, the global average pooling layer 560 is used to reduce the dimensions of the output tensor (e.g., combining the nembx nptensor into a length nembvector). That is, in some embodiments, one of the dimensions of the second tensor is reduced at the global average pooling layer 560. As a nonlimiting example, the global average pooling layer 560 can be configured to reduce the nembx nptensor into a length nembvector y[e]: 0,1, ... , nemb- 1

[0055] This is one example where all computed outputs from the transformer encoder layer 550 can be utilized with low computational costs.

[0056] As disclosed above, in some aspects, the output of the global average pooling layer 560 is optionally processed by a normalization layer 570. Hence, in some embodiments, the output of the global average pooling layer 560 is provided as input to the normalization layer 570. Using the normalization layer 570 before the multilayer perceptron or swish-gated linear unit 580 can improve training convergence speed and performance of trained models.

[0057] As disclosed above, the output of the global average pooling layer 560 (or the combination of global average pooling layer 560 and the normalization layer 570) can be further processed by a multi-layer perceptron or Swish-Gated Linear Unit 580 to produce final decision variables. Hence, in some embodiments, a multi-layer perceptron or swish-gated linear unit 580 is configured to compute the final decision variable, and the global average pooling layer 560 is provided between the transformer encoder layer 550 and the multi-layer perceptron or swish-gated linear unit 580.

[0058] The neural network architecture 500 can be used for both direct AI / ML positioning and AI / ML assisted positioning. Here, the final decision variable maybe different depending on whether the network architecture 500 is used for direct AI / ML positioning or AI / ML assisted positioning. For example, in case the network architecture 500 is used for AI / ML assisted positioning, the final decision variable that indicates the position in space of the UE no can be any, or any combination, of: time-of-arrival (ToA) estimates with respect to the plurality of TRPs 130a: 13 oN, time- difference-of-arrival (TDoA) estimates with respect to the plurality of TRPs i30a:i30N (for example, downlink reference signal time difference, and / or uplink relative time-of-arrival), timing advance (TA) estimates with respect to the plurality of TRPs 130a: 130N, angle-of-arrival (AoA) estimates with respect to the plurality of TRPs i30a:i30N, angle-of-departure (AoD) estimates with respect to the plurality of TRPs i30a:i30N, reference signal received power (RSRP), for example, downlink positioning reference signal (PRS) RSRP or uplink sounding reference signal (SRS) RSRP, reference signal received path power (RSRPP), for example, downlink PRS RSRPP or uplink SRS RSRPP, cell identifier (ID) and TRP related information (e.g., reference signal resource and / or resource set ID), carrier phase difference, round-trip time (RTT) measurements as obtained by combining TRP receive - transmit time difference and UE receive - transmit time difference. For example, in case the network architecture 500 is used for direct AI / ML positioning, the final decision variable that indicates the position in space of the UE no can be a two-dimensional coordinate (x,y) or a three-dimensional coordinate (x,y,z) of the UE no.

[0059] As an example, radio signal based positioning of a UE 110 as performed according to the herein disclosed embodiments was applied to the aforementioned 3GPP indoor factory (InF) model. In this scenario, 18 TRPs are deployed in a factory with TRP locations known by the network. With 60% clutter density and clutter height and width of 6 m and 2 m, respectively, this indoor factory scenario has less than 1% LoS probability from a UE to any TRP. It is assumed that the UE is requested to transmit an uplink reference signal, e.g., an SRS over a New Radio (NR) air-interface. This uplink reference signal is received by multiple TRPs at known positions to produce power delay profile (PDP) inputs, or other types of radio channel measurements, for the neural network architecture.

[0060] To demonstrate the performance of the disclosed exemplary embodiments, two ML models were trained for UE positioning. A first ML model follows the original vision transformer architecture as disclosed in the aforementioned document by Dosovitskiy, A., et al and uses a special class token to produce decision variables for UE positions. A second ML model is implemented by the neural network architecture 500. The architecture parameters / settings for the two ML models are tuned such that both have roughly the same 4.5 million floating point operations (M FLOPs). The 90- percentile position error is considered as the main performance criterion. It can be observed that the conventional vision transformer model has a 90-percentile position error of 0.620 m with respect to the position of the UE. However, the enhanced efficient transformer approach as disclosed herein has a lower 90-percentile position error of 0.597 m. The proposed neural network architecture 500 thus reduces the 90- percentile position error by almost 4%.

[0061] Fig. 7 schematically illustrates, in terms of a number of structural units, the components of a radio transceiver device 700 according to an embodiment. Processing circuitry 710 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 910 (as in Fig. 9), e.g. in the form of a storage medium 730. The processing circuitry 710 may further be provided as at least one application specific integrated circuit (ASIC), or field programmable gate array (FPGA).

[0062] Particularly, the processing circuitry 710 is configured to cause the radio transceiver device 700 to perform a set of operations, or steps, as disclosed above. For example, the storage medium 730 may store the set of operations, and the processing circuitry 710 may be configured to retrieve the set of operations from the storage medium 730 to cause the radio transceiver device 700 to perform the set of operations. The set of operations may be provided as a set of executable instructions. Thus the processing circuitry 710 is thereby arranged to execute methods as herein disclosed. The storage medium 730 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. The radio transceiver device 700 may further comprise a communications (comm.) interface 720 at least configured for communications with other entities, functions, nodes, and devices, such as network nodes, UEs, and transmission and reception points. As such the communications interface 720 may comprise one or more transmitters and receivers, comprising analogue and digital components. The processing circuitry 710 controls the general operation of the radio transceiver device 700 e.g. by sending data and control signals to the communications interface 720 and the storage medium 730, by receiving data and reports from the communications interface 720, and by retrieving data and instructions from the storage medium 730. Other components, as well as the related functionality, of the radio transceiver device 700 are omitted in order not to obscure the concepts presented herein.

[0063] The radio transceiver device 700 maybe provided as a standalone device or as a part of at least one further device. For example, the radio transceiver device 700, 830 may be provided in a node of the radio access network, in a node of the core network, or in a UE. Reference is here made to Fig. 8(a) and Fig. 8(b). In Fig. 8(a) is illustrated an example where the radio transceiver device 830 is part of a network node 810. Further, in Fig. 8(b) is illustrated an example where the radio transceiver device 830 is part of a UE 820. Alternatively, functionality of the radio transceiver device 700, 830 maybe 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 maybe spread between at least two such network parts. In general terms, instructions that are required to be performed in real time may be performed in a device, or node, operatively closer to the cell than instructions that are not required to be performed in real time. Thus, a first portion of the instructions performed by the radio transceiver device 700, 830 may be executed in a first device, and a second portion of the of the instructions performed by the radio transceiver device 700, 830 maybe 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 radio transceiver device 700, 830 may be executed. Hence, the methods according to the herein disclosed embodiments are suitable to be performed by a radio transceiver device 700, 830 residing in a cloud computational environment. Therefore, although a single processing circuitry 710 is illustrated in Fig. 7 the processing circuitry 710 maybe distributed among a plurality of devices, or nodes. The same applies to the computer program 920 of Fig. 9.

[0064] Fig. 9 shows one example of a computer program product 910 comprising computer readable storage medium 930. On this computer readable storage medium 930, a computer program 920 can be stored, which computer program 920 can cause the processing circuitry 710 and thereto operatively coupled entities and devices, such as the communications interface 720 and the storage medium 730, to execute methods according to embodiments described herein. The computer program 920 and / or computer program product 910 may thus provide means for performing any steps as herein disclosed.

[0065] In the example of Fig. 9, the computer program product 910 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 910 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 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 920 is here schematically shown as a track on the depicted optical disk, the computer program 920 can be stored in any way which is suitable for the computer program product 910.

[0066] 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

CLAIMS1. A method for radio signal based positioning of a user equipment, UE (no), wherein the method is performed by a radio transceiver device (700, 830), and wherein the method comprises: obtaining (S102) a multitude of radio channel measurements for the UE (110); providing (S104) the radio channel measurements as input (510) to a neural network architecture (500), wherein the neural network architecture (500) comprises a transformer encoder layer (550) and a global average pooling layer (560) for combining outputs from the transformer encoder layer (550); and obtaining (S106) an output (590) from the neural network architecture (500), wherein the output (590) is a final decision variable of the neural network architecture (500) that indicates a position in space of the UE (no).

2. The method according to claim 1, wherein the radio channel measurements pertain to any, or any combination, of: channel impulse response measurements, power delay profile measurements, and delay profile measurements.

3. The method according to claim 1 or 2, wherein the neural network architecture (500) further comprises a linear projection layer (530), wherein the transformer encoder layer (550) is configured to take its inputs from outputs from the linear projection layer (530) and positional embeddings (540).

4. The method according to claim 3, wherein the multitude of radio channel measurements are associated with transmission and reception points (130a: 130N), and wherein the positional embeddings (540) represent information of which of the transmission and reception points (130a: 130N) are associated with which of the radio channel measurements.

5. The method according to claim 3 or 4, wherein the outputs from the linear projection layer (530) are represented by a first tensor, wherein the outputs from the transformer encoder layer (550) are represented by a second tensor, and wherein the first tensor and the second tensor have same dimensions.

6. The method according to claim 4 or 5, wherein one of the dimensions of the second tensor is reduced at the global average pooling layer (560).

7. The method according to any preceding claim, wherein the neural network architecture (500) further comprises a normalization layer (570), and wherein an output of the global average pooling layer (560) is provided as input to the normalization layer (570).

8. The method according to any preceding claim, wherein the neural network architecture (500) further comprises a multi-layer perceptron or swish-gated linear unit (580) configured to compute the final decision variable, and wherein the global average pooling layer (560) is provided between the transformer encoder layer (550) and the multi-layer perceptron or swish-gated linear unit (580).

9. The method according to any preceding claim, wherein the multitude of radio channel measurements are obtained with respect to a plurality of transmission and reception points (130a: 130N), and wherein the final decision variable that indicates the position in space of the UE (no) is any, or any combination, of: time-of-arrival estimates with respect to the plurality of transmission and reception points (130a: 130N), time-difference-of-arrival estimates with respect to the plurality of transmission and reception points (130a: 130N), timing advance estimates with respect to the plurality of transmission and reception points (130a: 130N), angle-of- arrival estimates with respect to the plurality of transmission and reception points (130a: 130N), angle-of-departure estimates with respect to the plurality of transmission and reception points (130a: 130N).

10. The method according to any of claims 1 to 8, wherein the final decision variable that indicates the position in space of the UE (no) is a two-dimensional or three- dimensional coordinate (x,y,z) of the UE (no).

11. The method according to any preceding claim, wherein the radio transceiver device (700, 830) is part of a network node (120, 810).

12. The method according to claim 11, wherein the radio channel measurements are obtained from measurements on uplink reference signals as transmitted from the UE (no) and received at a plurality of transmission and reception points (130a: 130N).13- The method according to any of claims 1 to io, wherein the radio transceiver device (700, 830) is part of the UE (no, 820).

14. The method according to claim 13, wherein the radio channel measurements are obtained from measurements on downlink reference signals as transmitted by a plurality of transmission and reception points (130a: 130N) and received by the UE (no).

15. A radio transceiver device (700, 830) for radio signal based positioning of a user equipment, UE (no), the radio transceiver device (700, 830) comprising processing circuitry (710), the processing circuitry being configured to cause the radio transceiver device (700, 830) to: obtain a multitude of radio channel measurements for the UE (110); provide the radio channel measurements as input (510) to a neural network architecture (500), wherein the neural network architecture (500) comprises a transformer encoder layer (550) and a global average pooling layer (560) for combining outputs from the transformer encoder layer (550); and obtain an output (590) from the neural network architecture (500), wherein the output (590) is a final decision variable of the neural network architecture (500) that indicates a position in space of the UE (no).

16. The radio transceiver device (700, 830) according to claim 15, further being configured to perform the method according to any of claims 2 to 14.

17. A computer program (920) for radio signal based positioning of a user equipment, UE (no), the computer program comprising computer code which, when run on processing circuitry (710) of a radio transceiver device (700, 830), causes the radio transceiver device (700, 830) to: obtain (S102) a multitude of radio channel measurements for the UE (110); provide (S104) the radio channel measurements as input (510) to a neural network architecture (500), wherein the neural network architecture (500) comprises a transformer encoder layer (550) and a global average pooling layer (560) for combining outputs from the transformer encoder layer (550); andobtain (S106) an output (590) from the neural network architecture (500), wherein the output (590) is a final decision variable of the neural network architecture (500) that indicates a position in space of the UE (110).

18. A computer program product (910) comprising a computer program (920) according to claim 17, and a computer readable storage medium (930) on which the computer program is stored.

Images