Estimation of statistical parameters of a Rice channel in the presence of masking based on deep learning
The method addresses the challenge of estimating the Rice K factor in wireless networks with shadowing by employing a deep learning-based neural network architecture that accounts for permutation invariance and additional statistical information, achieving efficient and adaptable estimation.
Patent Information
- Application Number
- FR2022011185
- Authority / Receiving Office
- FR · FR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2022-10-27
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2042-10-27
AI Technical Summary
Existing methods for estimating the Rice K factor in wireless communication networks face challenges in the presence of shadowing, particularly due to high estimation variance when few channel samples are available and the complexity of conventional statistical methods like EM, which are difficult to implement in embedded systems, and deep learning approaches lack adaptability and granularity.
A method using deep learning with a tailored neural network architecture that accounts for invariance by permutation and includes additional statistical information, employing a two-step training process to estimate the Rice K factor in the presence of shadowing, reducing complexity and improving convergence and generalization.
The method provides a simple and practical estimation of Rice K factor with improved convergence and generalization capacity, suitable for embedded systems, by using a neural network architecture adapted to the specificities of the estimation problem and incorporating additional statistical information.
Smart Images

Figure 00000016_0000 
Figure 00000017_0000
Abstract
Description
Title of the invention: Estimation of statistical parameters of a Rice channel in the presence of masking based on deep learning. Technical field
[0001] The present invention relates to the field of wireless communications. More specifically, the invention relates to the estimation of statistical parameters of a propagation channel in order to perform a statistical estimation of a propagation channel. Prior art
[0002] For some wireless communication networks, there is no fast feedback channel that allows for a reliable estimation of the instantaneous impulse response of a propagation channel at a transmitting node. In this context, an optimal resource allocation adapted to the propagation conditions is generally based on the statistics of the propagation channel.
[0003] A frequently used statistical model of a propagation channel is the Rice channel. Conventional modeling of a Rice channel consists of writing each coefficient of the propagation channel's impulse response as the sum of a deterministic term, the Line of Sight (hereafter referred to as LdV), and a random term following a centered normal complex distribution (Rayleigh component) corresponding to the non-line-of-sight component, known as non-LdV. Such a propagation channel is notably illustrated in the paper "A new simple model for land mobile satellite channels: first- and second-order statistics," by A. Abdi, WC Lau, M.-. Alouini and M. Kaveh, published in IEEE Transactions on Wireless Communications, vol. 2, no. 3, pp. 519-528, May 2003. This statistical model accurately represents the distribution of a propagation channel amplitude when a direct Line of Sight exists between the transmitting node and a receiving node.In the model presented by Abdi et al., the line-of-sight (LOS) component is no longer deterministic but is itself subject to random fluctuations following a Nakagami-m distribution. An important indicator of the quality of the communication link is provided by a factor known as Rice's K factor. This K factor corresponds to a power ratio between the LOS component and the non-line-of-sight component between the transmitting and receiving nodes. This statistical propagation channel model encompasses the Rayleigh channel when K = 0 and the AWGN channel (for "Additive White Gaussian Noise") when K = 0, as special cases. Thus, the higher the K factor, the better the link between the transmitting and receiving nodes.
[0004] Knowledge of the K factor and the attenuation of the received power relative to the transmitted power related to the distance between the transmitting node and the node The receiver (called "pathloss" in English) allows for resource allocation adapted to propagation conditions. Estimating the K factor can become more complex in the presence of shadowing, which is caused by obstacles between the transmitting and receiving nodes. These obstacles generate random fluctuations in the LdV component value. This phenomenon is known as "shadowing." The paper by Abdi et al. describes, for example, a Rice channel with Nakagami-W type shadowing.
[0005] There exist in the literature many procedures for estimating the Rice K factor in the absence of “shadowing”. A relatively limited number of papers deal with the estimation of the K factor in the presence of “shadowing”. Among these documents, patent CN106850109B, the document “Estimation of Rician K-Factor in the Presence of Nakagamim Shadowing for the LoS Component by G. Giunta, C. Hao and D. Orlando in IEEE Wireless Communications Letters, vol. 7, no. 4, pp. 550-553, Aug. 2018 » this document Giunta et al. and the document « Estimation of the Ricean $K$ Factor in the Presence of Shadowing, by X. Leturc, P. Ciblât and CJ Le Martret in IEEE Communications Letters, vol. 24, no. 1, pp. 108-112, Jan. 2020,” says the paper Leturc et al. These various papers seek to improve the estimation of the Rice K factor of the spreading channel model described in the paper Abdi et al.These documents thus describe conventional statistical methods for estimating Rice's K factor in the presence of "shadowing", such as the method of moments in patent CN106850109B and in the Giunta et al. document or the EM method "for Expectation Maximization" in the Leturc et al. document.
[0006] The main disadvantage of the method of moments is that when the number of channel samples available to perform the estimation is small, the variance of the method's estimation is high, which is detrimental when using the estimation of the K factor within the framework of an appropriate resource allocation.
[0007] The performance of the EM method is better than that of the method of moments, but this method is difficult to implement in an embedded system because it is an iterative method whose convergence is not controlled and is potentially lengthy (several hundred iterations). Furthermore, in this EM method, each iteration involves special mathematical functions, for example, a cylindrical parabolic function, for which an embedded implementation is expensive. In addition, as shown in the Leturc et al. paper, the maximum likelihood estimator on the propagation channel model of the Abdi et al. paper is not implementable with reasonable complexity due to the mathematical complexity of this channel model.
[0008] Patent EP3503649 describes a method and device for calculating parameters Propagation channel statistics. This process involves several steps executed by a processor within a communication link in the propagation channel. Thus, from a training sequence contained in the communication signal received by a receiving node, the communication link processor estimates the propagation channel and then its statistics. This information is transmitted to a resource allocator to perform resource allocation. In a final step, the resource manager allocates resources for the next transmission. These allocated resources include, for example, power, modulation scheme, and coding.Optimized resource allocation across the various nodes of the communication network maximizes network performance dynamically, allowing the network to continuously adapt to propagation conditions that change with node mobility. This optimized resource allocation is particularly useful for nodes in embedded systems where propagation conditions are constantly evolving.
[0009] Patent EP3503649 describes the estimation of a Rice K factor solely in the absence of shadowing. It also describes the tracking of the K factor based on tools derived from break detection.
[0010] The document “Rician K-Factor Estimation Using Deep Leaming by M. Alymani, MH Alhazmi, A. Almarhabi, H. Alhazmi, A. Samarkandi and Y.-D. Yao in the 2020 29th Wireless and Optical Communications Conference (WOCC)” says Alymani et al., the document “CNN Based Rician K factor estimation for non-stationary industrial fading channel by G. Lu, Q. Zhang, X. Zhang, F. Shen and F. Qin in the 2018 IEEE Global Conference on Signal and Information Processing (GlobalSIP)” says Lu et al. and the document “Rician K-factor Estimation based on Channel Quality Indicator in OFDM Systems using Neural Network by Kun Wang published in 2018 and available on arxiv” says Wang, deal with the estimation of the Rice K factor using tools from “deep leaming”. However, these publications have limitations in their approach.
[0011] First, these documents are presented in a context without shadowing. Second, all these documents use classification tools based on Rice's K factor, where the estimation granularity is fixed by a grid of allowed values for classification. For example, in Wang's document, the classes considered range from 1 to 11 in steps of 1, but it can be useful to be able to estimate intermediate K factor values, since the performance between K = 1, K = 1.5, and K = 2 differs. Furthermore, the three documents describe conventional neural network architectures from the literature without attempting to adapt them to the specific characteristics of the estimation problem under consideration. In addition, these approaches They all assume that the number of input samples to a neural network architecture is identical from one scenario to another, which is not necessarily the case in practice. Finally, Wang's paper describes a "fully connected perceptron" architecture that is relatively simple, but this architecture can be considered inefficient in some cases because it does not account for permutation invariance of the input data.
[0012] There is therefore a need to propose a method for estimating statistics of a propagation channel in a communication network in the presence of "shadowing", based on the estimation of the K factor of the Rice channel and this from tools from "deep leaming" whose complexity is controlled. Description of the invention
[0013] The present invention aims to remedy at least partially this need.
[0014] More particularly, the present invention aims to reduce the complexity techniques based on conventional statistical methods.
[0015] This invention thus covers a method for estimating statistics of a propagation channel, said propagation channel enabling the transmission of a communication signal between a transmitting node and a receiving node, said propagation channel being a Rice channel in the presence of masking, said communication signal comprising a learning sequence, said estimation method comprising: - a step of determining a set of N estimates of the propagation channel from said learning sequence transmitted N times by the transmitting node; - a pre-processing step of said set of estimates to obtain intermediate data; - a step for processing intermediate data to obtain processed intermediate data, said processing step being capable of taking into account a permutation invariance of said intermediate data, for any N; - a step of determining statistical parameters from the intermediate data processed; said intermediate data processing step and said statistical parameter determination step being carried out by an architecture of at least one neural network.
[0016] The invention thus provides a method for estimating the statistics of a propagation channel that is simple and practical compared to techniques based on conventional statistical methods. This estimation method is based on the estimation of statistical parameters, such as Rice's K factor, in the presence of particularly "shadowing". Furthermore, unlike the prior art, this invention proposes to perform regression rather than enhanced classification. Finally, the invention incorporates permutation invariance of the estimation problem considered for the design of the architecture of at least one neural network. This drastically accelerates convergence during the training of the neural network(s) of the architecture and improves the generalization capacity of this architecture.
[0017] In a particular embodiment, the intermediate data processing step comprises: - a processing step by a first neural network having one input and 5 outputs, said first neural network being able to receive as input intermediate data obtained from a real part and an imaginary part of each estimation of the propagation channel and to provide as output vectors of S inputs; - a summation step of the 2N input vectors of S, said summation being commutative and applicable for all N.
[0018] In a particular embodiment, the pre-processing is a normalization of the set of N estimates.
[0019] In a particular embodiment, the estimation process includes a step of determining statistical information from the set of N estimates, said statistical information being intended to be transmitted to the architecture.
[0020] The invention thus allows the insertion of additional statistical information as input to the architecture.
[0021] In a particular embodiment, the statistical information determination step includes an estimation of a K factor of the propagation channel.
[0022] In a particular embodiment, the estimation process includes a step of determining a variance of the estimation noise, said variance of the estimation noise being intended to be transmitted to the architecture, said variance of the noise being associated with the number N of estimations or with the square root of said number N.
[0023] In a particular embodiment, the architecture comprises four neural networks.
[0024] In a particular embodiment, the architecture is pre-trained during a training step, said training step comprising a first pre-conditioning step on synthesis data and a second adjustment step on data from at least one measurement campaign.
[0025] The objective of the first step is to precondition the architecture, and more specifically the first neural network, in order to start training on real data with weights as close as possible to the optimal weights, with a view to having a limited penalty due to the small size of the databases from Measurement campaigns. The second step allows for an adjustment of these closest weights. This two-step training procedure optimizes the estimation of propagation channel statistics.
[0026] In a particular embodiment, the propagation channel is a frequency-flat channel.
[0027] In a particular embodiment, the propagation channel is a frequency-selective channel.
[0028] Another object of the invention relates to a statistical estimation device in a communication network for the implementation of a statistical estimation method for a propagation channel according to an object of the invention, said propagation channel allowing the transmission of a communication signal between a transmitting node and a receiving node, said propagation channel being a Rice channel type channel in the presence of masking.
[0029] The estimation device thus proposed is well suited to the problem addressed, unlike prior art approaches that use conventional architectures without adaptations. The device of the invention is also well suited to estimating the statistical parameters of the propagation channel, with the possibility of including additional statistical information beyond just the channel samples.
[0030] The present invention will be better understood upon reading the detailed description of embodiments taken by way of non-limiting examples and illustrated by the accompanying drawings, in which:
[0031] [Fig-1] [Fig. 1] illustrates a communication network between a sending node and a receiving node with a resource allocator;
[0032] [Fig.2] [Fig.2] illustrates the different stages of a method for estimating statistics of a propagation channel according to the invention;
[0033] [Fig.3] [Fig.3] represents a device for estimating statistics for the implementation of the process of [Fig.2].
[0034] The invention is not limited to the embodiments and variants shown and other embodiments and variants will be obvious to a person skilled in the art.
[0035] In the description that follows, we adopt the norm in which vectors are referenced in bold with respect to scalars.
[0036] Fig. 1 illustrates a communication network 1 with a resource allocator.
[0037] This communication network 1 comprises a transmitting node NE, a receiving node Nr, and an NGr terminal having a resource manager function.
[0038] The transmitting node NE and the receiving node NR communicate via a Rice channel propagation channel 10. The receiving node Nr communicates with the terminal NGr via another propagation channel 20.
[0039] The transmitter node NE is adapted to transmit N learning sequences SA to The receiving node NR is capable of estimating the propagation channel 10. These statistics are then transmitted to the terminal NGR so that it can determine a resource allocation. Resource allocation consists, for example, of determining a transmission power given the knowledge of the propagation channel 10 statistics. In another embodiment, the receiving node NR and the terminal NGR are the same entity.
[0040] The method for estimating the statistics of the propagation channel 10 is described in [Fig. 2]. It comprises: - a step El of determining a set of N estimates (hh . . .,h„ ... hv ); - a pre-processing step E2 of the set of N estimates ( h15 ,,hz, ... hv); - an E3 step of intermediate data processing; - a step E4 for determining statistical information; - a step E5 for determining a variance of the estimation noise; - a step E6 for determining statistical parameters K, m;
[0041] Step El is adapted to determine the set of N estimates ..., h;, ... hv of the propagation channel from the learning sequences.
[0042] In the case of a frequency-flat channel, the impulse response of the propagation channel consists at each instant i of a single random complex coefficient h; which can be written according to this model:
[0043] hi = + R:
[0044] where corresponds to an LdV part and R, corresponds to an nLdV part. More precisely, a corresponds to the amplitude of the LdV component, is a random variable that follows a Nakagami-OT distribution and represents the "shadowing" of the LdV component, is the phase of the LdV component, and ~ CN(0,2 <t£) où CN(0,2<t^) correspond à la loi normale centrée et de variance 2<7g. La densité de probabilité de c< peut s’écrire comme suit :
[0045] f (x) = “-x2"'-Vx> 0.
[0046] It is assumed that the "shadowing" and the nLdV component vary independently from one measurement instant to another, i.e. V i jct and ci are independent and identically distributed, just like R, and Rj. However, the approach proposed in the present invention can be extended to the case of block "shadowing", i.e., constant "shadowing" over Ts consecutive samples.
[0047] It should be noted that the channel model considered encompasses the Rice channel without "Shadowing" as a special case when m +
[0048] It is assumed that the propagation channel is estimated, for example via the use of the The learning sequence is also called the pilot sequence. Thus, an estimated propagation channel coefficient can be written as follows:
[0049] h^hi + b^
[0050] where ~ CN(0,2 <t^) représente un bruit d’estimation blanc complexe et gaussien, dont la variance peut être connue (par exemple via processus de calibration du matériel radio) ou inconnue.
[0051] It should be noted that the analytical expression of the likelihood of is provided in the prior art and involves parabolic cylindrical functions, making the maximum likelihood estimator too complex to implement for use in embedded systems.
[0052] This model can be extended as follows to the case of a frequency-selective channel. The impulse response of a frequency-selective channel can be written as h — , where L represents the number of paths in the propagation channel. At each instant i, each input of the vector h can be written
[0053] ^.= ^4 + R*.
[0054] where the LdV and nLdV components are independent between the different paths.
[0055] Also, the estimation of the channel impulse response is noted as follows
[0056] In the context of the present invention, it is assumed that N estimates of the propagation channel are available for estimating Rice's K factor. To adopt a common notation for the case of a flat channel and a frequency-selective channel, the set of available channel estimates will be denoted as h := [hb ..., h;, .... h;y], where h = hi in the case of a frequency-flat channel, and h, - h; in the case of a frequency-selective channel. The aim here is to estimate the different values of K for each path, denoted g, and potentially any other statistical parameter of the channel. such as the different values of the parameter m of the "shadowing" denoted niQ, ..., ïhL, from h.
[0057] It should be noted that the case of a frequency-selective channel can also be treated by considering each of the L paths as a frequency-flat channel.
[0058] In the following description, we are interested in the case of a frequency flat channel.
[0059] The estimation process also includes a step E3 of processing intermediate data obtained from the set of N estimates h3, ... , h ... h. The intermediate data datai come from a pre-processing E2 of this set of N estimates hb .... h.. ... h^.
[0060] The E3 step of processing the intermediate data datai is capable of taking into account counts a permutation invariance of the set of N estimates hb . ..,h(, ... h^.
[0061] Indeed, the estimation problem we seek to address is invariant under permutation, which means that the estimation of the values of K for each path is the same regardless of the permutation of the elements of h. Mathematically, this property can be written as °ù represents the estimation procedure, and ÿ(h) is a permutation of the elements of h which can be written 0 (h) — [h^_, .... ] °ù * 0 .V i * j, and 1 e
[0062] It should be noted that the pre-processing in step E2 which provides the intermediate data datai in the example of [Fig.2] is different from the processing applied to take into account the invariance under permutation in step E3.
[0063] Preferably, this preprocessing step E2 is a normalization of the set of jV estimates hj, ---,11,-, 1¼.
[0064] Step E6 is adapted to determine statistical parameters K, m from the processed intermediate data data2. This processed intermediate data data2 is obtained here as the output of step E3, which processes the intermediate data datai. The architecture A comprises several neural networks as illustrated in [Fig. 3].
[0065] The data3 data are derived from a step E4 of statistical information determination. This step E4 allows a rough estimation of a factor K from the set of A'estimates hj, ...h;. ... h,v. This estimation of the factor K will guide the architecture A in its searches.
[0066] The data4 data represents the variance of the estimation noise determined during step E5. This step E5 corresponds to a calibration of the radio equipment performed beforehand. The data4 data may also contain the number N of channel estimates. Alternatively, the data4 data may contain the square root of the number N of channel estimates. Transmitting such a number or the square root of this number improves the performance of the neural network architecture A.
[0067] The statistical estimation method of Figure 2 will be detailed in support of the description of a statistical estimation device 30 illustrated in Figure 3. It should be recalled that we are considering here a frequency-flat channel, i.e. V
[0068] The statistical estimation device 30 comprises: - a module 301 for determining the set of N estimates hh ..., h„ ... hv; - a pre-processing module 302 for said set of N estimates hb 11,, ... hv ; - a first neural network NN1; - a 303 module for storing data'l data from the first neural network NN1; - a module 304 for summarizing data'l from the first neural network NN 1; - a module 305 for determining statistical information data3; - a second neural network NN2; - a module 306 for determining the variance of the data4 estimation noise; - a third neural network NN3; - a fourth neural network NN4.
[0069] It should be noted from the outset that the neural networks NN1, NN2, NN3, NN4 are here deep neural networks.
[0070] Module 301 is adapted to provide the set of N estimates h; ... h, ... hv. We denote hy? and hy respectively the real part and the imaginary part of each h,.
[0071] Two branches originate from this module 301, an upper branch and a central branch.
[0072] The upper branch corresponds to a calculation of statistical information data3 by module 305 from the input data h], .... h;. ... h,v. We thus calculate: - the values and which correspond respectively to the empirical means of h;. / ? and hy ; - an estimate £ of the parameter K. This estimate can be obtained via the method of moments. These calculations correspond to step E4 of the determination of statistical information in [Fig.2].
[0073] We thus have three statistical pieces of information, which are provided as input to the second neural network NN2. This neural network NN2 comprises S2 outputs and allows the data data5 to be provided to the fourth neural network NN4.
[0074] In the middle branch, each h / . / ? and hy is provided to the preprocessing module 302. This module 302 is adapted to implement the preprocessing step E2 of Figure 2 and thus provide the data datai as output. This data datai feeds into the input of the first neural network NN1 to generate 2N input vectors of Si forming the data data'l. In other words, the first neural network NN1 is applied in parallel to the hy? and hy of each h. This allows for a number of inputs that can vary from one scenario to another. The data data'l are then stored in the storage module 303. This data data'l are then summed by the summation module 304 during a summation step of the 2N vectors of 5^ inputs, said summation being commutative and applicable for all N.
[0075] In a lower branch, the variance of the estimation noise of the data4 is determined by module 305 during step E5. This data4 is given as input to the third neural network NN3. This data4 also contains the number N of channel estimates or the square root of this number N. The third neural network then comprises two inputs, one for the variance of the estimation noise and one for the number N or its square root, and 53 outputs to constitute the dataô.
[0076] The outputs of module 304, the S2 outputs of the second neural network NN2 and the outputs of the third neural network NN3 are concatenated, resulting in a vector of size $1 + S2 + $3. Finally, this vector is passed as input to the fourth neural network NN4 to provide the statistical parameters K. m.
[0077] The architecture A of the neural networks NN1, NN2, NN3, NN4 is previously trained during a training step which allows this architecture to be better conditioned.
[0078] For the first neural network NN1 associated with the 304 summation module, the permutation invariance of the estimation problem is taken into account in the design of these elements in order to drastically accelerate the training procedure and improve the generalization capacity of the algorithm. Indeed, taking permutation invariance into account avoids having to present the first neural network NN1 with each permutation of each element of the training set.
[0079] The number N of inputs of h available to estimate the parameter K of the Rice channel can vary from one scenario to another (for example depending on the amount of resources allocated to the transmitter which will condition the number of channel estimates available), and thus it is desirable to have an approach allowing to perform a single learning taking into account different values of N. This avoids performing a different learning for each possible value of N in the first neural network, which would make the training procedure more cumbersome, and would require having to include as many first neural networks as there are values of N envisaged.
[0080] The training step of the first neural network NN1 includes a first pre-conditioning step on synthetic data and a second adjustment step on data from at least one measurement campaign.
[0081] Learning on synthetic data has two advantages, namely that it is possible to generate very large databases, and that the data labels are perfectly known.
[0082] Learning from data from at least one measurement campaign has the advantage of limiting the risk of learning simulation artifacts that are not representative of real communication. Conventional estimators from the literature can be used to obtain the labels. Alternatively, real communications can be performed on the propagation channel, and their performance measured in terms of packet error rate or bit error rate, in order to deduce the true value of the K factor by mapping a theoretical error curve to a measured error curve.
[0083] The invention thus provides the following advantages: - the use of "deep learning" for estimating the K factor of the Rice channel in the presence of "shadowing"; - the selection of a neural network architecture well suited to the problem being addressed, based on the identification of specific properties and constraints of that problem; - the insertion of additional statistical information as input to the neural network architecture; - a two-stage training procedure in the context of estimating propagation channel statistics.
[0084] The solution proposed by the invention can be adapted to any other propagation channel model for which the use of conventional estimation tools is too complex and / or whose performance is insufficient.
[0085] The method for estimating propagation channel statistics can advantageously be implemented in a basic transmission station of a wireless communication network.
Claims
Claims
1. Method for estimating statistics of a propagation channel (10), said propagation channel (10) allowing transmission of a communication signal between a transmitting node (NE) and a receiving node (NR), said propagation channel (10) being a Rice channel in the presence of masking, said communication signal comprising a training sequence (SA), said estimation method comprising: - a step (El) of determining a set of N estimates ( h1? ....h,-, ... hjV) of the propagation channel (10) from said training sequence (SA) transmitted N times by the transmitting node (Ne); - a step (E2) of pre-processing said set of N estimates ( h), ..., h,. ...h,vj to obtain intermediate data (datai); - a step (E3) of processing the intermediate data (datai) to obtain processed intermediate data (data2), said processing step (E3) being capable of taking into account an invariance by permutation of said intermediate data (datai) whatever N; - a step (E6) of determining statistical parameters (K, m) from the processed intermediate data (data2); said step (E3) of processing the intermediate data (datai) and said step (E6) of determining the statistical parameters (K, m) being carried out by an architecture (A) of at least one neural network (NN1, NN2, NN3, NN4).
2. Estimation method according to claim 1, wherein the step (E3) of processing the intermediate data (datai) comprises: - a step of processing by a first neural network (NN1) having an input and 5'i outputs, said first neural network (NN1) being capable of receiving as input intermediate data obtained from a real part (h / ^) and an imaginary part (Ky) of each estimation (h,) of the propagation channel (10) and of providing as output 2N vectors of inputs (data'l); - a step of summing the 2N vectors of Sj inputs (data'l), said summation being commutative and applicable for all N.
3. Estimation method according to one of claims 1 or 2, in which the pre-processing (E2) is a normalization of the set of N estimates (hh .... h,. ... hjy).
4. Estimation method according to any one of claims 1 to 3, wherein said estimation method comprises a step (E4) of determining statistical information (data3) from the set of jV estimations (h1? ..., hz, ... h^), said statistical information (data3) being intended to be transmitted to the neural network architecture (A).
5. Estimation method according to claim 4, wherein the step (E4) of determining statistical information comprises an estimation of a factor K of the propagation channel (10).
6. Estimation method according to any one of claims 4 or 5, wherein said resource allocation method comprises a step (E5) of determining a variance of the estimation noise (data4), said variance of the estimation noise being intended to be transmitted to the architecture (A), said variance of the noise being associated with the number N of estimations or with the square root of said number N.
7. Estimation method according to any one of claims 1 to 6, wherein the architecture (A) comprises four neural networks (NN1, NN2, NN3, NN4).
8. Estimation method according to any one of claims 1 to 7, in which the architecture (A) is previously trained during a training step, said training step comprising a first pre-conditioning step on synthetic data and a second adjustment step on data from at least one measurement campaign.
9. An estimation method according to any one of claims 1 to 8, wherein the propagation channel (10) is a frequency-flat channel.
10. An estimation method according to any one of claims 1 to 8, wherein the propagation channel (10) is a frequency selective channel.
11. Device for estimating statistics in a communication network (1) for implementing a method for estimating statistics of a propagation channel (10) according to any one of claims 1 to 10, said propagation channel (10) allowing transmission of a communication signal between a transmitter node (NE) and a receiver node (NR), said propagation channel (10) being a Rice channel type channel in the presence of masking.
