Method for training a machine learning model coupled to a digital twin, method for estimating an environment of a transceiver, and associated electronic devices

A digital twin trained with a machine learning model estimates dynamic entities in the transceiver's environment, enhancing beamforming and network connectivity by addressing the lack of environmental knowledge in existing systems.

WO2026238822A1PCT designated stage Publication Date: 2026-11-19ORANGE SA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/EP2026/063959
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-15
Filing Date
2026-05-13
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Existing wireless communication systems lack detailed knowledge of the transceiver's environment, limiting the optimization of beamforming and network efficiency.

Method used

A method involving a digital twin of the transceiver's environment, trained with a machine learning model to estimate the propagation channel and dynamic entities, allowing for improved beamforming and network connectivity.

Benefits of technology

Enhances the link budget and service area coverage by accurately modeling dynamic entities in the transceiver's environment, improving communication efficiency and reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure EP2026063959_19112026_PF_FP_ABST
    Figure EP2026063959_19112026_PF_FP_ABST
Patent Text Reader

Abstract

The invention relates to a method for training a machine learning model (210) coupled to a digital twin of an environment of a transceiver equipped with a plurality of antennas, the digital twin being used to improve the connectivity of communication devices with the transceiver, the method comprising: estimating the dynamic character of the environment so as to obtain an "estimated environment" (Ēs,ẽ), the estimation including determining a data item representative of the presence of a dynamic entity in a spatial unit resulting from the discretization of the environment; and adapting the machine learning model (210) on the basis of the estimated environment (Ēs,ẽ).
Need to check novelty before this filing date? Find Prior Art

Description

Description Title of the invention: Method for training a machine learning model coupled with a digital twin, method for estimating the environment of a transceiver, and associated electronic devices technical field

[0001] The present invention belongs to the general field of telecommunications, and in particular to wireless communications implemented on radio networks such as mobile networks (e.g., 3G, 4G, 5G, B5G, etc.), Wi-Fi, etc. More specifically, it relates to a method for training a machine learning model coupled with a digital twin of a transceiver's environment. It also relates to a method for estimating a transceiver's environment, as well as electronic devices configured to implement such methods.

[0002] The invention finds a particularly advantageous, although by no means limiting, application in the case of a transceiver equipped with a plurality of antenna elements. Previous technique

[0003] In a wireless communication system, it is common to use a precoder to simultaneously transmit, via the different transmitting antenna elements of a transceiver, one or more signals intended for a communication device of the same system.

[0004] Such a precoder is based on the transceiver's knowledge of the propagation channel (or transmission channel) that separates it from the communication device. It allows the transceiver to deliver data to the communication device with high spectral efficiency, thanks to beamforming (also called "beamforming" in English-language literature) that carries the signals.

[0005] Such a precoder is designed so that the emission of signals in the form of beams is optimized with regard to a determined operating criterion, such as a quality of service criterion of data transmission [maximization of the rate or power received at the level of the receiving device, absence of interference between time symbols at the level of the receiving device, etc.), or a spectral efficiency criterion of the network taking into account interference generated on other users, or even a network energy efficiency criterion, etc.Known examples of precoders include zero-forcing (ZF) precoders, maximum ratio transmission (MRT) precoders, eigenbeamforming precoders, Singular Value Decomposition (SVD) precoders, etc.

[0006] Nevertheless, a more detailed knowledge of the transceiver's environment would allow the transceiver to improve the link budget of the communication devices it serves. Description of the invention

[0007] The present invention aims to overcome all or part of the drawbacks of the prior art, particularly those described above, by proposing a solution based on a digital twin of the transceiver's environment. With such a digital twin, it is then possible to estimate the propagation channel at different points in this environment, or to determine an optimal geographical position for the transceiver that improves the service area covered by that transceiver.

[0008] To this end, and according to a first aspect, the invention relates to a method for training a machine learning model intended to be coupled to a digital twin of a transceiver environment, the method being implemented by an electronic device capable of executing the machine learning model and comprising: - Obtain a data, called "reference data", representative of a propagation channel between the transceiver and a communication device both located in an environment called the "reference" environment; - Estimate by executing the machine learning model a transmitter-receiver environment, called the estimated environment, the learning model taking as input the reference data, and the estimated environment including data representative of the presence of a dynamic entity in the environment; - Adapt the machine learning model according to the estimated environment.

[0009] The invention also relates to a method for training a machine learning model, referred to as the "first machine learning model", intended to be coupled to a digital twin of a transmitter-receiver environment, the method being implemented by an electronic device capable of executing the machine learning model and comprising:

[0010] - an adaptation of the first machine learning model, based on an environment estimated by the electronic device by executing the first machine learning model and using as input a data, called "reference data", representative of a propagation channel between the transmitter-receiver and a communication device both located in a so-called "reference" environment, the estimation of the estimated environment including a determination of a data representative of the presence of a dynamic entity in a spatial unit resulting from a discretization of said environment.

[0011] For the purposes of the invention, the transceiver corresponds, for example, to a base station of a mobile telecommunications network, or to a Wi-Fi access point capable of connecting wireless devices to a wired network or to the Internet.

[0012] In certain implementations, the transceiver is equipped with a plurality of antennas. It can thus perform beamforming detection. However, the invention can also be applied when the transceiver comprises only one antenna, even if the transceiver's detection capabilities are then reduced.

[0013] For the purposes of this invention, a machine learning model is a computer entity, that is, one capable of being executed by an electronic device, which has the particularity of being obtainable through a machine learning or training process and of being able to model an aspect of physical reality. A machine learning model may include a set of learning rules that can be adapted by the training process.

[0014] Such a machine learning model can, for example, estimate a situation based on input data. The model can be used during a learning or training process, where it performs its estimation function. Providing numerous input datasets and considering the estimation results allows the machine learning model to be trained. Once the machine learning model is refined at the end of the training process, it can then be used to estimate a situation based on input data, which corresponds to the intended use of the machine learning model once it has been trained.

[0015] In the invention, the first machine learning model allows for the estimation of the dynamic nature of a transceiver's environment, that is, for determining the presence of dynamic entities in the environment. This estimation is performed by the electronic device by executing the first machine learning model using as input a reference data point representing the propagation channel between the transceiver and a communication device. The presence of dynamic entities in the environment will modify the propagation of waves between the transceiver and communication devices, and it is therefore possible to train a machine learning model to estimate the presence of dynamic entities in the environment using as input data representative of wave propagation. In some embodiments, reference data points are obtained through simulation.11. It is thus possible to obtain a larger quantity of reference data in order to improve the training of the first machine learning model. This estimation can also use as input, in certain embodiments, data representative of the static nature of. the environment. The estimation by the device running the first machine learning model therefore gives an estimated environment.

[0016] Estimating the estimated environment involves determining a representative data point for the presence of a dynamic entity within a spatial unit derived from a discretization of said estimated environment. To estimate the dynamic nature of the environment, the machine learning model estimates an environment in which it specifies whether dynamic entities are present at a given location within the environment. To specify this presence, the environment is discretized; that is, the environment is divided into distinct zones, and the environment's characteristics (e.g., the presence of a dynamic entity) are attached to the different zones, thus discretely, and are not continuously defined for all possible points in the environment.

[0017] In some embodiments, the estimation of the estimated environment includes determining a data point representing the presence or absence of a dynamic entity within a spatial unit resulting from a discretization of said estimated environment. Generally, the estimation produces a data point representing the presence of a dynamic entity, and the absence of a dynamic entity can be inferred from the absence of such a representative data point. However, it may be useful to explicitly specify the absence of a dynamic entity, so that in some embodiments, the data point represents the presence or absence of such a dynamic entity.

[0018] In some embodiments, the estimation of the estimated environment includes a determination of data representative of the presence or absence of at least a portion of a dynamic entity in each spatial unit of at least a subset of a set of spatial units resulting from a discretization of said environment.

[0019] A dynamic entity will not necessarily occupy the entirety of a spatial unit resulting from a discretization of the environment. Therefore, in certain implementations, it is advantageous for the representative data present in the estimated environment to represent the presence of a portion of a dynamic entity. Furthermore, a dynamic entity can occupy several contiguous spatial units. In this case, the representative data represents the presence in each spatial unit of a given subset in order to represent such dynamic entities spanning multiple spatial units. Finally, the estimated environment will include spatial units through which dynamic entities can pass (the area above the ground surface, for example) and other spatial units where the presence of dynamic entities will not be considered (the area sufficiently far from the ground, for example).11 It is therefore interesting to consider subsets of a set of spatial units and not the totality of spatial units resulting from the discretization of the environment.

[0020] The communication device corresponds, for example, to a user terminal, a laptop, a personal assistant, a "smartphone" type mobile phone, or a connected object.

[0021] When an entity in the environment is considered "dynamic," it refers to an entity with characteristics, such as its state and / or position, that can change over time. Such an entity can represent various elements of the environment, such as a vehicle, an individual, or an object. It is therefore understood that the notion of a dynamic entity is defined in contrast to that of a static entity, which designates an element of the environment whose characteristics, such as state and position, remain constant and do not change over time. The notions of static and dynamic entities are to be interpreted in relation to a timescale that corresponds to the implementation of the invention when using the machine learning model to detect the presence of a dynamic entity in the environment of the transmitter-receiver.A static entity corresponds, for example, to an infrastructure - such as a building, a road, a bridge - or to a natural element or an object which is not subject to transformation or movement during the implementation of the invention, such as a mountain, a river, a tree or a parked vehicle.

[0022] For the purposes of the invention, a "spatial unit" corresponds to a voxel (e.g., a unit used to represent volumetric data in three dimensions) or to a set of pixels located in the same plane.

[0023] A "digital twin" differs from a conventional digital model in that it is configured to provide real-time information on the current state of the environment it represents, and / or to simulate a scenario or anticipate situations likely to occur within that environment. The term "environment" used in this document most often refers to a digital representation of a real environment, but can also refer to the physical environment in which a transceiver, communication devices, etc., are located. To fulfill its function, a digital twin relies on a physical model that must be continuously fed with data collected in real time. Therefore, one of the objectives of this invention is to feed a digital twin of a transceiver's environment with data that accurately reflects the dynamic nature of that environment.This data is present in the estimated environment obtained by the first machine learning model.

[0024] The digital twin can be used, in particular, to improve the connectivity of communication devices with the transceiver. Indeed, the representation that the digital twin provides of the transceiver's environment allows for simulations or any other calculations that can improve the transceiver's connectivity with the communication devices present.

[0025] Generally speaking, the steps of a process should not be interpreted as being linked to a notion of temporal succession.

[0026] In particular modes of implementation, the training process may also include one or more of the following characteristics, taken individually or in all technically possible combinations.

[0027] In particular implementation modes, a spatial unit corresponds to a voxel, and the environment is discretized into voxels in an irregular manner.

[0028] In particular modes of implementation, adaptation is implemented in such a way as to minimize a cost function between the estimated environment and the reference environment.

[0029] In certain implementation modes, the data representing the presence or absence of at least a portion of a dynamic entity in each spatial unit is a "one-hot" vector or a numerical value representative of the nature of the material [e.g., metallic, composite, plastic, organic, minerals] of said at least a portion present in the spatial unit.

[0030] In specific implementations, the method further includes the electronic device obtaining, for use as input to the first machine learning model, data representing the static nature of the environment. This data, representing the static nature of the environment, describes the static element(s) of the environment. For example, it takes the form of a three-dimensional matrix, each element of the matrix corresponding to a voxel in the space, to which is associated data representing the presence or absence of at least a portion of a static entity within that voxel.

[0031] In certain implementation modes, the communication device is static within the environment and is, for example, positioned within the environment to enhance the capabilities of the machine learning model. In one particular implementation mode, the process further includes the electronic device obtaining, for use as input to the first model, positioning data for the communication device within the environment.

[0032] In particular modes of implementation, the process further includes an estimation, by the electronic device, by executing the first machine learning model and using the reference data as input, of the estimated environment.

[0033] In particular modes of implementation, the process further includes obtaining, for use by the electronic device, as input to the first machine learning model, a piece of data, called "reference data", representative of a propagation channel between the transceiver and a communication device, both located in a so-called "reference" environment.

[0034] In specific implementation modes, the electronic device is capable of executing a second machine learning model, and the method further comprises:

[0035] - a determination, by the electronic device by executing the second machine learning model and using the estimated environment as input, of a piece of data, called "estimated data", representative of the propagation channel between the transceiver and the communication device located in the estimated environment; and,

[0036] - an adaptation of the second machine learning model, so as to minimize a cost function between the estimated data and the reference data.

[0037] The second machine learning model, like the first, is a computer entity, meaning it can be executed by an electronic device. Its distinguishing feature is that it can be obtained through a machine learning or training process and can model an aspect of physical reality. The electronic device is capable of executing both the first and second machine learning models to carry out the various steps of the processes according to the invention.

[0038] In this embodiment, the second model obtains representative data for the propagation channel from the transmitter-receiver environment. The second machine learning model should therefore obtain, from the estimated environment, an estimated value close to the reference data used as input for the first machine learning model. Minimizing the cost function between the estimated and reference data thus enables the training of the second machine learning model.

[0039] The first and second machine learning models correspond to the same model or to distinct models. The second machine learning model can be combined with the first machine learning model to form a single machine learning model which, when executed by the electronic device and given as input data representative of the propagation channel, is capable of estimating the transceiver environment and, conversely, when given an estimated environment as input, of producing estimated data representative of the propagation channel. In the embodiment where the first and second models correspond to the same model, the adaptation step of the second machine learning model, so as to minimize a cost function between the estimated data and the reference data, will also correspond to an adaptation step of the first model.11. This embodiment involves joint learning of the two machine learning models.

[0040] In certain implementation modes, the training process further includes:

[0041] - a determination, by the electronic device by executing a differentiable ray tracing function and using the estimated environment as input, of a data, called "estimated data", representative of the propagation channel between the transceiver and the communication device located in the estimated environment;

[0042] - an adaptation of the first machine learning model so as to minimize a cost function between the estimated data and the reference data.

[0043] In these embodiments, a ray tracing function provides an estimated value from the estimated environment. The ray tracing function must be differentiable to allow gradient propagation from the cost function between the estimated and reference values. The cost function then improves the fit of the first model only. Indeed, the ray tracing function, which in this embodiment fulfills the function of the second learning model seen in previous embodiments, is not a model that needs to be trained, but is given once and for all. The fit is therefore limited here to the first model.

[0044] In particular implementation modes, the determination of the estimated data is implemented by positioning the transceiver and the communication device, in the estimated environment, at positions identical to those taken in the reference environment, and by applying a differentiable ray tracing method.

[0045] In specific implementation modes, the first machine learning model is a recurrent neural network.

[0046] In particular implementation modes, the recurrent neural network, when deployed, comprises a plurality i = 1.. M of subnetworks of neurons linked together and having the same structure, the process further comprising the following steps:

[0047] - obtaining, by each subnetwork i = 2.. M of neurons, the reference data and a partial estimate of the environment generated by the subnet i — 1 immediately preceding subnet i, and

[0048] - a determination, by each subnetwork i = 2.. M of neurons, of a partial estimate E s e i of the environment based on the reference data and the partial estimate generated by subnetwork i — 1,

[0049] and the partial estimate E s e M the environment determined by the M-th subnetwork corresponds to the estimated environment considered during the adaptation of the first machine learning model.

[0050] As mentioned previously, partial environment estimations are performed by the electronic device running machine learning models, which in this case are subnetworks of interconnected neurons. These partial models take as input the reference data and the partial estimation generated by the previous subnetwork. The partial estimations thus produce a series of estimated environments that include, as seen previously, determinations of data representing the presence of a dynamic entity within a spatial unit derived from a discretization of said estimated environment. The partial estimations are therefore partial estimations of the dynamic nature of the transceiver's environment.

[0051] In particular modes of implementation, the process further includes a determination, by each subnetwork i = 1.. M — 1 of neurons, of a data, called "estimated partial data", representative of the propagation channel between the transceiver and the communication device located in the estimated environment, from the partial estimation of the environment determined by each subnetwork i = 1.. M — 1 of neurons;

[0052] and, the estimated partial data is further provided as input to subnetwork i+ 1 immediately following the i subnetwork of neurons.

[0053] In particular modes of implementation, the process further includes, for each subnetwork i = 1.. M of neurons, an adaptation of the subnetwork i = l.. M of neurons, so as to minimize a cost function between the partial estimation and the reference environment.

[0054] In specific implementation methods, the process further includes:

[0055] - a determination, by each subnetwork j = 1.. M of neurons, of a data point, called "estimated partial data", representative of the propagation channel between the transceiver and a communication device, based on the partial estimation of the environment determined by each subnetwork i = 1.. M — 1; and,

[0056] - an adaptation of the subnetwork j = 1.. M of neurons, so as to minimize a cost function between the estimated partial data and the reference data.

[0057] In specific implementation modes, the machine learning model is an autoencoder, the estimation step corresponding to an encoding of the reference data, so as to obtain an encoded representation corresponding to the estimated environment; the method further comprising:

[0058] - a decoding of the estimated environment, so as to obtain a data, called "estimated data", representative of the propagation channel between the transmitter-receiver and the communication device located in the estimated environment;

[0059] and the adaptation step corresponds to an adaptation of the encoder, so as to minimize a cost function between the estimated data and the reference data.

[0060] As mentioned previously, the machine learning model, in this case an autoencoder, is executed by the electronic device to first estimate the environment, thus obtaining an estimated environment. The same model can then be executed, again by the electronic device, to decode the estimated environment, yielding an estimated data point. The adaptation step is performed to minimize a cost function between the data estimated by the decoding and the reference data used as input for the encoding.

[0061] In specific implementation modes, the machine learning model is a reinforcement learning model comprising an agent and an interaction environment, and the process further includes the following steps, implemented iteratively until a stopping criterion is reached:

[0062] - a partial estimate, by the agent and based on the reference data, of the environment;

[0063] - a determination, based on the interaction environment and the partial estimation of the environment, of a data point, called "estimated partial data", representative of the propagation channel between the transceiver and the communication device located in the partially estimated environment; and,

[0064] - a revision of a strategy applied by the agent, so as to maximize a reward function representative of a similarity between the estimated partial data and the reference data;

[0065] and when the stopping criterion is reached, the obtaining and estimation steps are implemented by the agent by applying the revised strategy, and the adaptation step is an adaptation of the agent, implemented so as to maximize a reward function representative of a similarity between the estimated environment and the reference environment.

[0066] The machine learning model used here is a reinforcement learning model. The electronic device is capable of executing this reinforcement learning model in order to train it. The agent's partial estimation of the environment corresponds to an execution of the machine learning model using the reference data as input; this partial estimation yields an estimated environment. As before, the estimated environment includes data representing the presence of a dynamic entity, data determined during the estimation; the estimated environment is therefore an estimate of the dynamic nature of the environment.

[0067] In specific implementation modes, the reward assigned by the reward function represents a similarity between the environment estimated and the reference environment, depends on the number of iterations performed before reaching the stopping criterion.

[0068] In specific implementation modes, at a given time t, the reference data is a time series which represents a temporal evolution of the propagation channel between the transceiver and the communication device located in the reference environment.

[0069] According to a second aspect, the invention relates to a method for estimating the environment of a transceiver implemented by an electronic device capable of executing a machine learning model and comprising:

[0070] - obtaining data representative of a propagation channel between the transceiver and a communication device, both located in the environment; and,

[0071] - an estimation, by the electronic device by executing the trained learning model by implementing the training method according to the invention and using as input the data obtained, of an environment, called the estimated environment;

[0072] - a provision of the estimated environment as input to a digital twin of the transceiver's environment.

[0073] This second aspect of the invention corresponds to the use of a machine learning model trained according to one of the training methods which correspond to the first aspect of the invention.

[0074] The second aspect of the invention is a method for estimating the environment of a transceiver, thereby providing an estimated environment of the transceiver to a digital twin of the environment. Providing the estimated environment to the digital twin will enable the implementation of one or more other methods to improve the connectivity of the communication device with the transceiver.

[0075] The estimation process is implemented by an electronic device capable of running a machine learning model. The learning model performs an estimation of the dynamic nature of the environment; the estimation of the estimated environment includes determining a data point representative of the presence of a dynamic entity within a spatial unit resulting from a discretization of said environment.

[0076] In one embodiment of this second aspect, the machine learning model was trained according to one of the training methods of the first aspect of the invention.

[0077] Providing the estimated environment to the digital twin therefore allows the latter to have an estimate of the dynamic nature of the environment, so as to improve the connectivity of the communication device with the transmitter-receiver.

[0078] The estimation of the estimated environment uses as input data representative of a propagation channel between the transceiver and a communication device, both located in the environment. The machine learning model was trained to use such representative wave propagation data as input in order to detect dynamic entities in the environment, and thus provide an estimate of the dynamic nature of the environment in the form of an estimated environment. In one embodiment, the machine learning model was trained by the methods of the first aspect of the invention.

[0079] In particular modes of implementation, the process further includes a selection of data representative of communication channels between the transceiver and different communication devices located in the environment, so as to improve a diversity of the communication channels considered.

[0080] According to a third aspect, the invention relates to a computer program comprising instructions for implementing a training method or an estimation method according to the invention, when said program is executed by a processor.

[0081] This program 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.

[0082] According to a fourth aspect, the invention relates to a computer-readable recording medium on which the computer program according to the invention is recorded.

[0083] According to a fifth aspect, the invention relates to a computer program comprising instructions for implementing a training method or an estimation method according to the invention, when said program is executed by a processor.

[0084] The information or recording medium can be any entity or device capable of storing the program. 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.

[0085] On the other hand, the information or recording medium can be a transmissible medium such as an electrical or optical signal, which can be transmitted via an electrical or optical cable, by radio, or by other means. The program according to the invention can, in particular, be uploaded to a network such as the Internet.

[0086] Alternatively, the information or recording medium may be an integrated circuit in which the program is incorporated, the circuit being adapted to execute or to be used in the execution of the process in question.

[0087] According to a sixth aspect, the invention relates to an electronic device capable of executing a machine learning model, configured to implement a training method according to the invention and / or an estimation method according to the invention. Brief description of the drawings

[0088] 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:

[0089] [Fig.1] Figure 1 is an example of a communication system in which a method for estimating the environment of a transceiver is implemented;

[0090] [Fig.2] Figure 2 represents modules embedded in an electronic device for training a machine learning model, according to an example of implementation of the invention;

[0091] [Fig.3] Figure 3 schematically represents an example of the hardware architecture of an electronic device for training a machine learning model;

[0092] [Fig.4] Figure 4 represents modules embedded in an electronic device for estimating the dynamic nature of the environment of a transceiver, according to an example of an implementation of the invention;

[0093] [Fig.5] Figure 5 schematically represents an example of the hardware architecture of an electronic device for estimating the dynamic nature of the environment of a transceiver;

[0094] [Fig.6A] Figure 6A represents a method for training a machine learning model, according to a first example of implementation;

[0095] [Fig.6B] Figure 6B represents, in the form of a flowchart, the main steps of the process of training a machine learning model from Figure 6A;

[0096] [Fig.7] Figure 7 represents a method for training a machine learning model, according to a second implementation example;

[0097] [Fig.8] Figure 8 represents a method for training a machine learning model, according to a third implementation example;

[0098] [Fig.9] Figure 9 represents a method for training a machine learning model, according to a fourth implementation example;

[0099] [Fig.10] Figure 10 represents a method for training a machine learning model, according to a fifth implementation example;

[0100] [Fig.11 A] Figure 11A represents the 1.. N — 1 iterations of a machine learning model training process, according to a sixth implementation example;

[0101] [Fig. 11B] Figure 11B represents iteration N of a machine learning model training process, according to a sixth implementation example; and,

[0102] [Fig.12] Figure 12 represents, in the form of a logic diagram, a method for estimating the environment of a transmitter-receiver, according to an example of implementation. Description of the implementation methods

[0103] The terms "first[s]" [or first(s)], "second s)", etc. are used in this document by arbitrary convention to enable the identification and distinction of different elements (such as models, devices, messages, etc.) considered in the embodiments described below, and do not imply any particular sequencing unless explicitly stated.

[0104] Figure 1 is an example of a communication system in which a method for estimating the environment of a transceiver is implemented.

[0105] This system includes a wireless transceiver 10 connected to an electronic device 200, 300, and a plurality of communication devices 20. The functionalities of the electronic device 200, 300 are described in more detail with reference to figures 2 to 5. A communication device 20 corresponds, for example, to a user terminal, a laptop, a personal assistant, a "smartphone" type mobile phone or a connected object.

[0106] In this embodiment, and for the sake of simplicity in the description, the communication system is considered to comprise only the wireless transceiver 10 and three user terminals 20. However, it should be noted that there is no limitation on the number of wireless transceivers 10 or the number of user terminals 20. The following developments can indeed be easily generalized by a person skilled in the art if a different number of devices are considered.

[0107] In the example presented here, the wireless transceiver 10 comprises a plurality N of antenna elements, N being an integer strictly greater than 1, arranged in the form of an array of antenna elements and configured to transmit and receive beams. These beams serve as the medium for the simultaneous transmission of several data streams to serve one or more specific portions of space. It should be noted that the number of antenna elements equipping the wireless transceiver 10 is not a limiting factor of the invention.

[0108] In this embodiment, and for the sake of simplicity in the description, it is assumed that each communication device 20 is equipped with only one antenna, and therefore they do not communicate in full-duplex. It should be noted, however, that there is no limitation on the number of antenna elements of the communication devices 20.

[0109] In this embodiment, and for the sake of simplicity, it is also assumed that communication between the transceiver 10 and the communication devices 20 is unidirectional. However, the following developments can be readily generalized by those skilled in the art to cases where bidirectional communication is implemented between the transceiver 10 and the communication devices 20.

[0110] The transceiver 10 and the communication devices 20 belong to a wireless communications network [not shown in Figure 1] and are capable of communicating with each other in a frequency band associated with this wireless communications network. For the remainder of this description, this telecommunications network is considered, without limitation, to be a 5G mobile network.

[0111] It should be noted, however, that the invention remains applicable to other types of telecommunications networks, such as GSM, 3G (the third generation of mobile network standards), 4G (the fourth generation of mobile network standards), 5G (the fifth generation of mobile network standards), and / or B5G (Beyond 5G). The invention also remains applicable to Wi-Fi (Wireless Fidelity), WiMAX (Worldwide Interoperability for Microwave Access), satellite internet access, and other networks. Generally speaking, there are no limitations on the nature of the telecommunications network that can be considered within the scope of the present invention. Furthermore, the invention is applicable regardless of the nature of the data transmitted over this network.

[0112] As illustrated in Figure 1, the wireless transceiver 10 is a base station, for example of the "lAB-donor" type, connected to the core network. A base station is sometimes called a "nodeB" in 3G networks, an "eNodeB" according to the LTE standard (acronym for "Long Term Evolution"), and a "gNodeB" in 5G networks.

[0113] It should be noted, however, that the invention is applicable regardless of the nature of the wireless transceiver 10, provided that it is capable of wireless communication. Thus, as mentioned previously, there is nothing to preclude considering the wireless transceiver 10 as an access point of a local area network (LAN), such as a wireless LAN, for example, conforming to the IEEE 802.11 standards.

[0114] The transceiver 10 is positioned in a geographical environment, in this case an urban one. As discussed in more detail below, this environment plays a crucial role in telecommunications network coverage, and consequently in the reliability of the services offered.

[0115] As illustrated in Figure 1, the environment is composed of static and dynamic entities. When an entity in the environment is considered "dynamic," it refers to an entity with characteristics, such as its state and / or position, that can change over time. Such an entity can represent various elements of the environment, such as a vehicle (e.g., a truck or a car), an individual, or an object. It is therefore clear that the concept of a dynamic entity is defined in contrast to that of a static entity, which designates an element of the environment whose characteristics, such as state and position, remain constant and do not change over time, following a timescale corresponding to the implementation of the invention.A static entity corresponds, for example, to an infrastructure - such as a building, a road, a bridge - to a natural element or an object which is not subject to transformation or movement during the implementation of the invention, such as a mountain, a river, a tree or a parked vehicle.

[0116] The electronic device 300 includes at least one computer representation of the environment, a digital twin, which is not shown in Figure 1. This digital twin is used by the transmitter / receiver 10 to improve its communication with the communication devices 20, or for other functions such as accident detection. The device 300 is capable of running one or more machine learning models (not shown in Figure 1) that can feed the digital twin data representative of dynamic entities present in the environment, i.e., capable of estimating the dynamic nature of the environment and thus providing an estimated environment. The estimated environment allows the digital twin to improve its functions. The machine learning model(s) use as input data representative of wave propagation in the environment.The electronic device 200 is also capable of running one or more machine learning models in order to train them using real data or data from simulations. These simulations can also use a digital twin such as the one used by the electronic device 300, or even this specific digital twin. Computer representations of the real environment are often referred to as the environment for simplicity. The invention relates to a method for training a machine learning model and a method for estimating the environment using a machine learning model trained according to the training method.

[0117] Figure 2 represents modules embedded in an electronic device for training a machine learning model, according to an example of implementation of the invention.

[0118] As illustrated in Figure 2, the electronic training device 200 comprises a machine learning model 210 and includes a MODJJPD module for adapting the machine learning model 210, depending on the estimated environment E s e by this electronic device 200 by executing the first machine learning model 210 and using as input a data, called "reference data", representative of a propagation channel between the transceiver and a communication device both located in a so-called "reference" environment, the estimation of the estimated environment including a determination of a data representative of the presence of a dynamic entity in a spatial unit resulting from a discretization of said environment.

[0119] The functionalities of this module are described in more detail below with reference to different implementation methods.

[0120] Furthermore, in specific implementation modes, the machine learning model is implemented in the form of a neural network [convolution, perceptron, autoencoder, recurrent, etc.].

[0121] 11. It is also important to note that there are no limitations on the type of training technique used to obtain the machine learning model. Any technique implementing a machine learning algorithm and providing, as output, an estimate of the dynamic nature of the transceiver's environment based on a reference data point representing a propagation channel between the transceiver and a communication device, both located in that environment, can be considered within the context of the invention (e.g., support vector machine, logistic regression, etc.). In other words, the machine learning model is independent of the training method used to train it.

[0122] In addition, any training criterion known to a person skilled in the art can be considered during the training phase of this machine learning model, such as the least squares method or cross-entropy minimization.

[0123] Figure 3 schematically represents an example of the hardware architecture of an electronic device 200 for training a machine learning model. The electronic device 200 is capable of executing the machine learning model 210 to perform a training process according to the invention.

[0124] As illustrated in Figure 3, the electronic training device 200 has the hardware architecture of a computer. Thus, the electronic training device 200 includes, in particular, a processor 1, a random access memory 2, a read-only memory 3 and a non-volatile memory 4. It also has communication means 5.

[0125] The read-only memory 3 of the electronic training device 200 constitutes a storage medium according to the invention, readable by the processor 1, on which a computer program PROG_TR according to the invention is stored, comprising instructions for executing steps in the training process of a machine learning model according to the invention. The PROG_TR program defines functional modules of the electronic training device 200, which rely on or control the hardware elements 1 to 5 of the electronic training device 200 mentioned above. One of the functional modules is illustrated in Figure 2, without limitation, and is described in more detail below with reference to different implementation methods.

[0126] In the implementation modes described below, the communication means 5 enable the electronic training device 200 to obtain the reference data representing the propagation channel between the transceiver and the communication device. To this end, the communication means 5 include a wired or wireless communication interface capable of implementing any suitable communication protocol.

[0127] Figure 4 represents modules embedded in an electronic device 300 for estimating the dynamic nature of the environment of a transmitter-receiver, according to an example of implementation of the invention.

[0128] As illustrated in Figure 4, the electronic estimation device 300 comprises a machine learning model 310 previously trained according to the training method of the invention. This machine learning model 310 includes, in particular:

[0129] - a MOD_OBT module for obtaining representative data from a propagation channel between the transceiver and a communication device, both located in the environment;

[0130] - a MOD_EST estimation module by executing the machine learning model trained by implementing the training method according to the invention and using as input the data obtained from an environment, called the estimated environment;

[0131] - a MODJN module for supplying the estimated environment as input to a 320 digital twin of the transceiver environment.

[0132] Their functionalities are described in more detail below with reference to different implementation methods.

[0133] The electronic estimation device 300 also includes the previously mentioned digital twin 320 of the environment. As previously discussed, unlike conventional digital modeling, a digital twin is configured to provide real-time information on the current state of the environment it represents, and / or to simulate a scenario or anticipate certain situations that may occur within that environment. To do this, a digital twin relies on a physical model that must be continuously updated with data collected in real time. The digital twin 320 therefore has a representation of the static environment, or more precisely, of the static elements of the environment of the transceiver 10. However, this representation, limited to the static environment, lacks the precision required for the intended uses of the digital twin 320.

[0134] Therefore, one of the objectives of this invention is to power this digital twin 320 of an environment of a transceiver with data produced by the trained model 310 and faithfully reflecting the dynamic nature of this environment.

[0135] Once fed with the estimated environment produced by the trained model 310, the estimated environment comprising data representative of the presence of dynamic entities at spatial units of the environment, the digital twin 320 can be used by the transceiver 10 for various functionalities. For example, it can be used to improve the connectivity of the transceiver 10 with communication devices 20 using beamforming techniques. Simpler uses of the digital twin 320, thus fed with an estimated environment, include traffic accident detection, in the case where the transceiver 10 is an element of a mobile telecommunications network, or the presence detection of an individual, in the case where the transceiver 10 is a home gateway for a wireless local area network.

[0136] Figure 5 schematically represents an example of the hardware architecture of an electronic device 300 for estimating the dynamic nature of the environment of a transceiver.

[0137] As illustrated in Figure 5, the electronic estimation device 300 has the hardware architecture of a computer. Thus, the electronic estimation device 300 includes, in particular, a processor 1, random access memory 2, read-only memory 3, and non-volatile memory 4. It also has communication means 5.

[0138] The read-only memory 3 of the electronic estimation device 300 constitutes a storage medium according to the invention, readable by the processor 1, on which is stored a computer program PROGJNF according to the invention, comprising instructions for executing steps in the process of estimating the environment of a transceiver according to the invention. The PROGJNF program defines functional modules of the electronic estimation device 300, which rely on or control the hardware elements 1 to 5 of the electronic estimation device 300 mentioned above. These modules Functional features are illustrated in Figure 4 by way of no limitation, and are described in more detail below with reference to different modes of implementation.

[0139] Training

[0140] Figure 6A represents a method for training a machine learning model 210, according to a first example of implementation.

[0141] More precisely, the machine learning model is trained to enrich a model of the transmitter-receiver environment comprising only static entities, with dynamic elements also present in that environment.

[0142] As mentioned previously, this environment is discretized into a set of spatial units organized in the form of a matrix. In the present embodiment, and for the sake of simplicity, we consider the case of an environment represented as a 3D voxel grid.

[0143] In one particular implementation mode, the grid is regular. Alternatively, the grid is irregular and is then made up of voxels of varying sizes, arranged in a non-uniform way.

[0144] The difference in resolution is adjusted, for example, according to the sensing capabilities of the transceiver: larger voxels are used in areas requiring low precision, while smaller voxels are used in areas where detection must be optimal. Resolution can also be modulated according to specific needs: smaller voxels at sidewalk level for a more detailed representation of children or animals, and larger voxels on roads, where vehicles do not require high-resolution representation. In this case, a compromise can be reached between system complexity and the required accuracy.

[0145] According to a particular implementation, sets (or "batches" in Anglo-Saxon terminology) of scenes representing realistic situations are generated by simulation. Dynamic entities, such as pedestrians and cars, are placed probabilistically within a 3D grid. Thus, the ZI For example, vehicles are positioned on roads, and pedestrians on sidewalks. Various situations likely to cause changes in the propagation channel are considered to train this machine learning model. Thus, pedestrians and vehicles of varying sizes and in different geographical locations are considered. These vehicles are equipped with communication devices that can be simulated in order to collect information representative of a propagation channel between an environmental transceiver and these communication devices. One possible implementation involves using the 320 digital twin, illustrated in Figure 4, to perform these simulations.

[0146] In one particular implementation, some user terminals are positioned inside buildings, and the environment model accurately reproduces the layout within those buildings. Alternatively, simple presence models, located near exterior walls and windows on different floors, are considered.

[0147] The transceiver is also positioned within this environment. This transceiver combines sensing and communication capabilities within a single infrastructure, enabling more efficient use of network resources. In other words, this transceiver implements both a sensing function within its environment and a communication function. This emerging technology is sometimes called "Integrated Sensing and Communication" (1SAC) and relies on the use of a single frequency band and hardware infrastructure, thereby reducing the costs and operational complexities associated with managing two separate systems. Thus, the same radio signal is used not only to transmit information but also to detect objects or events within the environment.

[0148] In a particular implementation, the transceiver is positioned to optimize network coverage and / or its sensing capabilities. These sensing capabilities rely on the transceiver transmitting waves to determine information about its environment, such as the presence of dynamic entities. To do this, the transceiver uses information known as Channel State Information (CS1), which describes how a signal propagates, taking into account factors such as diffusion, attenuation, and / or signal power loss with distance.

[0149] In a particular implementation mode, an optimal position of the transmitter-receiver is determined using the 320 digital twin and "offline" simulations, prior to deployment.

[0150] In a particular implementation mode, the transceiver can only be installed in specific environmental positions, such as on the roof of a building or on a previously installed pylon, and only these positions are then tested.

[0151] For each possible transceiver position, the associated coverage map and a representative map of the detection capabilities are generated and then evaluated. In a particular implementation mode, the voxel positioning error and / or the error in the radio properties of the materials are considered to evaluate the detection capabilities.

[0152] The process further includes a step in which a ray tracing method is implemented for the ray emitted by the transceiver towards the communication device(s). The ray tracing takes into account the effects of reflection, diffraction, refraction, and / or scattering. Considering a transceiver with beamforming capabilities, this yields a triplet [CS1, UE, beam], where CS1 represents the state information of the propagation channel between the transceiver and the communication device.

[0153] The ray tracing methods considered here are deterministic, meaning that if the environment is known, the propagation channel between a transmitter and a receiver will also be known. Conversely, knowledge of a propagation channel can be used to determine the environment, and therefore the presence and position of dynamic elements within it. However, this inverse operation, performed by the machine learning model being trained, can lead to multiple solutions. It is shown below that elements derived from the time sequences allow us to simplify the possible solutions, ultimately arriving at a single solution.

[0154] In a particular implementation mode, the communication device has beamforming capabilities, and the propagation channel state information then also depends on the receiving beam.

[0155] In a particular implementation mode, the propagation channel state information between the transceiver and various communication devices is considered, and each propagation channel state information is then associated with a unique identifier of the communication device, such as its 1MS1 (International Mobile Subscriber Identity). When the state information is derived from simulations, other identifiers of the simulated communication devices can be used.

[0156] In the rest of the description, we call H s e the representative matrix of the propagation channel between the transceiver and one or more communication devices, for a sample e of a given set ["batch") s, this matrix being generated from the previously mentioned propagation channel state information.

[0157] This matrix H s eis provided as input to the machine learning model 210 during a step referenced S100 in Figure 6B. The machine learning model 210 is then trained so as to generate, when the model is executed, during a step referenced S110 in Figure 6B, an environment called "estimated" E s e This "estimated" environment E s e includes at least one data point representing the presence of a dynamic entity within a spatial unit resulting from a discretization of said environment. In an example implementation, the estimated environment associates with each voxel of the environment a data point representing the presence or absence of at least a portion of a dynamic entity within that voxel, based on this matrix El s e In order for the estimated environment to be used later in conjunction with a digital twin, it is necessary to take into account the static elements of the environment. The estimated environment E s eIt can thus integrate these static elements, or limit itself to representing only the dynamic elements of the environment. Consequently, the training of the learning model can therefore use or not use the static elements of the environment as input.

[0158] In the present embodiment, we consider the case of a machine learning model 210 taking the form of a supervised neural network.

[0159] In a particular implementation mode, the data representing the presence or absence of at least one portion of a dynamic entity in each spatial unit is a "one-hot" vector or a numerical value representing the nature of the material [e.g., metallic, composite, plastic, organic, mineral] of said at least one portion present in the spatial unit.

[0160] As is well known, a "one-hot" vector is a suitable representation for encoding classes or categories. This vector is composed of zeros, with the exception of a single value equal to one, positioned at the index corresponding to the class represented.

[0161] The estimated environment E s e is expressed for example in the form of a matrix, each element of the matrix corresponding to a voxel of the discretized environment, and each element being further associated with a "one-hot" vector characterizing the nature of the material [e.g., metallic, composite, plastic, organic, mineral] of at least a portion of a dynamic entity occupying that voxel.

[0162] In a particular implementation mode, this estimated environment E s e includes only the dynamic entities that make up the transceiver's environment. Alternatively, this estimated environment E s e includes the dynamic and static entities of the environment.

[0163] In a particular implementation mode, the machine learning model 210 also takes as input a data E INIT representative of the static nature of the transceiver's environment. In a particular implementation mode, this data E INIT For example, it takes the form of a three-dimensional matrix, each element of the matrix corresponding to a voxel in the space to which is associated a data point representing the presence or absence of at least a portion of a static entity within that voxel. In a particular implementation mode, this data point E INIT is accompanied by a vector describing the geographical position of the transmitter-receiver within the environment.

[0164] As illustrated in Figure 6A, this estimated environment E s eis then used to adapt the machine learning model 210, during a step referenced S120 in Figure 6B. This adaptation is implemented, for example, to minimize a cost / loss function between the estimated environment E s e and the reference environment E s e corresponding to a "ground truth" (according to Anglo-Saxon terminology). The reference environment E s e It can also correspond to one or more simulated environments in order to have a larger amount of data and make the machine learning model training more complete. This cost function (Loss) is, for example, a mean squared error (MSE) loss function. Then a backpropagation step is implemented during which the weights of the neural network are adjusted to minimize this cost function (Loss).

[0165] Finally, it is important to note that this process is iterated over multiple samples e from different sets s, thus allowing the neural network 210 to learn and adjust to improve its performance.

[0166] Figure 6B represents, in the form of a flowchart, the main steps of the process of training a machine learning model from Figure 6A.

[0167] As illustrated in Figure 6B, the learning process includes a first step S100 of obtaining, by the machine learning model 210, a data point El s e , called "reference data", representative of a propagation channel between the transceiver and a communication device, both located in a so-called "reference" environment.

[0168] The process further includes an S110 estimation, using the machine learning model 210 and based on the reference data El s e, of the dynamic nature of the environment, so as to obtain an "estimated environment", said estimation including the determination of a data point representing the presence of a dynamic entity in a spatial unit resulting from a discretization of said environment. In particular implementation modes, the estimation includes the determination of a data point representing the presence or absence of at least a portion of a dynamic entity in each spatial unit of at least a subset of a set of spatial units resulting from a discretization of said environment.

[0169] Finally, the process includes an S120 adaptation of the machine learning model 210, based on the estimated environment E s e .

[0170] Figure 7 represents a method for training a machine learning model, according to a second implementation example.

[0171] This second example includes the steps previously mentioned with reference to Figures 6A and / or 6B, as well as a determination, by a machine learning model 710 and from the estimated environment E s e , of a given El s e , referred to as "estimated data", representative of the propagation channel between the transceiver and the communication device when the latter are located in the estimated environment E s e This estimated data El s e is then considered within the framework of an adaptation step of this machine learning model 710, which is represented in Figure 7 by a return arrow from the Loss H function se , H s e ') to the machine learning model 710. In some embodiments, this Loss H function se , H se') is also used for adapting the machine learning model 210, although this feedback is not shown in Figure 7. This is particularly the case when a single joint learning model fulfills the roles of both learning models 210 and 710.

[0172] In this embodiment, and for the sake of simplifying the description, we consider the special case where the estimated environment E s e includes both static and dynamic entities in the sender-receiver environment.

[0173] The machine learning models 210 and 710 correspond to the same machine learning model (e.g., the same neural network), or to distinct models (e.g., distinct neural networks).

[0174] The estimated data H s eFor example, it takes the form of a propagation matrix between the transceiver and a communication device, when the latter are located in environment E s e estimated by the machine learning model 210 during step S110.

[0175] In a particular implementation mode, the data H s e is determined by positioning the transceiver and the communication device in the estimated environment E s e at positions identical to those taken in the reference environment. Once the transceiver and the communication device are positioned, a differentiable ray tracing method is implemented. Thus, the propagation channel between the transceiver and the communication device, when positioned in the estimated environment, is evaluated.

[0176] As illustrated in Figure 7, this estimated data El s eis then used to adapt the machine learning model 710. This adaptation is, for example, implemented to minimize a costLoss() function between the estimated data El s e and a reference datum corresponding to the matrix El s e previously mentioned. Then a backpropagation step is implemented during which the weights of the 710 neural network are adjusted to minimize this costLoss() function. This costLoss() function is, for example, a mean squared error (MSE) loss function.

[0177] When the data El s e is determined using a ray-tracing function; this function is a fixed element that does not need to be modified by the training process. The difference between the reference data and the estimated data El s eis then used in the adaptation step of the learning model 210 in addition to the difference between the estimated environment and the reference environment. To use a backpropagation technique, the ray tracing function must be differentiable.

[0178] Figure 8 represents a method for training a machine learning model, according to a third implementation example.

[0179] In this example, we consider the special case where the machine learning model 210 is a recurrent neural network. As illustrated by Figure 8, this recurrent neural network 210 comprises, when unfolded, a plurality i = l.. M of subnetworks of neurons 210J = 1.. M chained together and having the same structure.

[0180] The training process according to this third example includes a first step during which the matrix H s eA representative of the propagation channel between the transceiver and a communication device is provided as input to each 210J = 1 neuronal subnetwork. This step is similar to step S100 in Figure 6B. The first 210J = 1 neuronal subnetwork in the chain then generates a first partial estimate E. s e l of the dynamic nature of the environment, which is provided as input to the second subnetwork of neurons 210J = 2 of the chain. This first partial estimate is expressed, for example, in a similar way to that of the estimated environment E s e as previously mentioned, i.e. in the form of a matrix, each element of the matrix corresponding to a voxel of the discretized environment, and each element being further associated with a "one-hot" vector characterizing the nature of the material [eg, metallic, composite, plastic, organic, mineral] of at least a portion of a dynamic entity occupying that voxel.

[0181] This second subnetwork of neurons, 210J = 2, takes as input the matrix El s e previously mentioned, as well as the first partial estimate E s e l , and generates as output a second partial estimate E s e 2 of the dynamic nature of the environment.

[0182] More generally, each 210J = 2..M subnetwork of neurons generates a partial estimate E s e i the dynamic nature of the environment based on the El matrix s e and the partial estimate generated by the subnetwork i — 1.

[0183] Finally, the 210J = M subnetwork of neurons generates a partial estimate E s e U the dynamic nature of the environment based on the H matrix s e and the partial estimate E s e U-1 generated by the subnetwork 210J = M — 1.

[0184] As illustrated in Figure 8, this partial estimate E s e Mis then used to adapt the machine learning model 210. This adaptation is implemented, for example, in a similar way to the step referenced S120 in Figure 6B, that is, in such a way as to minimize a cost function Loss() between the partial estimation E s e M generated by the subnetwork 210_i = M and the reference environment E s e .

[0185] In other words, only the latest E output s e M is considered to calculate the difference(s) and perform backpropagation of the gradient.

[0186] The use of a recurrent neural network is advantageous because it decomposes the initial problem into M subproblems corresponding to M neural subnetworks, and offers an architecture that favors an iterative construction of the matrix E s e M .

[0187] In a particular implementation, the process further includes the steps mentioned with reference to Figure 7. More precisely, the process then further includes a determination, by a machine learning model 710 and from the partial estimation E s e M , of a given El s e M , referred to as "estimated partial data", representative of the propagation channel between the transceiver and the communication device when the latter are located in the partially estimated environment E s e M This estimated partial data El s e M is then considered within the framework of an adaptation step of this machine learning model 710. As mentioned previously, this adaptation is implemented, for example, in such a way as to minimize a cost function (Loss) between the estimated partial data El s e M and the matrix H s epreviously mentioned. Then a backpropagation step is implemented during which the weights of the 710 neural network are adjusted so as to minimize this costLoss() function.

[0188] In a particular implementation method, each partial estimate s e i is provided as input to a subnetwork of 810J = 1.. M — 1 neurons configured to determine a data H s e i , referred to as "estimated partial data", representative of the propagation channel between the transceiver and the communication device located in the estimated environment E s e i And this estimated partial data H s e i is then provided as input to the i + 1 subnetwork immediately following the i subnetwork of neurons.

[0189] Figure 9 represents a method for training a machine learning model, according to a fourth implementation example.

[0190] This fourth example includes the steps previously mentioned with reference to Figure 8. It further includes a step in which each 210J = 1..M subnetwork of neurons is adapted to minimize a cost function between the partial estimate E s e i ) and the reference environment (E s e This step is implemented, for example, in a similar way to the previously mentioned S120 step.

[0191] In a particular implementation mode, each partial estimate E s e i is provided as input to a subnetwork of 810J = 1.. M — 1 neurons configured to determine a data El s e i , referred to as "estimated partial data", representative of the propagation channel between the transceiver and the communication device located in the estimated environment E s e i And this estimated partial data H se iis then provided as input to the i + 1 subnetwork immediately following the i subnetwork of neurons.

[0192] In a particular implementation, the process further includes a step in which each 810J = 1.. M — 1 subnetwork of neurons is adapted so as to minimize a cost function between the estimated partial data El s e i and the reference data corresponding to the channel matrix El s e .

[0193] In this way, gradient propagation is improved within successive 210J = 1.. M — 1 subnetworks. This particular implementation method also allows observation of the system's performance improvement over iterations, and even the determination of a number of iterations in which the process is implemented, for example, as a function of a complexity / accuracy ratio.

[0194] Figure 10 represents a method for training a machine learning model, according to a fifth implementation example.

[0195] In this example, we consider the case where the machine learning model 210 is an auto-encoder, which consists of two main parts, an encoder 1010 and a decoder 1020.

[0196] The 1010 encoder is configured to take the H matrix as input s e representative of the propagation channel between the transceiver and the communication device. This step is similar to step S100 in Figure 6B. In a particular implementation, this encoder is further configured to take E data as input INIT representative of the static nature of the transceiver's environment. This 1010 encoder is configured to generate a representation corresponding to the estimated environment E s efrom this input or these inputs. To do this, the encoder is composed of several layers of neurons that progressively reduce the size of the input to obtain a representation E s e compact.

[0197] The 1020 decoder, on the other hand, is configured to take as input the representation corresponding to the estimated environment E s e , and to determine the estimated data El s e representative of the propagation channel between the transceiver and the communication device located in the estimated environment E s e In other words, the 1020 decoder is trained to learn how to generate an approximation of the El input. s e based on representation E s e In a particular implementation mode, the decoder is a differentiable ray tracing model.

[0198] As illustrated in Figure 10, this estimated data El s eis then used to adapt the machine learning model 210. This adaptation is, for example, implemented to minimize a costLoss() function between the estimated data H s e and a reference datum corresponding to the matrix H s e previously mentioned. This cost function (Loss) is, for example, a Mean Squared Error (MSE) loss function. Then, a backpropagation step is implemented during which the encoder weights (1010) are adjusted to minimize this cost function (Loss). In other words, the decoder layers are "frozen".

[0199] 11. It is also important to note that at the inference stage, only the trained 1010 encoder is used to determine the estimated environment E s e .

[0200] Figures 11A and 11B represent a method for training a machine learning model, according to a sixth implementation example.

[0201] In this example, we consider the case where the machine learning model 210 is a reinforcement learning model, which consists of two main parts, an agent 1110 and an interaction environment 1120.

[0202] Conventionally, reinforcement learning involves an autonomous agent 1110 learning the best action or sequence of actions to perform through experience. An autonomous agent is understood to be a software agent, that is, a computer program that performs tasks like an automaton. More precisely, the autonomous agent 1110 observes an interaction environment 1120 and decides on an action or sequence of actions to perform on this interaction environment 1120 based on its current state.

[0203] A state corresponds to a representation of the interaction environment 1120 that said agent 1110 receives from this interaction environment 1120, for example, after an action is performed on this interaction environment 1120. In return, the interaction environment 1120 provides agent 1110 with a reward REC, which can be positive if the action performed had a positive impact on the interaction environment 1120, negative, which then corresponds to a penalty, or zero. The interaction environment 1120 can correspond to a computer representation of a real environment, and in this case, The interaction environment 1120 also corresponds to a computer program.

[0204] These steps of determining an action to be carried out, transitioning to a new current state, and receiving a reward by agent 1110 are repeated until a stopping criterion is reached, and allow agent 1110 to learn or determine optimal behavior, also called "strategy" or "policy" n, in the sense that it maximizes a gain corresponding to a sum of the rewards received over time.

[0205] A policy n is a function that maps a given state to actions that can be performed. The reinforcement learning method then allows us to learn the best action or sequence of actions to perform when the interaction environment 1120 is in a given state.

[0206] Since this training process is iterative, Figure 11A illustrates iterations 1.. N — 1, and Figure 11B illustrates iteration N.

[0207] As illustrated in Figure 11A, agent 1110 is configured to take the matrix H as input s e representative of the propagation channel between the transceiver and the communication device. This step is similar to step S100 in Figure 6B. In a particular implementation, this agent 1110 is further configured to take E data as input INIT representative of the static nature of the transceiver environment. During an iteration n, this agent 1110 is configured to generate, as output, a partial estimate E s e n of the environment from this or these inputs.

[0208] During this same iteration n, the interaction environment 1120 is configured to take as input the partial estimate E s e n , and to determine the estimated partial data H se nrepresentative of the propagation channel between the transceiver and the communication device, both located in this partially estimated environment E s e n In a particular implementation mode, the interaction environment 1120 is a ray tracing model [differentiable or not].

[0209] As illustrated in Figure 11A, this estimated partial data H sen is then used so that agent 1110 can revise the strategy he is applying. More precisely, the strategy is adapted to maximize a reward function REC1 that represents a similarity between this estimated partial data H s e n and the reference data corresponding to the matrix H s e Then a backpropagation step or another technique used in reinforcement learning is implemented to update the strategy.

[0210] These steps are iterated until a stopping criterion is reached, which may correspond to a predetermined number of iterations or a convergence criterion. The process illustrated in Figure 11B is implemented when this stopping criterion is reached.

[0211] As illustrated in Figure 11B, the matrix H s e A representative of the propagation channel between the transceiver and the communication device is provided as input to agent 1110, whose strategy is revised (e.g., by applying the process illustrated in Figure 11A). This step is similar to step S100 in Figure 6B. In a particular implementation, this agent 1110 is further configured to take data E as input. INIT representative of the static nature of the transceiver's environment. This agent 1110 is configured to generate a partial estimate E s e N of the environment, based on this or these inputs.

[0212] As illustrated in Figure 11B, this partial estimate E s e N is then used to adapt agent 1110 [or, more precisely, the strategy it implemented in step N]. This adaptation is, for example, implemented to maximize a REC2 reward function representative of a similarity between this partial estimate E s e N and the reference environment E s e Then a backpropagation step or another technique used in reinforcement learning is implemented to update the strategy.

[0213] The use of such an architecture is advantageous because it decomposes the initial problem into sub-problems, and promotes an iterative construction of the matrix E s e N In this way, the system is deployable.

[0214] In a particular implementation mode, the reward assigned by the REC2 reward function represents a similarity between the estimate partial E s e N and the reference environment E s e depends on the number of iterations performed before reaching a convergence criterion. In other words, the lower the number of iterations, the greater the reward.

[0215] Inference

[0216] Figure 12 shows, in the form of a logic diagram, a method for estimating the environment of a transceiver, according to an example implementation. This method is implemented, for example, by an electronic device 300 for estimating the dynamic nature of the environment of a transceiver, such as the device shown in Figures 4 and 5. This estimation electronic device 300 corresponds to the training electronic device 200, or to a separate device.

[0217] In one particular implementation, the estimation electronic device 300 is located within the transceiver or directly connected to it. Alternatively, the electronic device 300 is located within the telecommunications operator's core network. Alternatively, the electronic device 300 is configured to estimate the environment of several transceivers relatively close to each other [e.g., in the case of base stations, two kilometers apart], and a centralized architecture is then adopted. In this particular case, the use of a single digital twin and a single machine learning model is then feasible. Such an architecture offers the advantage of being able to benefit from cross-referencing information between the different transceivers, thus improving the system's fidelity to reality.

[0218] As illustrated in Figure 12, the estimation process includes a first step S200 of obtaining a data point H s e representative of a propagation channel between the transceiver and a communication device, both located in the environment. This step is implemented, for example, by the MOD_OBT module of the 300 electronic estimation device.

[0219] To achieve this, the electronic estimation device 300 receives channel state information, CS1, from the connected communication devices. Some of these communication devices may be anchors, i.e., static receivers, intentionally placed in the environment to improve the sensing capabilities of the electronic estimation device 300. During the training process, anchors identical to those placed in the real environment were incorporated into the simulations to ensure the accuracy of the simulation-derived data used in the training process. All or part of the channel state information can be used for detection.In a specific implementation method, a selection of data representative of communication channels between the transmitter-receiver and various communication devices located in the environment is used, ensuring greater diversity among the communication channels considered. This diversity is assessed, for example, by calculating a correlation value between the different channels used. By prioritizing channels with a certain degree of diversity, the system maximizes the chances of having measurement points at varied locations in the environment, and thus obtaining a digital twin consistent with reality.

[0220] The process further includes a selection of H data se ) representative of communication channels between the transceiver and different communication devices located in the environment, so as to improve a diversity of the communication channels considered.

[0221] The process further includes an S210 estimation, by a learning model trained using the previously mentioned training process and based on the data obtained, of the environment E s e This step is implemented, for example, by the MOD_EST module of the electronic estimation device 300.

[0222] Finally, this estimate is provided as input to a digital twin of the transceiver's environment during an S220 step, thereby improving the connectivity of the communication device with the transceiver. This step is implemented, for example, by the MODJN module of the 300 estimation electronic device.

[0223] Temporal aspect

[0224] In a particular implementation mode, the machine learning model is configured to process a time series of reference data representative of a propagation channel between the transceiver and a communication device.

[0225] Indeed, the transceiver environment is inherently dynamic due to the nature of the entities that compose it. Furthermore, the communication device(s) can move within this environment, resulting in variations in the propagation channel depending on the different positions adopted by these devices. Therefore, this particular implementation method, which takes the temporal dimension into account, has the advantage of improving the quality of the predictions made by the machine learning model.

[0226] Indeed, starting from a reference data point representing a propagation channel between the transmitter and one or more communication devices, there can be several estimated environments whose dynamic entities could give rise to the propagation channel. In particular, estimated environments linked by a symmetry relationship can result in an identical propagation channel. The use of temporal data makes it possible to identify the correct solution among the possible symmetrical solutions, eliminating impossible evolutions among the symmetrical solutions.

[0227] More formally, we consider a time series H s e [t — T],.., H s e [t], which represents a temporal evolution of the propagation channel between the transceiver and the communication device.

[0228] The time gap between times t and t+1 is defined, for example, based on the temporal consistency of the dynamic environment—that is, the speed at which dynamic entities move—and the complexity introduced by the model's consideration of the temporal sequence. It is also important to note that this time gap can vary during deployment, depending on the needs and observations. For instance, if an abnormal situation is detected in the environment, the processing frequency is increased.

[0229] In a particular implementation mode, sets s of samples e that incorporate this temporal dimension are generated by simulation. Each set then provides a coherent evolution of the environment over a period long enough to allow for a relevant training step [e.g., from a few seconds to a few minutes].

[0230] The temporal aspect is considered, for example, as follows for the machine learning models in Figures 6A, 7, and 10. At the current time t, these machine learning models 210 then take as input a time series of matrices H s e [t — T],..., H s e [t], and generate, as output, an estimated environment E s e [t] at this instant t.

[0231] In a particular implementation mode, these machine learning models also take as input a history of estimated environments E s e [t — T],..., E s e [t — 1] between times t — 1 and t — T. During initialization, this history corresponds, for example, to a series of matrices E init representative of the static nature of the environment, which are progressively updated by the values ​​of E s e [t].

[0232] The temporal aspect is for example considered as follows for the machine learning models in figures 8 and 9 which, for the record, correspond to recurrent neural networks comprising, when unfolded, M neural subnetworks.

[0233] In this particular case, rather than providing the same El input s e For each 210J neural subnetwork = 1..M, each 210J neural subnetwork receives as input the matrix H se [t — M + i]. Thus, the 210_l neural subnetwork receives, for example, the matrix H s e [t — M + 1], the 210_2 neuronal subnetwork receives the matrix El s e [t — M + 2],..., and the 210_M neuronal subnetwork receives El s e [t].

[0234] The outputs and / or states of each 210J neural subnetwork are therefore those corresponding to the input time. For example, if the 210J neural subnetwork takes the matrix H as input se[t — M + i], then the outputs / states of this 210J neuronal subnetwork will be E s e i = E s e [t — M + i] and H se i = H se [t — M + i].

[0235] Finally, the temporal aspect is taken into account in the following way for the machine learning model in Figures 11A and 11B, which, for the record, corresponds to a reinforcement learning model.

[0236] The reinforcement learning model will naturally retain its current state E s e [t]. Because it takes time to reach the stopping criterion, the reinforcement learning model performs several actions before updating the channel state information, CSL

[0237] We therefore consider that between times t and t + 1, the model performs A actions, the value of A being defined according to the temporal consistency of the dynamic environment considered. A suitable difference between times t and t + 1 is approximately one second. Such a timescale allows us to capture the dynamic entities of interest for consideration by the uses of the 320 digital twin, such as vehicles and pedestrians. Given the current computing power offered by the hardware used in machine learning, it is realistic to consider a value of A between 10 and 100 actions. An action, which corresponds to an execution of the machine learning model, then covers a period between 10 and 100 milliseconds.Such a scale allows for real-time execution of the machine learning model and therefore real-time input to the digital twin, enabling it to perform its functions of connectivity optimization or event detection, in simpler variations. However, the model's inputs and outputs remain unchanged and are always those of time t.

[0238] Furthermore, the size of the spatial units considered will vary depending on the required accuracy. For example, in the case of an implementation for a base station transceiver in a mobile communication network, the size can vary from a few tens of centimeters on each side where maximum accuracy is required, i.e., at ground level, to several meters on each side where accuracy is not critical, for example, several meters above ground. In the context of a residential gateway, where the transceiver's environment is indoors, the required spatial accuracy could remain constant throughout the room, to within one or more tens of centimeters.

[0239] Other variations

[0240] In a specific implementation, either as a complement to or replacement for communication devices, the transceiver environment includes strategically placed anchors to facilitate the sensing process. These anchors are added both to the digital twin during the training phase and to the transmitter's real-world environment. These anchors offer the advantage of enabling sensing without knowing the position of the communication devices based on the channel state information (CS1). In a particular implementation, these anchors are optimally positioned, for example, to maximize sensing capabilities. These positions are determined, for instance, using the digital twin and offline simulations.

[0241] In a particular implementation method, the neural network model is initially trained on a set of consistent external environments, such as dense urban areas. Subsequently, a fine-tuning operation is performed, during which this generalist model is specialized for a specific environment. For example, it can be adjusted and optimized for a particular street for future deployment.

[0242] In a particular implementation mode, the machine learning model is enriched with additional information, such as:

[0243] - meteorological data which can for example allow the ray tracing method to be dynamically adapted according to the state [dry, wet] of the materials;

[0244] - temporal data [e.g., season, day, time]; and / or

[0245] - Contextual geographic data that provides, for example, clues about the nature of the dynamic elements present in the environment under consideration. For example, a park will likely contain children, parents, strollers, etc.

[0246] It is important to note, however, that the model's use of this data requires that it has been trained to integrate it. Therefore, it is essential to be able to simulate these different parameters when simulating environments.

[0247] Finally, it should also be noted that the invention has so far been described in cases where channel state information is estimated at the communication device level before being sent back to the transceiver, but the invention remains applicable even when the channel state information is estimated by the transceiver itself. Furthermore, the transceiver and the communication device may have different channel state information, and in this case, it may also be possible to exploit the channel state information of the uplink and downlink paths, particularly in the context of frequency-division multiplexing (FDD), given that materials interact differently depending on the frequency.

[0248] Finally, in a specific implementation mode, the environment is equipped with multiple sensors, and the learning model is retrained considering these additional modalities. Indeed, a dynamic entity might not have been correctly detected by the machine learning model, but it could have been detected by a video surveillance camera, which provides data with a high degree of certainty. In this particular case, the machine learning model is informed of the presence, or even the exact position, of this dynamic entity, and a retraining step (also called "online" training) is then implemented.

Claims

Demands

1. A method for training a machine learning model (210) intended to be coupled to a digital twin (320) of a transceiver environment (10), the method being implemented by an electronic device (200) capable of executing the machine learning model (210) and comprising: - Obtain (S100) a data point H se ), called "reference data", representative of a propagation channel between the transceiver and a communication device, both located in an environment called the "reference" environment; - Estimate (S110) by running the machine learning model (210) a transceiver environment E se ), said estimated environment, the learning model taking as input the reference data (ff s e ), and the estimated environment E se) including data representing the presence of a dynamic entity in the environment; - Adapt (S120) the machine learning model according to the estimated environment

2. A training method according to claim 1, the electronic device (200) being capable of executing a second machine learning model (710), the method further comprising: - a determination, by the electronic device (200) by executing the second machine learning model (710) and using as input the estimated environment E se ), of a given H se ), referred to as "estimated data" representing the propagation channel between the transceiver and the communication device located in the estimated environment; and, - an adaptation of the second machine learning model (710), so as to minimize a cost function between the estimated data (ffs e ) and the reference data (ff s e ).

3. A training method according to claim 1 further comprising: - a determination, by the electronic device [200] by executing a differentiable ray plotting function and using as input the estimated environment E se ), of a given H se ), called "estimated data" representative of the propagation channel between the transceiver and the communication device located in the estimated environment; - an adaptation of the first machine learning model (210) so as to minimize a cost function between the estimated data H se ) and the reference data

4. A training method according to any one of claims 1 to 3, wherein the first machine learning model (210) is a recurrent neural network.

5. A training method according to claim 4, wherein the recurrent neural network comprises, when unfolded, a plurality i = 1.. M of interconnected neural subnetworks having the same structure, the method further comprising the following steps: - obtaining, by each subnetwork i = 2.. M of neurons, the reference data H s e and a partial estimate of the environment generated by the subnet i — 1 immediately preceding subnet i, and - a determination, by each subnetwork i = 2M neurons, of a partial estimate E s e i ) of the environment from the reference data (ff s e ) and the partial estimation generated by subnetwork i — 1, and the partial estimate E s e MThe environment determined by the Mth subnetwork corresponds to the estimated environment E se ) considered during the adaptation of the first machine learning model.

6. Training method according to claim 5, further comprising, for each subnetwork i = 1.. M of neurons, an adaptation of the subnetwork i = 1.. M of neurons, so as to minimize a cost function between the partial estimate E s e i ) and the reference environment

7. A training method according to claim 1, wherein the machine learning model (210) is an auto-encoder, the estimation step corresponding to an encoding of the reference data (W s e ), so as to obtain an encoded representation corresponding to the estimated environment E se ), the process further comprising: - an estimated environmental decoding E se), so as to obtain a data point H s e , referred to as "estimated data", representative of the propagation channel between the transceiver and the communication device located in the estimated environment; and the adaptation step corresponds to an adaptation of the encoder (1020), so as to minimize a cost function between the estimated data (ff s e ) and the reference data (ff s e ).

8. A training method according to claim 1, wherein the machine learning model (210) is a reinforcement learning model comprising an agent (1110) and an interaction environment (1120), and the method further comprises the following steps, implemented iteratively until a stopping criterion is reached: - a partial estimate, by agent (1110) and based on the reference data, - a determination, by the interaction environment (1120) and from the partial estimation data said to be "estimated partial data", representative of the propagation channel between the transceiver and the communication device located in the partially estimated environment; and, - a revision of a strategy applied by the agent, so as to maximize a reward function representing a similarity between the estimated partial data and the reference data H se ) and when the stopping criterion (IV) is met, the obtaining and estimation steps are implemented by the agent by applying the revised strategy, and the adaptation step is an adaptation by the agent, implemented so as to maximize a reward function representative of a similarity between the estimated environment and the reference environment

9. A training method according to any one of claims 1 to 8, wherein, at time t, the reference data is a time series H s , e [t which represents a temporal evolution of the propagation channel between the transceiver and the communication device located in the reference environment.

10. A method for estimating the environment of a transceiver [10] implemented by an electronic device (300) capable of executing a machine learning model and comprising: - obtaining (S200) a data point H se ) representative of a propagation channel between the transceiver and a communication device, both located in the environment; and, - an estimation (S210), by running the machine learning model and using the obtained data as input, of an environment, called the estimated environment ; - a supply (S220) of the estimated environment as input to a digital twin (320) of the environment of the transceiver (10).

11. Computer program comprising instructions for carrying out a training method according to any one of claims 1 to 9 or an estimation method according to claim 10, when said program is executed by a processor.

12. Computer-readable recording medium on which a computer program according to claim 11 is recorded.

13. Electronic device [200, 300] capable of executing a machine learning model, configured to implement a training method according to any one of claims 1 to 9 and / or an estimation method according to claim 10. I