Method for transmitting information, method for receiving information and devices configured to implement these methods

Abstract

The invention relates to a method for transmitting information by a first device (3) to a second device (2) via a communication network, comprising: - a step of receiving reference signals transmitted by the second device; and - a step of transmitting to the second device information relating to predictions regarding the state of a propagation channel between the second device and the first device, made by the first device on the basis of the received reference signals using at least one prediction technique based on machine learning and another prediction technique.

Description

Description Title of the invention: Method for transmitting information, method for receiving information, and devices configured to implement these methods Technical Field

[0001] The present invention belongs to the general field of telecommunications. It relates more particularly to techniques for acquiring knowledge of the state of the propagation channel separating two devices communicating via a network.

[0002] The invention has a preferred but not limiting application in the context of telecommunications networks defined by the 3GPP standard and in particular 4G, 5G, 6G, etc. networks. However, it can be applied in other contexts whenever it becomes necessary to estimate the state of a propagation channel between two devices communicating via a network (for example, in the context of an IEEE802.11x network). Previous technique

[0003] Many telecommunications networks, such as those defined by the 3GPP standard, now rely on multi-antenna communication devices, whether they are user equipment (UEs) or network access points (e.g., base stations). The propagation or transmission channels separating these communication devices are then called multiple-input multiple-output (MIMO) channels.

[0004] In MIMO systems, knowledge of the propagation channel state, or CSI (Channel State Information), between a transmitter and a receiver is crucial because it allows the transmitter to adapt its transmission strategy and, in particular, to appropriately select a precoding scheme to apply to the data it wishes to transmit to the receiver in order to optimize communication with that receiver. Various techniques exist to enable a transmitter to acquire knowledge of the propagation channel state separating it from a receiver.

[0005] The 3GPP standard defines, particularly for downlink communications, a CSI acquisition technique based on the use of a limited feedback channel between the receiver and the transmitter, i.e., in the downlink between the UE and the base station. According to this technique, the base station transmits reference signals (for example, CSI-RS signals for "Channel State Information - Reference Signals" in a 5G network) which allow the UE to estimate the transmission channel between the base station and the UE, i.e., in the transmitter-to-receiver direction.Based on this estimate, the UE determines a precoder or equivalently a precoding matrix to be used by the base station and sends its choice back to the base station via the return channel, for example in the form of a PMI (Precoding Matrix Indicator) type indicator pointing to a matrix from a finite dictionary (or codebook) of precoding matrices specified by the 3GPP standard.

[0006] Two types of dictionaries, known as Type-I and Type-II, as well as variants (e.g., eType-II), are defined by the 3GPP standard for 5G networks, specifically in the 3GPP TS 38.214 document entitled "3GPP Technical Specification Group Radio Access Network; NR; Physical layer procedures for data (Release 18)", V18.6.0, March 2025. A Type-I dictionary is associated with low-resolution channel knowledge, while a Type-II dictionary corresponds to higher-resolution channel knowledge. Although Type-II dictionaries are generally more efficient than Type-I dictionaries, they are considered less resilient to the mobility of network users, particularly the speed at which their UEs (User Units) move.In a context of UE mobility, it is possible that the precoding matrix applied by the base station may no longer be suitable for the propagation channel between the base station and the UE due to channel variations.

[0007] To address the problem of outdated knowledge of the propagation channel at the base station, a new variant of the Type-II dictionary, called eType-II Doppler or eType-II for predicted PMIs, was introduced in Release 18 of the 3GPP standard (see TS 38.214 cited above). This variant allows the UE to make predictions of the propagation channel state at different times separated by a predetermined number of slots, based on reference signals sent by the base station. The number of predictions made is denoted N4, where N4 = {1, 2, 4, 8}. Based on the N4 predictions thus obtained, N4 precoding matrices are selected by the UE and then compressed in the time domain using a DFT (Discrete Fourier Transform) type temporal compression matrix, thus forming an eType-II Doppler precoder.This pre-encoder is transmitted back to the base station so that it has up-to-date knowledge of the channel status.

[0008] Predictions are made by the EU using a machine learning (AI / ML) prediction technique (or equivalently, a model), such as a transformer-based technique or a long short-term memory (LSTM) model. The introduction of an AI / ML prediction model for channel state aims to optimize the estimation of the transmission channel, and more specifically its accuracy, from limited reference signals, with the goal of reducing the resources required for this estimation. However, such a prediction model faces certain challenges in 5G networks.

[0009] Indeed, in a non-stationary environment, CSI predictions made using AI / ML models can fail due to rapidly changing propagation conditions related to the UE's environment (for example, due to the mobility of scatterers located near the UE). Furthermore, the effectiveness of such a model is strongly linked to its training: if it is not sufficiently trained, the CSI predictions it makes may not be reliable. Consequently, there is a risk of obtaining lower performance with an eType-II Doppler codebook than with a Type-I codebook, which is less precise but more resilient to UE mobility.

[0010] To support the deployment of AI / ML channel state prediction models and ensure their effectiveness in the context of 5G networks, ongoing studies at the 3GPP standard level are focusing on the implementation of new functionalities known as Life Cycle Management (LCM) of AI / ML prediction models in the network and at the UE level: data collection (for model training, inference, monitoring, activation, etc.), model learning and updating, model storage and management (monitoring and evaluation of its performance, activation / deactivation, etc.), exchange of supported AI / ML models, identification, etc.

[0011] The document Rl-2500057, entitled "AI / ML for CSI prediction," Ericsson, 3GPP TSG-RAN WG1 Meeting #120, February 17-21, 2025, focuses specifically on monitoring the performance of AI / ML prediction when it is activated at the UE level. To this end, it proposes that after activation of the eType-II Doppler CSI acquisition technique based on an AI / ML prediction, the UE calculates an NMSE (Normalized Mean Squared Error) or SGCS (Subband Gain Computation Score) performance metric of the AI / ML predictions it has made, and transmits this performance metric to the base station.Depending on the value of the metric reported to it, it can then decide to switch from an AI / ML prediction-based eType-II Doppler CSI acquisition technique to a "classic" CSI acquisition technique ("legacy CSI reporting" in English, i.e. to a Type-I or Type-II codebook for example) if the performance of the UE's AI / ML predictions is not satisfactory.

[0012] The decision to revert to a classic CSI acquisition technique has consequences for the accuracy of the CSI estimation benefiting the base station, and incidentally, for the performance of the MIMO techniques implemented by the base station.

[0013] Furthermore, a considerable amount of time can elapse before the network is able to detect that the AI / ML predictions made by the UE are incorrect, and that the CSI estimate sent by the UE to the base station is not suited to its environment. Indeed, to obtain an accurate performance metric, a significant number of measurements are required after the AI / ML prediction is activated at the UE level. Description of the invention

[0014] The invention makes it possible, in particular, to overcome the aforementioned drawbacks by proposing a method for transmitting information from a first device to a second device via a communications network, this method comprising: a step of receiving reference signals transmitted by the second device; and a step of transmitting to the second device information relating to predictions of a state of a propagation channel between the second device and the first device, made by the first device from the received reference signals using at least one prediction technique based on machine learning and another prediction technique.

[0015] Correspondingly, the invention also relates to a first device capable of communicating with a second device via a communications network, said first device comprising: a receiving module, configured to receive reference signals transmitted by the second device; and a transmission module, configured to transmit to the second device information relating to predictions of a state of a propagation channel between the second device and the first device, made by the first device from the reference signals received using at least one prediction technique based on machine learning and another prediction technique.

[0016] The invention also relates to a method for receiving information from a first device via a communications network, said method being implemented by a second device and comprising: a step of sending reference signals to the first device; and a receiving step, from the first device, of information relating to predictions of a state of a propagation channel between the second device and the first device, made by the first device from the reference signals received using at least one prediction technique based on machine learning and another prediction technique.

[0017] Correspondingly, the invention also relates to a second device capable of communicating with a first device via a communications network, said second device comprising: a sending module, configured to send reference signals to the first device; and a receiving module, configured to receive from the first device information relating to predictions of a state of a propagation channel between the second device and the first device, made by the first device from the received reference signals using at least one prediction technique based on machine learning and another prediction technique.

[0018] The invention therefore advantageously proposes, in order to take full advantage of the accuracy of CSI acquisition techniques based on prediction, to integrate several prediction techniques into the first device, exploiting the reference signals sent by the second device, and to send back to the second device, for each of these prediction techniques, information relating to the predictions made by the first device using the reference signals. The prediction techniques integrated into the first device advantageously include at least one prediction technique based on machine learning and at least one more conventional prediction technique that does not use such learning. For the sake of simplicity, these techniques are hereafter referred to as the AI / ML prediction technique and the non-AI / ML prediction technique, respectively.One such non-AI / ML prediction technique is, for example, a prediction technique based on an autoregressive model known to a person skilled in the art.

[0019] Thanks to the information on predictions made by the first device and transmitted to the second device, the latter can make an informed decision regarding the suitability of choosing, or if necessary, maintaining, a prediction-based channel knowledge acquisition technique, whether AI / ML or not, taking into account the environment of the first device. The invention thus ensures that the implemented channel knowledge acquisition technique is appropriate for the environment of the first device.

[0020] It should be noted that, given the technical problem mentioned above, the invention is particularly applicable when the first device is a user device in a communications network and the second device is a network device such as a base station or, more generally, a network access point. However, the invention is also applicable in other contexts. In particular, it can also be applied to uplink connections, as well as to other devices separated by a propagation channel that must be estimated to enable these devices to communicate with each other.

[0021] Most advantageously, the methods according to the invention can be implemented at any time.

[0022] They can be implemented, in particular, during a preliminary testing phase, before enabling the use of an AI / ML prediction technique at the first device level to estimate the state of the propagation channel. This makes it possible to determine whether the environment of the first device is suitable for the use of such an AI / ML prediction technique.

[0023] Thus, the reception process may further include a step of activating the prediction technique based on machine learning at the level of the first device or said other prediction technique depending on the information received.

[0024] However, the transmission and reception methods according to the invention can also be implemented after the activation of an AI / ML prediction technique, as part of performance monitoring of the latter, to determine whether its use is appropriate compared to another prediction technique or whether it is better not to use any prediction technique at all. For example, a non-AI / ML prediction technique may lead to more reliable predictions compared to an AI / ML prediction technique if the latter is poorly trained, typically if it is trained with data too far removed from the actual propagation conditions of the first device. The invention thus makes it possible to select the most suitable prediction technique at any given time for the environment of the first device without having to change the precoder dictionary, or even to disable prediction altogether.

[0025] There are no limitations attached to the form that the information relating to said predictions takes when transmitted from the first device to the second device.

[0026] Thus, for example, in a particular embodiment, the information relating to said predictions may include prediction quality metrics evaluated for these predictions by the first device.

[0027] Various quality metrics can be considered, including metrics that measure the difference between predictions provided by prediction techniques and ground truth information reflecting the "true" values ​​of the predicted quantities. Examples of such quality metrics include NMSE or SGCS metrics, evaluated over a time interval (or slot) or averaged over a window of a defined dimension (e.g., equal to a plurality of slots), etc.

[0028] Alternatively, it can be envisaged that the information relating to the predictions sent from the first device to the second device includes a prediction quality metric evaluated for one of the prediction techniques considered and a difference between this quality metric and a quality metric evaluated for the other prediction technique considered.

[0029] The quality metrics and / or difference metrics thus reported can also be quantified over a determined number of bits or bytes.

[0030] Alternatively, it can also be considered that the information relating to the predictions includes, for each prediction technique, an index designating in a table known from the first device and the second device, an interval of values ​​in which lies a prediction quality metric evaluated for that prediction technique.

[0031] This reduces the amount of signage needed in the return lane to improve the quality metrics assessed by the first device.

[0032] In yet another embodiment, the information relating to the predictions sent from the first device to the second device may include information representative of said predictions and ground-truth information reflecting the state of the propagation channel (or "ground-truth CSI" in English).

[0033] The information representing the predictions can, for example, be a quantized and compressed form of said predictions to facilitate their transmission to the second device. This could include indices pointing to elements of a precoder dictionary or a specific precoding matrix, such as PMIs pointing to an eType-II Doppler dictionary defined by the 3GPP standard. This embodiment further reduces the signaling required to implement the invention.

[0034] In this embodiment, the reception process can then include a step of estimation by the second device of a prediction quality metric of said prediction technique based on machine learning and of a prediction quality metric of said other technique, based on the information received from the first device.

[0035] As mentioned previously, the information relating to predictions sent from the first device to the second device can be used in different ways by the second device.

[0036] Thus, in a particular embodiment, the reception process includes a configuration step at the first device of a dictionary of precoders used by the first device to send information on the state of the propagation channel to the second device based on information relating to the predictions received.

[0037] For example, in the context of a 5G network, the second device can decide, based on the information received, to configure the first device with an e-Type II Doppler dictionary using an AI / ML prediction technique if a prediction quality metric of the AI / ML prediction technique is better (e.g., higher or lower depending on the metric considered) than the same prediction quality metric evaluated for a non-AI / ML prediction technique. Conversely, the second device can decide, based on the information received, to configure the first device with an e-Type II Doppler dictionary using a non-AI / ML prediction technique if the prediction quality metric of the non-AI / ML prediction technique is better than the prediction quality metric evaluated for the AI / ML prediction technique.It is also possible to consider that the second device configures at the level of the first device a dictionary that does not require prediction of a propagation channel state by the first device, such as for example an eType-II, or Type-I or Type-II dictionary.

[0038] The second device can also take into account other parameters in addition to the information received during its decision-making, such as an estimate of the speed of the first device, etc.

[0039] Furthermore, the configuration of the first device by the second device can be achieved by any means.

[0040] For example, in one particular embodiment, the configuration step includes sending a control message to the first device indicating the dictionary to be used by the first device.

[0041] Such a control message can be, in particular, an RRC (Radio Resource Control) message, in which the dictionary is indicated in an Information Element (IE). Sending an RRC message to indicate the dictionary is particularly well-suited when the dictionary choice is relatively static and is not likely to change too frequently over time.

[0042] However, other control messages adapted to more dynamic changes can also be considered, such as a MAC-CE (Medium Access Control Element) or a DCI (Downlink Control Information) message. These are, of course, only illustrative and non-limiting examples of the invention.

[0043] In one particular embodiment, the transmission and reception processes are implemented by a computer.

[0044] The invention also relates to a computer program on a recording medium, this program being capable of being implemented in a computer or more generally in a first device conforming to the invention and comprising instructions adapted to the implementation of a transmission method as described above.

[0045] The invention also relates to a computer program on a recording medium, this program being capable of being implemented in a computer or more generally in a second device according to the invention and comprising instructions adapted to the implementation of a reception process as described above.

[0046] Each of these programs can use any programming language, and be in the form of source code, object code, or code somewhere between source code and object code, such as in a partially compiled form, or in any other desirable form.

[0047] The invention also relates to an information carrier or a recording medium readable by a computer, and comprising instructions for a computer program as mentioned above.

[0048] The information or recording medium can be any entity or device capable of storing programs. For example, the medium may include a storage means, such as a ROM, for example a CD-ROM or a microelectronic circuit ROM, or a magnetic recording means, for example a hard drive, or a flash memory.

[0049] On the other hand, the information or recording medium can be a transmissible medium such as an electrical or optical signal, which can be carried via an electrical or optical cable, by radio link, by wireless optical link or by other means.

[0050] The program according to the invention can in particular be downloaded onto an Internet-type network.

[0051] Alternatively, the information or recording medium may be an integrated circuit in which a program is incorporated, the circuit being adapted to execute or to be used in the execution of the transmission and reception processes according to the invention.

[0052] The invention also relates to a communication system comprising a first device and a second device according to the invention.

[0053] It can also be envisaged, in other embodiments, that the transmission and reception processes, the first and second devices and the communication system according to the invention have in combination all or part of the aforementioned characteristics. Brief description of the drawings

[0054] Other features and advantages of the present invention will become apparent from the description below, with reference to the accompanying drawings which illustrate an example of an embodiment without being limiting in any way. In the figures:

[0055] [Fig.1] Figure 1 represents a communication system according to the invention, in a particular embodiment;

[0056] [Fig.2] Figure 2 represents the hardware architecture of the communication system devices of Figure 1, in a particular embodiment;

[0057] [Fig. 3] Figure 3 represents in call-flow form the main steps of a transmission process and a reception process according to the invention as implemented by the devices of the system of Figure 1, in a particular embodiment;

[0058] [Fig. 4] Figure 4 represents the different steps implemented by the devices of the system in Figure 1 during the testing phase; and

[0059] [Fig.5] Figure 5 illustrates a reference signal configuration that can be used in the context of the invention.

[0060] Annexes 1, 2 and 3 provide examples of algorithms that can be implemented during the transmission process according to the invention. Description of the invention

[0061] Figure 1 represents, in its environment, a communication system 1 according to the invention, in a particular embodiment.

[0062] In this embodiment, system 1 comprises: at least one base station 2 of a NW communications network; and at least one user device or UE 3 attached to base station 2 and capable of communicating with it via the NW communication network.

[0063] It is assumed here that the NW communication network is a 5G NR access network, as defined by the 3GPP standard, with the necessary adaptations for implementing the invention. Furthermore, in the embodiment envisaged here, the invention is applied in a downlink configuration: in this configuration, the UE 3 includes all the functionalities of a first device according to the invention, and the base station 2 includes all the functionalities of a second device according to the invention.

[0064] These assumptions are not, however, limiting in themselves, and the invention can be applied in other contexts. In particular, it can be applied to other networks, such as a 4G or 6G access network, a proprietary network, etc., but also in uplink, in which case base station 2 includes all the functionalities of a first device as defined in the invention, and UE 3 includes all the functionalities of a second device according to the invention. It is understood that since the invention can be applied to both downlink and uplink, base station 2 and UE 3 can each integrate the functionalities of both a first and a second device according to the invention.

[0065] We consider the example shown in Figure 1 in a MIMO context: base station 2 is equipped with a plurality (N T > 1) of transmitting antennas and the UE 3 is equipped with at least one (N R1) Receiving antennas. The acquisition by base station 2 of knowledge of the propagation channel (CSI) 4 separating it from UE 3 allows it, in a manner known per se, to adapt its transmission strategy and in particular to select a precoder (or equivalently a precoding matrix) from the aforementioned dictionaries. In accordance with the 3GPP standard, knowledge of the downlink propagation channel 4 (i.e., in the direction from base station 2 to UE 3) is acquired by base station 2 via a feedback channel 5 used by UE 3 to send it information relating to the state of the downlink propagation channel.In order to limit the amount of information sent in the return path to base station 2, the 3GPP standard defines three pieces of information, transmitted by UE 3 to base station 2, and representing the state of the channel, namely: a precoding matrix indicator (or PMI) pointing to a precoding matrix belonging to a dictionary of precoders with which base station 2 and UE 3 are configured, a channel quality indicator (or CQI for "Channel Quality Indicator") and a rank indicator (or RI for "Rank Indicator"). It is assumed here that base station 2 and UE 3 are each configured with a plurality of dictionaries (codebooks) defined by the 3GPP standard and in particular with Type-I, eType-II codebooks (with R=1 or R=2 depending on the number of PMI indicators sent by UE 3 to base station 2 for a sub-band) and eType-II Doppler as defined in paragraph 5.2.2 of the 3GPP TS 38 document.214 cited previously. It is noted that the invention applies regardless of the type of feedback envisaged for the CSI, whether it is frequency-selective (in which case a PMI indicator and a CQI indicator are reported per sub-band and the RI indicator is common to all sub-bands (of which there are N3)) or broadband (in which case a single PMI / CQI / RI triplet is reported for the whole band).

[0066] These assumptions are linked to the 5G NR network context considered here and are not limiting in themselves. It is of course possible to consider other ways for the UE 3 to transmit channel state information to base station 2 depending on the context in which the invention is applied, as well as other dictionaries based or not on predictions, or even other variants of the aforementioned dictionaries.

[0067] In the embodiment described here, base station 2 and UE 3 have the hardware architecture of a computer 6 as shown in Figure 2. This computer 6 includes in particular a processor PROC, a random access memory MEM, a read-only memory ROM, a non-volatile memory NVM, and COM means of communication adapted to communicate in particular on the NW communication network.

[0068] The non-volatile NVM memory of base station 2 constitutes a storage medium according to the invention, readable by the PROC processor of base station 2, and on which a PROG2 program according to the invention is stored. It is also assumed here that the codebooks (generally referred to as CODEB in Figure 2) with which base station 2 is configured are stored in its NVM memory.

[0069] The PROG2 program includes instructions defining the main steps of a reception process according to the invention, and more specifically defines the functional modules of base station 2 that rely on and / or control all or part of the PROC, MEM, ROM, NVM, and COM elements of computer 6 mentioned previously. These functional modules include, in particular, in the embodiment described here, as illustrated in Figure 1: a transmitting module 2A, configured to send reference signals to UE 3 to enable it to estimate the state of the downlink channel 4 between base station 2 and UE 3. In the context of a 5G NR network as envisaged here, these reference signals are CSI-RS type signals. Their configuration for implementing the invention is detailed further below; and a receiving module 2B, configured to receive from UE 3, via the return channel 5, information relating to predictions of the state of the downlink channel 4, made by UE 3 from the received reference signals.

[0070] The functions of modules 2A and 2B of base station 2 are described in more detail later.

[0071] The non-volatile NVM memory of UE 3 constitutes a recording medium according to the invention, readable by the PROC processor of UE 3, and on which a PROG3 program according to the invention is stored. It is also assumed here that the CODEB codebooks (identical to the CODEB codebooks stored by base station 2) with which UE 3 is configured are stored in its NVM memory.

[0072] The PROG3 program includes instructions defining the main steps of a transmission method according to the invention, and more specifically defines the functional modules of the UE 3 that rely on and / or control all or part of the PROC, MEM, ROM, NVM, and COM elements of the computer 6 mentioned previously. These functional modules include, in particular, in the embodiment described here, as illustrated in Figure 1: a 3A receiving module, configured to receive CSI-RS reference signals transmitted by base station 2 via its 2A sending module; A module 3B evaluates information related to predictions of the propagation channel state 4 between base station 2 and UE 3, made by UE 3 based on received CSI-RS reference signals. In the embodiment described herein, the information evaluated by the evaluation module 3B consists of performance (i.e., quality) metrics for the predictions made by UE 3. According to the invention, UE 3 has, for making such predictions, at least one prediction technique TECH1 based on machine learning (TECH1 is an AI / ML prediction technique) and another prediction technique TECH2 that is not based on machine learning (TECH2 is a non-AI / ML prediction technique). There are no limitations on the prediction techniques TECH1 and TECH2 that can be used by UE 3.For example, the AI / ML technique TECH1 can be an LSTM technique or a transformer-based technique, and the non-AI / ML technique TECH2 can be a technique based on an autoregressive (AR) model, a Kalman filter, etc.; and. a 3C transmission module, configured to transmit to base station 2 the information evaluated by the 3B evaluation module.

[0073] For the sake of simplicity, we consider here two prediction techniques, TECH1 and TECH2, implemented by UE 3. However, a larger number of prediction techniques can be considered, provided that these techniques include at least one AI / ML prediction technique and one non-AI / ML prediction technique. It should be noted that the 3GPP standard, when defining the eType-II Doppler codebook, assumes that UE 3 uses an AI / ML prediction technique to predict the channel at different times and select the precoding matrices for this codebook associated with the predictions obtained. The invention therefore proposes integrating at least one other non-AI / ML prediction technique into UE 3, in addition to the AI / ML prediction technique, to perform channel predictions.Regardless of the prediction technique used, it is assumed here that UE 3 selects the precoding matrices corresponding to the predictions made in the eType-II Doppler dictionary. In other words, the eType-II Doppler dictionary is agnostic with respect to the prediction technique used by UE 3 to predict the propagation channel 4 separating base station 2 from UE 3.

[0074] The functions of modules 3A to 3C of EU 3 are described in more detail later.

[0075] Figure 3 illustrates the main steps of the reception and transmission processes according to the invention as implemented respectively by base station 2 and by UE 3, in a particular embodiment.

[0076] A preliminary discovery phase is assumed by the NW network (and more specifically by base station 2) of the capabilities of UE 3. For this purpose, in the embodiment described here, base station 2 sends UE 3 an RRC control message UECapabilityEnquiry asking for its RRC capabilities, for example during the procedure of attaching UE 3 to the NW access network (step 10).

[0077] UE 3 responds with an RRC UECapabilityInformation control message to base station 2, sending it its capabilities, including the categories of CODEB precoder dictionaries it supports (implements), for example here the Type-I, eType-II and eType-II Doppler dictionaries (step E20).

[0078] Upon receiving this information, Base Station 2 can transmit additional network condition details to UE 3, for example, in an RRC NWSideAdditionalInformation control message (step E30). Such conditions may include information about the antenna configuration of Base Station 2, such as the vertical tilt value of its antennas. This information can enable UE 3 to identify inconsistencies with the data used to train the AI / ML prediction model employed by the TECH1 technique (for example, if the model was trained with data corresponding to a 3° vertical tilt, while the information provided by Base Station 2 in the NWSideAdditionalInformation message indicates a 10° vertical tilt). If necessary, UE 3 can declare that it does not support an AI / ML prediction model trained with the correct parameters.

[0079] The UE 3 can, in turn, send an additional RRC control message, SupportedFunctionality, indicating, where applicable, its support for prediction techniques for prediction-based precoder dictionaries such as the eType-II Doppler dictionary (step E40). According to the invention, the UE 3 is capable of using, to perform predictions of the propagation channel 4 state, at least one AI / ML prediction technique (TECH1 in the example considered here) and at least one non-AI / ML prediction technique (TECH2 in the example considered here). It therefore informs base station 2 that it supports the TECH1 and TECH2 prediction techniques.

[0080] Note that the RRC message names proposed here are given for illustrative purposes only and are not limiting to the invention. Other RRC message names may, of course, be considered as alternatives.

[0081] In the embodiment described here, following the receipt by base station 2 of the RRC control message from UE 3 indicating that it supports an AI / ML prediction technique (and a non-AI / ML prediction technique), base station 2 triggers a test phase φtest to configure the technique to be used by UE 3 to transmit to base station 2 information representative of the state of the propagation channel 4 (dictionary and where applicable prediction technique) (step E50).

[0082] Figure 4 illustrates the main steps of the φtest phase. The φtest phase is triggered here by base station 2 before any prediction technique is activated by base station 2 at the level of UE 3.

[0083] In the embodiment described here, the test phase φtest begins with the configuration, by base station 2, of CSI-RS reference signals to allow the evaluation of the performance of the prediction techniques implemented by UE 3 (step E51). In the embodiment described here, this performance is evaluated using prediction quality metrics (information relating to predictions as defined in the invention), derived by UE 3. However, in another embodiment, such prediction quality metrics can be derived by base station 2 from information relating to predictions made by UE 3, provided to base station 2 by UE 3. Other metrics can also be used to evaluate the performance of the prediction techniques available at UE 3.

[0084] More specifically, in the embodiment described here, Base Station 2 configures the CSI-RS signals as proposed in the aforementioned Rl-2500057 document to monitor the performance of the AI / ML prediction technique within an LCM lifecycle management context, and as illustrated in Figure 5. Sending such reference signals is intended to enable UE 3 to collect measurements both over an observation window (OW) and a prediction window (PW). Thus, Base Station 2 configures at least one pair (i.e.a pair) of signal sets or equivalently of CSI-RS resources, called a sample (“sample” or “example” in English), and comprising: a first set of resources consisting of a number K > 1 of CSI-RS resources used to create inputs of the prediction techniques (or equivalently of the model) TECH1 and TECH2 that we wish to test. The temporal behavior of these K CSI-RS resources can be either periodic (CSI-RS resources sent every P time intervals or slots, where P is an integer greater than or equal to 1), semi-persistent (activation / deactivation of CSI-RS resources via a MAC-CE control element, then periodic sending of activated CSI-RS resources every P slots) or aperiodic (sending a burst of K = {4,8,12} CSI-RS resources spaced 1 or 2 slots apart after receiving an activation signal), in accordance with the predictions of 3GPP TS 38.214; et. A second set of resources consisting of 1 < n < N4 CSI-RS resources used to create a ground-truth label intended to be used here to calculate a performance metric to evaluate the TECH1 and TECH2 prediction techniques being tested. In the example considered here, n = N4.

[0085] Note that it is possible to configure a single sample of CSI-RS resource sets to limit the resources required for the invention. The same applies to the value of n, which can be strictly less than N4. However, configuring a plurality of samples and a second set of N4 resources enhances the reliability of the evaluated performance metrics. Similarly, the same reference signals can be used for the φtest phase and for training the inference model of the AI / ML TECH1 technique.

[0086] With reference to Figure 4, the CSI-RS reference signals thus configured are sent by base station 2 to UE 3 via its sending module 2A (step E52).

[0087] Upon receiving the reference signals via its receiver module 3A, UE 3 estimates the state of propagation channel 4 in a self-knowledgeable manner, based on measurements taken on the observation window OW from the K CSI-RS resources of the first set of each sample, as illustrated in Figure 5 (step E53). These estimates are provided by UE 3 as input to the two prediction techniques TECH1 and TECH2 that it implements to generate predictions of the propagation channel state (step E54). Both techniques provide N4 predictions of the propagation channel 4 state, corresponding to N4 distinct prediction slots. UE 3 also uses measurements taken on the n = N4 CSI-RS resources of the second set of each sample in the prediction window PW to determine ground truth labels reflecting the propagation channel 4 state at the N4 prediction slots (step E55).

[0088] Note that the state of the propagation channel can be predicted in different forms: for example, it can be predicted as the matrix of complex coefficients representing the channel, or as precoding matrix indicators (PMIs) adapted to the channel, etc. The same applies to the corresponding ground truth labels.

[0089] Then, in the implementation described here, UE 3, via its assessment module 3B, evaluates a quality metric for the predictions made using each of the TECH1 and TECH2 techniques (step E56). Different prediction quality metrics can be considered, depending in particular on the form of the predictions considered.

[0090] For example, if the propagation channel coefficient matrix 4 was predicted, the evaluation module 3B can be configured to evaluate, for each technique TECH1 and TECH2 respectively, an NMSE metric between the N4 channel state predictions made by UE 3 using technique TECH1 and TECH2 respectively, and the N4 ground truth labels representing the channel state at the prediction times obtained during step E55. Such an NMSE metric, evaluated for each prediction slot n4 = {1, 2..., N4], and for each sample s, is defined by: fl V- / Vn n ~ Hn S n II \ ) NMSE^ (TECH) = EJ — Y f - ( h to-A IK.dl / J where | | | |F denotes the Froebenius norm, and where H„3, n4 and H® 3i "4th c N R* N T fc denoting the field of complex numbers, represent the matrix of coefficients of the predicted channel and the ground truth label representing the channel respectively, for sub-band n3, prediction slot n4, sample s, and TECH denotes the prediction technique being tested (i.e. TECH=TECH1 or TECH2 in the example considered here).

[0091] Alternatively, when the channel state is predicted as a precoding matrix indicator (PMI), the evaluation module 3B can be configured to evaluate, for each technique TECH1 and TECH2 respectively, an SGCS metric between the N4PMIs predicted by UE 3 using technique TECH1 and TECH2 respectively, and the N4 ground truth labels representing the channel state at the prediction times obtained in step E55. Thus, the SGCS metric is, for each prediction slot n4 = {1, 2, ..., N4], for each sample s, and for each spatial layer l = {1, ..., v}, where v denotes the number of spatial layers considered in the MIMO system, defined by: N32 | l|li v v v n s 3,n4,lj H wn s 3,n4,l, Illl \ 1 3 n3 = l 11 W n3,n4,l 1111 W n3,n4,l 11 / where w^,.^ e < C WrX1is the predicted PMI and w n s 3 lli l is the PMI ground truth label for spatial layer l, prediction slot n4, sample s.

[0092] Of course, other quality metrics can be considered as alternatives.

[0093] Furthermore, quality metrics can be evaluated per prediction slot [n4 = {1, 2..., N4}] as proposed above for NMSE and SGCS metrics, or averaged over the PW prediction window (i.e., over the values ​​of n4 = {1, 2..., N4}). Knowing a quality metric per prediction slot provides more information about the predictions made with either prediction technique and can also be used by the network to configure the number of predictions to monitor (i.e., to appropriately size n or N4).

[0094] Similarly, the SGCS metric can be reported per spatial layer, as in the example above, or averaged across all spatial layers. Knowing a metric for each spatial layer allows the network to adapt the number of spatial layers.

[0095] In addition, quality metrics can be assessed by sample (pair of CSI-RS resource sets), as above, or averaged over multiple samples to obtain more accurate metrics.

[0096] For example, the SGCS metric can be averaged over an S number of samples, over the v spatial layers, over the N3 sub-bands and over the N4 prediction slots as follows: 1 yyyy / \\ w n3, n4 ,i SGCSÇTECH') = SV W4. W3ZJ ZJ ZJ ZJ 4 3 s=l 1=1 n4= l n3= ll ||iv s ,||||wn s 3,n4,l\\ /

[0097] Annex 1 provides, in pseudocode form, an example of an algorithm to calculate (and send back to base station 2) an NMSE metric for each of the prediction techniques TECH1 and TECH2, averaged over S samples, N3 sub-bands and N4 prediction slots.

[0098] The UE 3 then transmits, via its 3C transmission module, the prediction quality metrics thus evaluated for the TECH1 and TECH2 techniques to base station 2 (step E57).

[0099] The quality metrics evaluated for the TECH1 and TECH2 techniques can be transmitted in different forms by the UE 3 to the base station.

[0100] For example, UE 3 can send to base station 2 both metrics as calculated by its 3B evaluation module or quantified.

[0101] Alternatively, it can send the absolute value of the metric evaluated for one of the prediction techniques (possibly quantified), for example TECH1, and the difference between the metric sent and the metric evaluated for the other prediction technique, for example in the illustrative example envisaged TECH2.

[0102] In yet another variant, UE 3 can send, for each metric evaluated for each prediction technique, an index designating, in a table known to UE 3 and base station 2, an interval of values ​​in which the metric in question lies.

[0103] Appendix 2 provides, in pseudocode form, an example algorithm that can be used to implement this variant, for three value intervals indexed by {1, 2, 3} such that the interval indexed by 1 corresponds to metric values ​​between 0 and a threshold thl, the interval indexed by 2 corresponds to metric values ​​between the threshold thl and a threshold th2, and the interval indexed by 3 corresponds to metric values ​​above the threshold th3. In this example, an index of 1 associated with a prediction technique indicates good prediction performance.

[0104] In the embodiment described here, it is the UE 3 via its evaluation module 3B that evaluates performance metrics (prediction quality metrics in the example considered here) for the prediction techniques TECH1 and TECH2, and then transmits them to base station 2. In another embodiment, it can be envisaged that the UE 3 transmits, via its transmission module 3C, to base station 2, other information than performance metrics relating to the predictions it has made with the prediction techniques TECH1 and TECH2 from the reference signals transmitted by base station 2 during step E52.For example, UE 3 can transmit to base station 2 information representing the predictions it has made using the TECH1 and TECH2 prediction techniques, such as the N4 channel matrices predicted with each of the TECH1 and TECH2 prediction techniques at the different prediction slots, or the PMI indices representing the precoding matrices corresponding to these predicted channel matrices, quantized and time-compressed as defined by the 3GPP standard for the eType-II Doppler dictionary to minimize the necessary signaling. It can also accompany this representative prediction information with corresponding ground truth information that reflects the state of propagation channel 4, as previously mentioned.

[0105] Annex 3 provides, in pseudocode form, an example of an algorithm that can be used in this embodiment to report predictions to the base station in the form of PMI indices.

[0106] Information regarding the predictions of the propagation channel 4 state using prediction techniques TECH1 and TECH2 is received by base station 2 (step E58). If this information is transmitted as representative prediction data rather than performance metrics, base station 2 evaluates the performance metrics of each prediction technique based on the received information. It may evaluate NMSE, SGCS, or any other metric for this purpose, in a manner identical or similar to that described previously for step E56. Note that the test phase can be repeated periodically, at predetermined times, or continuously, etc.

[0107] With reference to Figure 3, based on the information available to it, base station 2 configures the precoder dictionary to be used by UE 3 to retrieve future information on the state of propagation channel 4 and / or prediction techniques to be used if the eType-II Doppler dictionary is selected by base station 2 (step E60).

[0108] For example, in the embodiment described here, if we consider a prediction quality metric averaged over the samples, the prediction slots and, where applicable, the layers: if the prediction quality metric of the AI / ML prediction technique TECH1 is better (in this case, given the example metrics provided, is less than or indicates better performance) than the prediction quality metric of the non-AI / ML prediction technique TECH2, and reflects satisfactory prediction performance (e.g., the metric is less than a given THR threshold), base station 2 configures the eType-II Doppler dictionary based on the AI / ML prediction technique TECH1; if the prediction quality metric of the non-AI / ML prediction technique TECH2 is better (in this case, given the proposed example metrics, is lower than or indicates better performance) than the prediction quality metric of the AI / ML prediction technique TECH1, and reflects satisfactory prediction performance (e.g., the metric is below the THR threshold), base station 2 configures the eType-II Doppler dictionary based on the non-AI / ML prediction technique TECH2; If none of the prediction quality metrics evaluated for the TECH1 and TECH2 prediction techniques reflect satisfactory prediction performance (e.g., none are below the THR threshold, given the example metrics considered), then Base Station 2 configures a conventional precoder dictionary, such as a Type-I dictionary, which is more resilient to the Doppler effect. In this case, Base Station 2 can terminate the measurements performed by UE 3 to evaluate the performance of the prediction techniques, as the Type-I dictionary does not require predictions.

[0109] Of course, other selection strategies can be considered as alternatives, depending in particular on the metrics being considered. In particular, when the metrics are not averaged, more complex selection functions can be considered, allowing base station 2 to take into account the individual values ​​of the metrics and their subtleties, and to adjust certain parameters as discussed previously (number of spatial layers, size of the prediction window, etc.).

[0110] The configuration selected by Base Station 2 is then activated by Base Station 2 with UE 3 (step E70). To configure the dictionary to be used by UE 3 for reporting channel state information to Base Station 2, and / or the prediction technique to be used by UE 3 if an eType-II Doppler dictionary is selected (or any other prediction-based dictionary), Base Station 2 sends an RRC control message to UE 3, as described herein. The RRC control message sent during this configuration step includes an information element indicating the dictionary to be used. Furthermore, if a prediction-based dictionary is selected (for example, an eType-II Doppler dictionary), Base Station 2 activates the selected prediction technique from among the AI / ML TECH1 techniques and not AI / ML TECH2.This activation can also be done via an RRC control message (the same as for dictionary configuration or another).

[0111] Alternatively, other control messages can be considered to configure UE 3, such as via a DCI message or a MAC-CE message.

[0112] Upon receiving this control message, UE 3 activates the use of the dictionary and / or prediction technique indicated by base station 2 for its future reports of information relating to the state of propagation channel 4 between base station 2 and UE 3 (step E80).

[0113] It should be noted that following this activation, when a prediction technique is enabled at the UE 3 level, a measurement phase can be considered to monitor the performance of this prediction technique, as proposed, for example, in document Rl-2500057, or a new testing phase such as the one just described, to potentially change the prediction technique and / or dictionary (step E90). The testing and measurement phases can be deployed simultaneously or independently of each other.

[0114] It should also be noted that in the embodiment described here, only two distinct prediction techniques that can be used at the EU 3 level have been considered. However, in another embodiment, a different number of prediction techniques can be considered, including at least one AI / ML prediction technique and one non-machine learning-based prediction technique. The invention then applies in this context by evaluating, for each of the prediction techniques available at the EU 3 level, performance metrics representative of the quality of the predictions offered by each of these techniques. Base station 2 can then select the precoder dictionary and the prediction technique to be used, if necessary, based on the various metrics thus evaluated. A person skilled in the art would have no difficulty adapting the embodiment just described to other prediction techniques.

[0115] Furthermore, in the embodiment described here, the selection of the dictionary and, where applicable, the prediction technique, is based on the performance metrics evaluated for each of the prediction techniques supported by UE 3. Alternatively, it is possible for base station 2 to take into account other factors in addition to these metrics, such as the environment in which UE 3 is located, etc. APPENDIX 1 - Example of an algorithm enabling UE 3 to evaluate and send NMSE performance metrics to base station 2 for the AI / ML prediction techniques TECH1 and non-AI / ML TECH2. Jk Algorithm Inputs: H̃ˢ(observed channel matrix) ∈ ℂ NR×NT×Ns×K , H̃ˢ(ground truth) ∈ ℂ NR×NT×Ns×N4 for s ∈ {1,..., S}. Outputs: NMSE(TECH1) and NMSE(TECH2) Init: NMSE(TECH1) ← 0 NMSE(TECH2) ← 0 1 For s = 1: S do 2 H predTECH1 ← TECH1prediction(H̃ˢ, N4) 3 H predTECH2← TECH2prediction(H̃ˢ, N4) 4 NMSE TECH1_tmp ← 0 5 NMSE TECH2_tmp ← 0 6 For n4= 1: N4do 7 For n3= 1: N3do 8 NMSE TECH1_tmp ← NMSE TECH1_tmp + norm(H predTECH1 (:,:,n3,n4) − H̃ˢ(:,:,n3,n4)) / norm(H̃ˢ(:,:,n3,n4)) 9 NMSE TECH2_tmp ← NMSE TECH2_tmp + norm(H predTECH2 (:,:,n3,n4) − H̃ˢ(:,:,n3,n4)) / norm(H̃ˢ(:,:,n3,n4)) norm «3,114) 10 End for 11 End for 12 NMSE(TECH1) ← NMSE(TECH1) + NMSE TECH1_tmp / N3N413 NMSE(TECH2) ← NMSE(TECH2) + NMSE TECH2_tmp / N3N414 End for 15 NMSE(TECH1) ← NMSE(TECH1) / S 16 NMSE(TECH2) ← NMSE(TECH2) / S 17 Upload NMSE(TECH1) and NMSE(TECH2) to base station 2ANNEX 2 - Example of an algorithm allowing UE 3 to evaluate NMSE performance metrics for AI / ML prediction techniques TECH1 and non-AI / ML TECH2 and to retrieve indices pointing to ranges of values ​​representative of the metrics evaluated. Inputs: thresholds (th1, th2), H̃ˢ(observed channel matrix) ∈ ℂ NR×NT×Ns×K , H̃ˢ(ground truth) ∈ ℂ NR×NT×Ns×N4 for s ∈ {1, Exits: ind TECH1 and ind TECH2 ∈ {1,2,3} 1 NMSE(TECH ) s NMSE(TECH2) ^apply Algorithm 1 (see appendix 1) up to step 16 knowing Hf •= and H i for s E (1,..., S). 2 If WMSEfTECHl) < tftl do 3 ind TECH1 ← 1 4 Else if NMSEÇTECH1) > tftl and WMSE(TECffl) < tft2 do 5 ind TECH1 ← 2 6 Else 7 ind TECH1 ← 3 8 End if 9 ff WMSMTEOÏ2) < tftl do 10 ind TECH2 ← 1 11 Else if NMSE(TECH2) > tftl and NMSE(TECH2) < 02 do 12 ind TECH2 ← 2 13 Else 14 ind TECH2 ← 3 15 End if 16 Go back up ind TECH1 and ind TECH2 at base station 2 APPENDIX 3 - Example of an algorithm allowing UE 3 to send back to base station 2 predictions made via AI / ML TECH1 and non AI / ML TECH2 techniques and ground truth information. Inputs; W-channel fynatrice observed) € HHfi'érité terrain) E N3 r >,v4paramCbmfitrMitioîi p©MF the calculation of rank matarnuni v. 1 hour TECH1 ← TECH1(H̃, N4) # H TECH1 ∈ ℂ NR×NT×Ns×N4 2 hours TECH2 ← TECH2(H̃, N4) #H TECH2 ∈ ℂ NR×NT×Ns×N4 3 W groundTruth ← PMIcalculation(H̃, v, paramCombination) 4 W TECH1 ← PMIcalculation(H TECH1 , v, paramCombination) 5 W TECH2 ← PMIcalculation(H TECH2 , v, paramCombination) 6 Go back up W groundTruth , W TECH1 , and W TECH2 at base station 2 In this algorithm, the "PMIcalculation" function is used to obtain a PMI index pointing to a specific dictionary from a channel matrix (predicted, observed / estimated, "ground truth") known to a person skilled in the art. This function is parameterized using a set of paramCombination configuration parameters relating to the compression / quantization corresponding to the dictionary in question. In the example considered here, the dictionary is an e-Type II Doppler dictionary as defined by the 3GPP standard in document TS 38.214 (Release 18).

Claims

1. Claims

1. A method for transmitting information from a first device (3) to a second device (2) via a communication network, comprising: a reception step (E52) of reference signals transmitted by the second device; and a transmission step (E57) to the second device of information relating to predictions of a state of a propagation channel between the second device and the first device, made by the first device from the reference signals received using at least one prediction technique based on machine learning and another prediction technique.

2. A method for receiving information from a first device (3) by a second device (2) via a communication network, said method comprising: a sending step (E52) of reference signals to the first device; and a receiving step (E58), from the first device, of information relating to predictions of a state of a propagation channel between the second device and the first device, made by the first device from the received reference signals using at least one prediction technique based on machine learning and another prediction technique.

3. A receiving method according to claim 2 further comprising a configuration step (E70) at the first device of a dictionary of precoders used by the first device to send information on the state of the propagation channel to the second device based on the information relating to the predictions received.

4. A receiving method according to claim 3 wherein said configuration step (E70) comprises sending to the first device a control message indicating said dictionary to be used by the first device.

5. A receiving method according to any one of claims 2 to 4 further comprising an activation step (E70) of the machine learning-based prediction technique at the level of the first device or of said other prediction technique based on said received information.

6. A receiving method according to claims 2 to 4 wherein the configured dictionary does not require prediction of a propagation channel state by said first device.

7. A method according to any one of claims 1 to 6 wherein said other prediction technique is a prediction technique based on an autoregressive model.

8. A method according to any one of claims 1 to 7 wherein said prediction information includes prediction quality metrics evaluated for said predictions by the first device.

9. A method according to any one of claims 1 to 7 wherein said prediction information includes a prediction quality metric evaluated for one of said prediction techniques and a difference between this quality metric and a quality metric evaluated for the other prediction technique.

10. A method according to any one of claims 1 to 7 wherein said prediction information includes, for each prediction technique, an index designating in a known table of the first device and the second device, a range of values ​​in which lies a prediction quality metric evaluated for that prediction technique.

11. A method according to claim 1 to 7 wherein said prediction information includes information representative of said predictions and ground truth information reflecting the state of said propagation channel.

12. Device (3) in a communications network, referred to as the first device, comprising: a receiving module (3A), configured to receive reference signals transmitted by a second device in the communications network; and a transmission module (3C) configured to transmit to the second device information relating to predictions of the state of a propagation channel between the second device and the first device, made by the first device from received reference signals using at least one machine learning-based prediction technique and another prediction technique.

13. Device (2) in a communications network, referred to as the second device, comprising: a sending module (2A), configured to send reference signals to a first device in the network; and a receiving module (2B), configured to receive from the first device information relating to predictions of a state of a propagation channel between the second device and the first device, made by the first device from the reference signals received using at least one prediction technique based on machine learning and another prediction technique.

14. Communication system (1) comprising: at least one first device (3) according to claim 12; and at least a second device (2) according to claim 13.

Images