Method for improving the determination of the estimated value of a telecommunications channel, associated receiving device

Abstract

The invention relates to a method for improving, in a receiving device (10), the determination of the estimated value of a telecommunications channel based on at least one signal frame comprising predefined pilot sequences, comprising the steps of: providing, as input to a neural network of the receiving device, initial input data based on the correlation between the received frame and said predefined pilot sequences, and obtaining, at the output of said network, an estimate of a binary mask, determined as a function of said initial input data provided, with: = 0, if the path with index k is determined to contain only noise; = 1, otherwise; determination of second estimates of the channel, according to the following rule: Figure for the abstract: Fig. 4

Description

Title of the invention: Method for improving the determination of the estimated value of a telecommunications channel, associated receiving device technical field

[0001] The invention is in the field of telecommunications and relates more specifically to the processing carried out in a receiving device. Previous technique

[0002] Information transfer from a source (the telecommunications transmitter) to a destination (the telecommunications receiver) involves the propagation of the signal through a telecommunications channel, which can be, for example, a radio channel, a wired channel (e.g., a coaxial cable), etc. (see [Fig. 1]). Some propagation methods generate inter-symbol interference (ISI) on the received signal. In other words, the received signal, sampled at a given instant, after compensation for propagation and processing delays and having suitable synchronization, does not contain only the sent symbol (possibly with amplitude and phase disturbances) along with noise, but rather a mixture (linear combination) of sent symbols.

[0003] In such cases, it is common practice to use processing methods, for example equalization and detection algorithms, to reduce the harmful effect of inter-symbol interference. These methods require, beforehand, an estimate of the channel, which is used when calculating filters and other parameters related to the processing performed. The closer this channel estimate is to the actual channel impulse response (CIR) that occurs during transmission, the more efficient the receiving algorithms will be. Thus, channel estimation algorithms play a crucial role in communication system receivers.

[0004] The usual operating principle of these channel estimation algorithms involves exploiting the transmission of so-called "pilot" sequences, also known as "training sequences" (TS), "reference sequences," or simply "references," which accompany the data during transmission and are, by design, perfectly known to the receiver, as illustrated in Figure 3. The frame is divided into several sections comprising data symbols (called data blocks or resource blocks, among others) and TS sequences that can be inserted before and / or after these sections. The number of TS sequences in a frame being equal to N^g. Data and driver symbols can be derived, for example, from real or complex constellations such as quadrature amplitude modulation (QAM), phase-shift keying (PSK), amplitude-phase-shift keying (APSK), or pulse amplitude modulation (PAM). These techniques are also applicable to periodically rotated constellations such as ü / MM-PSK, where M is the constellation order and n / M radians is the periodic rotation angle.Without loss of generality, it is also possible to obtain these symbols with quasi-rectilinear constellations such as Minimum Shift Keying (MSK) or minimum Gaussian keying (GMSK), Offset Quadrature Amplitude Modulation (OQAM), or with memory modulations such as Continuous Phase Modulation (CPM). The drivers can also belong to families of sequences specifically designed for synchronization and estimation or spectrum spreading tasks, such as Gold, Hadamard, or Golay sequences, or Constant Amplitude Zero Autocorrelation (CAZAC) sequences, or any other family of sequences used in communication systems (see J.M. Velázquez-Gutierrez and C.(Vargas-Rosales, "Sequence Sets in Wireless Communication Systems: A Survey," in IEEE Communications Surveys & Tutorials, vol. 19, no. 2, pp. 1225-1248, 2nd quarter 2017). Once the data and driver symbols are obtained, the signal to be transmitted is constructed using various modulation techniques such as orthogonal frequency division multiplexing (OFDM), or single carrier (SC) modulation, as illustrated in [Fig. 3]. Multiple-access generalizations of these modulations can also be considered, such as orthogonal frequency division multiple access (OFDMA) or single carrier frequency division multiple access (SC-FDMA), or spectrally shaped generalizations.

[0005] Channel estimation algorithms exploit prior knowledge of the pilot sequences present in these emitted signals to obtain an estimate of the CIR, as well as the signal-to-noise ratio (SINR) of the channel. Linear and non-linear channel estimators exist. Linear estimators include, notably, the least squares (LS) algorithm and the linear minimum mean square error (LMMSE) algorithm. These linear estimators can be implemented in the time or frequency domain, for example, to reduce the complexity of these algorithms.

[0006] With further reference to Figure 3, the impulse response of the channel on each of the TS, assumed to be spread over a maximum of L symbols, is written as follows:

[0007] h[n] = ...; hL[n]], n = 1, ...,NTS.

[0008] Suppose that the transmission channel Ht] (replacing the index of the TS sequences n with the instant of continuous time t in this paragraph) considered is a stationary random process characterized by: • its time average: E[A[ t] ] — with E[ • ] the expectation function, • its autocorrelation matrix: E[h[ t]hH[ t] ] with ( • )H denoting the Hermitian transpose of a vector or matrix, • and its time autocorrelation functions for each of its paths 2=1.....with E[h^[t]hj[t + At] ] ±phl(At).

[0009] The first two statistical characteristics correspond to what is also called the power-delay profile (PDP), and the last characteristic corresponds to the dynamic properties generated by the carrier frequency offset (CFO) of the equipment and by the mobility of the equipment, characterized by Doppler spread or shift. In the remainder of this document, "channel profile" refers to the set of three statistical characteristics for the channel under consideration, while "PDP" or "PDP profile" refers to the first two characteristics.

[0010] Suppose we have a linear observation model of each TS, (which is valid in the time or frequency domain: only the form of the matrices and vectors changes):

[0011] y[n] = S[n]h[n] + wf n],

[0012] where y[n] is the observed received signal corresponding to the jth TS, S[u] is the matrix of reference symbols for the jth TS, known to the receiver, h[n] = [hjn]; ...;hk[n]; hKrs[n]] is the column vector corresponding to the impulse response of the channel to be estimated, but extended to the length of the TS sequence, KTS, with added zeros (by design we work with K^s > L), and wfn] the independent additive Gaussian noise and identically distributed (iid) with variance. Note that an extension of this model to the multi-antenna case in reception (i.e., "single-input multiple-output" - SIMO) gives rise to an impulse response consisting of spatial column vectors. hk[n] = [hkl[n]; ...;hk^R[n]] with NR the number of receiving antennas. In this case, the observation model concerns vectors of size KTSNR, with A[n] = [hjn]; hk[n]; ...;hKTS[n] ] and the noise will then be characterized by the spatial covariance matrix Rw of size NrX. In the single-antenna case (to keep the description simple), the kth component of the vector h[n]> therefore corresponds to the kth path of the impulse response of the teletransmission channel that affects the nth TS of the frame.

[0013] The path h^

[11] has a delay of k symbols if the observation model used is symbol-time; otherwise, the delay is k samples if the observation model used is obtained with a sampling frequency higher than the symbol frequency. Indeed, there are often several indirect paths between the transmitter and the receiver due to reflections from surrounding objects, in addition to the direct path in the case of line of sight (see Fig. 2).The received signal is therefore the result of several components associated with distinct delay values ​​and characterized by different amplitudes, phase angles, and arrival directions. Finally, the inter-symbol interference affecting the received signal is not solely caused by the physical propagation channel, since the shaping filters and the frequency response of the electronic components in the transmitter and receiver also generate interference by spreading the response of the physical channel.Thus, the equivalent channel generated by the sequence of these phenomena, and by the sampling rate and timing of the received signal (determined by an upstream synchronization algorithm), is modeled by the multiplication S[n]h[n], where S[n] is the convolution matrix corresponding to the nth TS and h[il] is the baseband impulse response of the equivalent channel. This model represents the behavior of all the processing blocks and the propagation environment between the sampled transmitted signal and the sampled received signal. The sampling rate can be the symbol rate or a higher rate, and this is determined by the precise locations in the transmission and reception chain between which the equivalent channel is to be calculated. In the following, we will use the abbreviation "channel" to refer to the equivalent channel described here.

[0014]

[0015]

[0016]

[0017]

[0018]

[0019]

[0020]

[0021]

[0022]

[0023]

[0024] To simplify the equations, by considering the single-antenna case and omitting the indices n related to the TS, the unbiased LS and MMSE (or LMMSE; on a linear channel model like this one, MMSE and LMMSE coincide) estimators are written respectively as: f ls =s(s h s)' k F mmse =S(ai,1+j where Rh is the channel autocorrelation matrix, cr? is |a variance of the additive Gaussian noise (assumed white here for simplicity of expression), and is a vector that allows the correction of the intrinsic bias of the LMMSE estimator with inputs The matrix of the LS estimator, respectively LMMSE, is then applied to calculate the estimate, LS, of the channel by LS and respectively the estimate of the channel, ?MMSE, by LMMSE: 11 ^LS h h [n] = F^sy[n], ~ MMSE w h [n] =FgMSFy[n], Note that if the channel expectation is known (i.e., the deterministic paths, related to the Rice factors of the paths, which, in practice, occur in the presence of a clear line of sight between transmitter and receiver), a more efficient LMMSE estimator can be obtained: F MMSE' = S(a^I + S H S( R h -m^nÿ) ) ^diag ') fMMSE r -, zr -, n -, h [n]=mh+FMMSE(y[n]-S[n]mh). It should be noted that LMMSE channel estimation can achieve optimal performance, in terms of mean squared error (MSE) criteria, when the propagation channel does not contain significant nonlinearities. This is often the case in terrestrial wireless communications. Furthermore, with reference sequences that satisfy gH g = j, it is possible to rewrite the LMMSE estimator with F mmse = Fls( ali + RJ^diag^') In other words, the linear MMSE estimator can be seen as a correction of the LS estimator.

[0025] Nevertheless, the LMMSE algorithm remains little used due to its exorbitant implementation complexity, which includes:

[0026] - the need to have prior knowledge of the channel statistics and the noise, and in particular the channel covariance matrix: obtaining such statics would require monitoring the channel's behavior on a larger scale than a frame of the PHY layer, thus introducing an additional exchange burden between the PHY layer and higher layers;

[0027] - the inversion of the covariance matrix of the observations (for example of size 16x16, 32x32, 64x64, etc.), which brings a significant computational load.

[0028] For this reason, practical implementations are limited to using an LS channel estimator, which only requires knowledge of the TS and is often implemented in the frequency domain to benefit from lower computational complexity. However, these estimates are highly affected by channel noise and can significantly limit the performance of subsequent processing.

[0029] Figure 11 shows the absolute values ​​of the exact impulse response (IIR) of channel / 2]|, the estimated LS of the channel IIR, and the estimated The LS of the perfectly denoised channel's CIR, Ir 11, is plotted on the ordinate as a function of the path index k on the abscissa, ranging from 1 to KTS, in the case of a cyclic channel. This case corresponds to models where the channel's impulse response is KTS-periodic (typical for OFDM or single-carrier systems with frequency-domain processing) and anti-causal paths are found at the end of the impulse response. Figure 11 presents the case of a channel with few multipaths (subfigure (a), with two unnuisanced paths), called a sparse channel, and the case of a channel with more paths (subfigure (b), with six unnuisanced paths).

[0030] With the rise of deep learning techniques, channel estimation algorithms aided by neural networks have emerged. However, these techniques have generalization problems with respect to the PDP profile of the channel, and while they manage to provide behaviors close to the MMSE estimator, this only occurs for channel families limited to a single statistical profile or to very similar statistical profiles.

[0031] For example, D. Luan and J. Thompson, “Attention based neural networks for wireless channel estimation,” arXiv preprint, arXiv:2204.13465, April 2022, use a “Transformer” type encoder with “multi-head attention” and a decoder that has a convolutional neural network (CNN) structure with residual connections. The Transformer and also the CNN Residuals use a type of normalization called "layer normalization," defined in JL Ba, JR Kiros, and GE Hinton, "Layer normalization," arXiv preprint, arXiv: 1607.06450, July 2016. This structure demonstrates generalization capabilities based on noise power and mobility (Doppler), but not on the channel profile. The model remains very complex, with over 100,000 parameters just for channel estimation and interpolation.

[0032] In D. Luan and J. Thompson, “Low complexity channel estimation with neural networks solutions,” arXiv preprint, arXiv:2201.09934vl, January 2022, the authors study networks with residual connections, based on K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2016. These networks are trained on an Extended Pedestrian A (EPA) profile and then tested in channels based on Extended Typical Urban (ETU) profiles, with a maximum spread of 5 ps, and Extended Vehicular A (EVA) profiles, with a maximum spread of 2.5 ps. These are profiles that are clearly different from the EPA profile and generate significant interference at the receiver. The authors show that, like their network and other similar networks in the literature, lack generalization capabilities, with this training method, the performance in MSE of channel estimators saturates very quickly.

[0033] MV Lier, A. Balatsoukas-Stimming, H. Corporaal and Z. Zivkovic, "OPTCOMNET: Optimized Neural Networks for Low-Complexity Channel Estimation," ICC 2020 - 2020 IEEE International Conference on Communications (ICC), pp. 1-6, June 2020, discusses the improvement of channel estimation and its interpolation in the frequency domain. The authors use a simple residual structure with fully connected hidden layers between the input and the residual connection, so that the network learns the noise that must be subtracted to denoise the channel estimation. The results of this work also show that it is more advantageous to use networks with parameters optimized over specific signal-to-noise ratio (SNR) sub-intervals to achieve better performance.

[0034] CN 113572708A proposes a neural network that outputs denoised channel estimates based on input channel estimates. However, this solution is not satisfactory in terms of efficiency and generalization. Indeed, the structure used is a large network with a chain of several fully connected layers, with approximately 328,832 parameters and almost as many operations. The neural network is trained by setting the denoising threshold to 0.0001, which will limit its generalization to different signal-to-noise ratios (SNRs) and fading channels. Furthermore, the use of an L2 norm cost function (or MSE criterion) does not allow for fine-grained learning at high SNRs.

[0035] In summary, in the prior art:

[0036] - classical improved channel estimation methods are either limited by their computational complexity (e.g. matrix inversions) or not feasible due to the difficulty of obtaining metrics (e.g. transmission channel statistics) which are required by these algorithms;

[0037] - improved channel estimation methods that use neural networks encounter various generalization problems, particularly when studying networks that attempt to improve LS estimates:

[0038] when considering networks that are trained on a database with certain specific channel profiles similar to each other, these networks learn to handle them well, to approach the performance of the MMSE estimator, but they degrade performance on channels that are not part of the database,

[0039] alternatively, when considering networks that are trained on a database with a wide variety of channel profiles, these networks lose their interest and bring little or no improvement compared to the LS estimates.

[0040] In conclusion, although some references are beginning to address the generalization of neural network-based channel estimators to substantially different channel profiles, a solution with reasonable complexity has not yet been revealed.

[0041] There is therefore a need to improve the quality of CIR estimates while limiting the complexity of estimation. Summary of the invention

[0042] To this end, according to a first aspect, the present invention describes a method for improving the determination of the estimate of a telecommunication channel between a transmitting device and a receiving device, as a function of at least one signal frame comprising iVjg predefined pilot sequences, each of length KTS, transmitted via said channel to the receiving telecommunication device and

[0043] a first estimate of the telecommunication channel having been carried out, following the reception of said frame by said receiving device, as a function of said received frame and said NTS predefined pilot sequences and having provided at least first estimates of the impulse response of the channel ji NTS; [h [n]j . k LJJ 21=1

[0044] with h'[n] = (¾n], h2»L ...,¾^]] °ù 2^ [ 22] corresponds to the kth path of the impulse response of the teletransmission channel relative to the jth pilot sequence of the frame;

[0045] said method being characterized in that it comprises the following steps implemented by the receiving device:

[0046] - supplying input to a neural network of the first receiving device input data, and obtaining at the output of said network an estimate of a binary mask nik, k = 1 to be determined as a function of said first input data provided, such that:

[0047] Ülk= 0' if the path with index k is determined to consist only of noise;

[0048] Êlk= 1, otherwise;

[0049] said first input data being a function of the correlation between the received frame and said predefined pilot sequences;

[0050] - determination of estimated second channel lengths, r NTS, as a function of first estimated channel, r ] NTS, and determined binary mask, according to the The following rule:

[0051] hk[n] = îik[n], simk=l 0, if nik = 0

[0052] The invention thus proposes a channel estimation method with sparsity detection (i.e. of the channel support) by neural network, followed by the determination of an estimate of the impulse response of the denoised channel, as a function of this sparsity, so as to approach the ideal denoising behavior of the [Fig.11].

[0053] The “channel sparsity detection” is the estimation of the “support” of the channel’s IRC in the delay domain, at the sample or symbol time, depending on the application: it is a matter of estimating at which delays (which correspond to indices of the symbol or sample time, i.e. to distinct paths) the channel has a non-negligible power.

[0054] In embodiments, such a method will further comprise at least one of the following features:

[0055] - the first input data of the neural network include at least one of the data sets from among the following sets relating to said framework:

[0056] - the first estimates of the channel r > -i ^ts,

[0057] - synchronization criterion(a) {c[n]}^sync, characterizing the response Zî=l impulse of the channel;

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068] - the first estimates of the channel r < i Nys and the estimates of the SINR through the U[n]| 1 variance of the noise '2 relative to said frame or of the covariance matrix J? ; - the first estimates of the channel r > ■» Nts and the estimates of the SINR through U

[13] } , VLJJ J2=l the noise variances r,2 rn nts or the covariance matrices rrn ^rs relating to each pilot sequence; - synchronization criterion(a) r ^ [ ] 1 NSych, characterizing the response 1 JJ n=l channel impulse and SINR estimates through noise variance g or covariance matrix R; - The neural network was previously trained during a learning phase using a set of initial input data from training frames such that, for initial input data corresponding to a given training frame, the desired output mj^ of the neural network, to which the output provided by the neural network during the learning phase is compared, is fixed at: - 1 if and only if a predefined criterion critest is greater than a predefined threshold T - 0 otherwise; where the CTit^ criterion is a function of the low or no noise channel realizations on the pilot sequences and at least one element, among the Cfit^ criterion and the threshold T is determined as a function of the SNR or SINR on said telecommunication channel for said training frame considered; - the threshold Th is chosen equal to {J^r / KTS, where (j2, is the variance of the noise on the training frame considered; - the threshold Tb is chosen equal to ^>2 p , where ^2 is the estimate of the noise variance over 11 w1 cuw the frame considered and Fc is a constant which, over a range of SNR of interest for the profile of the channel considered, minimizes the root mean square error expressed in the form: / / v \ H f N \ 1 minE ([n] -h [n] ) ) ] where E is the statistical expectation, on stochastic realizations of the channel within of the profile and the SNR range considered, and r is a channel estimate obtained through the denoising of the previously estimated h'[n] with the ideal mask m* via h[n] = m*0h[n] '

[0069] the criterion critk = n ] hRw[ n] -1hk [ n ] , with Rjn] the covariance matrix of the noise affecting the nth pilot sequence of the considered learning frame;

[0070] The criterion crit^ will be calculated from the formula critk= [ n ] HRW[ n ] , with the covariance matrix of the noise affecting the nth pilot sequence of the considered learning frame;

[0071] but replacing in said formula:

[0072] [n] by the average value Rw of the noise covariance matrix affecting the learning framework considered and / or

[0073] the impulse responses of the channel by their time average over the frame;

[0074] - during the learning phase, the weights of the neural network are obtained in function of a cost function based on crossed binary entropy;

[0075] The neural network comprises:

[0076] - one or more layers of aggregation of the first input data to have Nine aspects of 1X size tensors

[0077] - one or more layers for going from Naspects to Nc aspects of tensors of size 1 X KkS, where Nc is the number of intermediate aspects of the network;

[0078] - Nk ResNet-type layers, and

[0079] - one or more layers for going from Nc to NTS aspects of size tensors 1X Kts;

[0080] The method for improving the determination of the estimate of a telecommunications channel includes a step of:

[0081] - supplying, as input to another neural network of the receiving device, of second input data of said other neural network comprising at least the second estimated values ​​of the channel, r > i Nts, and obtaining at output of said other LLJJ estimated third-order neural network of the channel, rri Nts;

[0082] according to which said other neural network results from a learning phase of the other network aimed at minimizing a cost function, said cost function being a function of the mean squared error between the outputs obtained from said other network in the learning phase and the desired values ​​of the true impulse response of the channel, obtained by simulation;

[0083] The method for improving the determination of the estimate of a telecommunications channel includes at least one of the following provisions i, j:

[0084] i / the second input data of said other neural network further comprise at least one element among probabilistic information of the channel support delivered at the output of the neural network for estimating the channel support and an estimate of the noise variance;

[0085] j / said other neural network comprises:

[0086] - one or more layers for going from aspects to Nc aspects of tensors of size 2 X Kyg, where Nc is the number of intermediate aspects of the network and greater than NTS;

[0087] - Nl ResNet type layers, and

[0088] - one or more layers for going from Nc to NTS aspects of size tensors 2 x Kts.

[0089] According to another aspect, the invention describes a telecommunications receiving device adapted to determine the estimate of a telecommunications channel between said receiving device and a transmitting device, as a function of at least one signal frame comprising predefined pilot sequences, each of length KTS, transmitted via said channel to the telecommunications receiving device, in which

[0090] a first estimation of the telecommunication channel having been carried out, following the reception of said frame by said receiving device, as a function of said received frame and said predefined pilot sequences and having provided at least first estimates of the impulse response of the channel r • ■> NTS;

[0091] with h'[n] = [^[nhKjn], ]

[0092]

[0093]

[0094] where corresponds to the klcmc path of the impulse response of the channel teletransmission relating to the jth pilot sequence of the frame; said receiving device being characterized in that it comprises a neural network adapted to receive initial input data, and to obtain at the output of said network an estimate of a binary mask fhk, k = 1 to KTS, determined as a function of said initial input data provided, such that: mk=Q, if the path with index k is determined to consist only of noise; lïlk= 1, otherwise;

[0095] said first input data being a function of the correlation between the frame received and the said predefined pilot sequences;

[0096] said receiving device (10) being adapted to determine second estimated values ​​of the channel, ri ^ts, as a function of first estimated values ​​of the channel, (h In] j. ^ ts, and the determined binary mask, according to the following rule:

[0097] hk[n], simk=l 0, simk = 0 n = 1, Nts Brief description of the drawings

[0098] The invention will be better understood and other features, details and advantages will become clearer from the following description, given by way of non-limiting reason, and from the accompanying figures, given by way of example.

[0099] [Fig-1] Fig. 1 represents a wireless teletransmission system with a a transmitter and a receiver;

[0100] [Fig.2] Fig.2 schematically represents the origin of a multipath channel, the paths being generated by reflectors encountered by the electromagnetic waves of the signal;

[0101] [Fig.3] The [Fig.3] is an example of the structure of a frame in the case of a single-carrier transmission;

[0102] [Fig.4] The [Fig.4] schematically represents processing blocks in a telecommunications receiving device in one embodiment;

[0103] [Fig.5] The [Fig.5] is a flowchart representing a method for improving the determination of the estimate of a telecommunication channel in an embodiment of the invention;

[0104] [Fig.6] The [Fig.6] is an example of the architecture of a neural network implemented in block 12 (of the [Fig.4]) of mask estimation in an embodiment of the invention;

[0105] [Fig.7] The [Fig.7] is an example of structures implemented in the architecture of the [Fig.6] in one embodiment of the invention;

[0106] [Fig.8] The [Fig.8] is an example of the architecture of a neural network implemented in block 16 (of the [Fig.4]) of channel estimation refinement in an embodiment of the invention;

[0107] [Fig. 9] Figure 9 is an example of a neural network implemented in block 12 (in Fig. 4) for mask estimation in an embodiment for single-carrier frames with four pilot sequences tN^s = 4) of K?s = 32 symbols;

[0108] [Fig. 10] Figure 10 is an example of a neural network implemented in block 16 (in Fig. 4) of channel estimation refinement in an embodiment of the invention for single-carrier frames with four pilot sequences (Nts = 4) of K^g = 32 symbols;

[0109] [Fig. 11] The [Fig. 11] illustrates the result of an ideal channel denoising solution compared to an LS method.

[0110] Identical references may be used in different figures when they refer to identical or comparable elements. Description of the implementation methods

[0111] A telecommunications receiver device 10 is partially represented in [Fig.4], in one embodiment of the invention. It comprises, in addition to a data reception block on a telecommunications channel (not shown), for example radio frequency (UHF, HF, 4G, 5G ...), a synchronization and deframing block 31, named SYNC_DETRAME 31, a first channel estimation block 11, named EST_CNL_0 block 11, a channel estimation improvement module 20, named AML_EST_CNL 20 and a subsequent processing block 30, named EGAL_DET_DEC 30.

[0112] Channel estimation improvement module 20 includes a mask estimation block 12, named IS_MASK, a second channel estimation block 14, named IS_CNL_DEBRTGE and a channel estimation refinement block 16, named IS_CNL_RAFF.

[0113] The IS_MASK 12 block includes a neural network, called a sparsity detection neural network, adapted to identify the channel support, in the form of the estimation of a binary mask in indicating the useful paths of the impulse response.

[0114] The optional EST_CNL_RAFF 16 block is adapted to further improve the quality of the denoised estimates, using, in the example described below, information from the channel support detection neural network. This EST_CNL_RAFF 16 block is, for example, implemented using a neural network architecture.

[0115] The SYNC_DETRAME block 31, the EST_CNL_0 block 11, the AML_EST_CNL module blocks 20, and the subsequent processing block 30 are adapted to implement their respective operations, detailed below with reference to [Fig. 5], representing the steps of a method 100 for improving the estimation of a teletransmission channel in an embodiment of the invention. Steps 101 to 105 The 100 steps of the process are implemented relative to a frame received by the receiving device 10, from a transmitting device, via a telecommunication channel. The frame structure is, for example, that described with reference to [Fig. 3].

[0116] A phase 200, prior to these steps 101-105, will be described later.

[0117] In step 101, the SYNC_DETRAME 31 block performs the operations of Synchronization and deframing, in particular to extract the TS, are performed based on a newly received signal frame by the receiving device 10. These operations consist of determining the sampling time and estimating the residual phase and frequency shifts. After correcting these shifts, this block 31 then extracts, from the received signal, the parts corresponding to the TS, yjn] Il = 1 ct 'cs corresponding parts of the data blocks (not shown in Figure 4) are provided to the relevant sub-blocks of the receiving device 10, after possible subsampling. In order to find the sampling time that ensures proper receiver operation (and to estimate phase and frequency shifts), block 31 calculates a synchronization criterion {cn where NSVnch corresponds to the search window for the sampling instant, and the c[u] vectors are often derived from the cross-correlations of the pilot sequences (opportunistically offset) with the received signal, and possibly from the autocorrelation of the received signal. The precise techniques for obtaining these vectors are highly specific to each system and are well known to those skilled in the art, without directly affecting the inventiveness of the method revealed in this document. What is important to note here is that the synchronization criterion contains information correlated with the channel's impulse responses, information that a neural network can extract and utilize.

[0118] The EST_CNL_0 block 11 determines, in a first estimation of the channel performed by the receiving device 10, the estimates, named î,r _ 1 > [nj, u — i, ••• / 1v jr§ k = 1, ..., Kts, of the channel impulse response on the received frame after processing by block 31, as well as an estimate of the variance (or the wide-sense covariance matrix n) of the noise on the frame, as a function of yfll] U = 1, ..., Nts where u's the received signal corresponding to the jth TS in the frame. It uses an LS algorithm or any other channel estimation algorithm. The algorithm is, for example, a simplified LS type (e.g., when the drivers have ideal autocorrelation and the associated autocorrelation matrix is ​​the identity), or a regularized LS algorithm (to reduce noise sensitivity), or a simplified MMSE algorithm (which, for example, assumes a determined statistical profile), or an algorithm which uses a signal containing interference from an unknown data block (i.e., "data-aided"), etc.

[0119] In step 102, the parsimony detection neural network of the IS_MASK 12 block identifies the channel support, that is, the set of positions on the impulse response vector that correspond to a noisy useful path, and not just noise: it outputs an estimate of a mask; this mask, output from the IS_MASK 12 block, takes the form of a binary vector m = [^KTS] where the index fhk = 0 means that the j^th entries of the vectors of the frame channel estimates with n = 1, ..., N correspond only to noise, whereas on the contrary, mk = 1 means that the ^th inputs of the channel estimate vectors î,' r _ 1 at are noisy paths containing useful signal from the 22^- L12 J, 12 — 1, ., •, IV j, g frame. This mask, in one embodiment, is implicitly provided as a binary probabilistic or binary log-probabilistic estimate of the presence of a useful path. In other words, the neural network provides the following output in one embodiment: - a positive probabilistic vector Pm that indicates useful path positions with a value greater than 0.5, and noise positions with a value less than 0.5, or - a log-probabilistic vector that would indicate "useful path" positions with a positive value, and "noise" positions with a negative value, such that the absolute value of these indicates the confidence of the network in its prediction. Then the value ïhk of the mask at position k will be obtained with >0.5} or with 1 ( > 0) where the indicator operator 1(P) is 1 if the condition P is true, and 0 otherwise.

[0120] In a step 103, the EST_CNL_DEBRTGE 14 block determines a denoised estimate of the channel, Cri „ * ,k=l, .... as a function 22^ [ 12 J , 12 — lz ..., 1V 'pg 1 of the mask delivered in step 102 and the estimated r-, tv, ^kLJz n — L, k= 1, ..., Kj-g, which had been delivered at step 101 by the EST_CNL_0 block 11 with :

[0121] 4(n] = . hk[n], mk=l | 0, m k =Q n — 1, Nts Note also: ^"r -, î'rh [n] =h [n]Om In vector format, the operator 0 indicates the product element by element.

[0122] In an optional step 104 (and therefore shown in dashed lines in Fig. 5), the EST_CNL_RAFF 16 block further improves the quality of the denoised estimates, for example by using information from the sparsity detection neural network in the EST_MASQUE 12 block. This EST_CNL_RAFF 16 block is also implemented, for example, using a neural network architecture. This neural network receives as input hk[nL 21=1, ...,Nts, k=l, ...,Kts, at least and delivers at output, rori NTS, an estimate of the best quality channel ih[n] (i.e., closer to the actual construction of the canal).

[0123] In a step 105, further processing is performed on the frame by the EGAL_DET_DEC 30 further processing block, for example equalization and / or useful symbol detection and / or decoding of detected symbols based on the symbols in the frame data blocks and the channel estimates delivered by the AML_EST_CNL 20 module, i.e. according to embodiments [ r fl] or hk[nL 22=1, ...,Nts, k=l,

[0124] Steps 101 to 105 described above are implemented in the inference phases of the neural networks considered.

[0125] The parsimony detection network

[0126] To achieve satisfactory use of neural networks, it is important to clearly define the problem to be addressed, to choose the inputs and outputs of the neural network appropriately, and to define the target outputs (sometimes called labels in English on supervised learning methods, such as the one described in this solution) that we want the network to reach.

[0127] The inputs to the neural network of the IS_MASK 12 block include one or more of the three quantities below and must include at least one of the first two quantities described below, for each frame considered successively (both in the learning phase and in the inference phase):

[0128] - pre-estimated channel r • ■> Nrs obtained on the different sequences U

[22] } , LLJJI / =l pilots of the received frame; these estimates are a function (at least) of the correlation between the received frame and the predefined pilot sequences;

[0129] - synchronization criterion(a) { jj]}^synch, for example from a correlation of pilot sequences with the received signal or any other algorithm that allows the calculation of a synchronization criterion implicitly containing a

[0130]

[0131]

[0132]

[0133]

[0134]

[0135]

[0136]

[0137] information on the channel impulse response, potentially in oversampled domain; - SINR estimates through the noise variance (or the covariance matrix Rw in the broad sense), case illustrated in Figure 4. More generally, these estimates can also vary within the frame, for example it is possible to provide an estimate by TS, ।} ^TS (°U { rS^' œ CaS n is Pas ^ustr^ in the [Fig.4]. From a subset of these inputs, the neural network in block 12 estimates a binary mask: the desired output of the neural network in block 12 is, in one embodiment, a log-probabilistic vector that indicates the positions of the "useful paths" with a positive value, and the positions of "noise" with a negative value, such that the absolute value of these indicates the network's confidence in its prediction. Then the IS_MASK block 12 determines a binary mask in that indicates the positions of the "useful paths" with: hnjt 0 • . 0, otherwise In embodiments of this sparsity detection network, one or more of the provisions detailed below are implemented for the following three aspects: - strategy for building training, validation and testing databases; - training strategy; - architecture. The first two steps in building the databases and training strategy include: • The use of a channel impulse response database (as perceived by the telecommunications system receiver at the input of the signal pre-estimation block 11 in the absence of noise, or with a noise level negligible compared to the minimum noise level in the telecommunications system's operating range). This database can be generated by a simulator or can be obtained from electromagnetic channel survey campaigns. • The use of the EST_CNL_0 11 block to provide the preliminary channel estimates and possibly the noise variance estimate, and additional operations which are described in the section "Strategy for building training, validation and testing databases" of the sparsity detection network" to obtain a training dataset. and the channel support detection neural network of the block IS_MASK 12 being trained, whose structure and parameters are used during the training phase described in the section "Training strategy for the sparsity detection network".

[0138] Step 200 outputs the final parameters of block IS_MASK 12, obtained through training. Blocks 14 and 16 are not used in the training of the sparsity detection network, but only for channel refinement.

[0139] In one embodiment, one and / or the other of these provisions are implemented in a pre-step 200 to steps 101-105 by an electronic construction and training platform as illustrated in [Fig.5].

[0140] Strategy for constructing training, validation and testing databases for the sparsity detection network:

[0141] Even if the physical propagation channel is very parsimonious, shaping filters and receiver filters spread the useful paths over several symbol-time positions, according to the definition of the sampling frequency and the characteristics of the filter.

[0142] Initially, it is therefore important to define what a "useful path" is at this level of abstraction of the channel's impulse response. Several definitions are possible; in the present case, the path 13] at the jth position of the impulse response of the ideal channel is considered "useful" if and only if Cfit^ > T where Th is a threshold (not necessarily fixed) and Cfit^ is a criterion that aggregates the channel realizations on the pilot sequences. By For example, the criterion is chosen to be equal to | nts I 2 on a critk = low-mobility system, for example, or •. _ 1 rnll2n~2 When the system is likely to see channels that vary rapidly during the time frame, with (j^) the noise variance. Indeed, if the CIR changes rapidly enough in a frame that between two TS n and n' a negligible path becomes non-negligible, or vice versa, the average power will give more weight to considering this path as useful, while the average power will tend to "absorb" it into the average, in the hope of better denoising the channel. Therefore, this choice lies in the assumptions about the frame duration with respect to the channel coherence time, which describes the duration over which the The channel would change relatively little. In the case of a multi-antenna receptor, these two criteria will be written respectively critfr= ) And critk = ^^[hk[n] HR^hk[n] , with Rw the noise covariance matrix. The presence of the noise variance cr^ (or respectively the noise spatial covariance matrix Rw) within a criterion involving the noise-free channel hk[n] (or respectively hk[n]) will allow the definition of a "useful path" to be adapted to the noise level affecting the previously estimated f_r, which we will seek to denoise. In a mode of "kl fl J In practice, we could further generalize the criteria of this process, where we could have a variance ¢7^22] (or respectively of the spatial covariance matrix of the noise, different by TS of index n. Therefore, Iïl*k, which is the desired output of the network of neurons (used in the learning phase), is defined as

[0143] ^k = fl, critk>Tfi . 0, otherwise

[0144] where the criterion Cfit^. is to be chosen according to operating assumptions of the system, among for example the options mentioned above.

[0145] The choice of the best threshold Tk is, for example, carried out through the minimization of The mean squared error (MSE) is expressed by the following formula:

[0146]

[0147] minE[ - £ [ n] ) * & n ] -k[n] ) ] j L ' £2— J. / ' 12—1 f J n where the expectation E[ • ] was calculated over several noise configurations (i.e., SNR or SINR), channel profile, channel velocity, and relative position of the pilot sequences. In the formula, denotes an ideal channel estimate in terms of sparsity, obtained through denoising the previously estimated j?' with the ideal mask m* via h' [ n] ' and is the true answer The channel impulse response is therefore a function of the threshold Tk. Indeed, since propagation channels are simulated on computers, or emulated by channel emulators, it is possible to achieve the exact realization of the IRC in a laboratory environment. Therefore, by exploiting this opportunity, these databases can be built.

[0148] The inventors have determined that a constant mask m for a channel profile over all operating SNRs (or SINRs) does not lead to a very attractive optimization of the MSE. However, an adaptive mask, obtained through either a criterion or a threshold dependent on the SNR (or SINR), makes it possible to reach better operating points. The examples of Cflt^ given previously allow us to consider this case, even when the thresholds are fixed such as T = 1 / KTS. Nevertheless, if the criterion Critne does not depend on the noise variance (j^, (or respectively on the spatial covariance matrix of the noise Rw), as for example the criterion _ 1 -y^rsi r il |2 (or respectively C1- NtsZ72=1 I n JI critk= ), then a threshold that depends on the SNR (or SINR) such as (or T = ÏT(Rw) / Kys (in the multi-antenna case, where tr is the trace of the matrix) allows for a satisfactory result over a wide range of channel profiles. For better performance, in one embodiment, a constant Fc is determined for each statistical channel profile from among a plurality of predefined profiles, such that T = 0^vjFc optimizes the MSE for the channel profile in question over an operating SNR range. The operating SNR range depends on the target communication system and corresponds to the SNR range within which this communication system operates, i.e., allows information exchange with acceptable transmission reliability for the services provided by the communication system itself.In this case, T also becomes a function of the channel's statistical profile, but this does not hinder the practical feasibility of the solution, since these processes are done offline to create the database of labels (target outputs) to be used for training.

[0149] Once this work is completed, databases are created containing the entries (at least one of the quantities among i NTS v and possibly {h [n]} ,, Mn]} , ' also an additional input q- '^) and targets 111, where m* is the objective mask, obtained with the choice of the threshold corresponding to each underlying channel profile used when creating the database.

[0150] Finally, to obtain the most robust learning possible, it is recommended to extract the aforementioned quantities for the databases by means of a simulator representative of the waveform and physical layer processing, and in the presence of potential imperfections and randomness.

[0151] For example, it is used in one embodiment, a simulator representing the behavior of the waveform and the physical layer, where the following hazards and imperfections have been represented: - a mobile transmitter and receiver (e.g., up to 60 km / h), - a CFO at the transmitter (e.g., 2 ppm), - a random propagation delay up to the maximum delay manageable by the waveform, - a set of PDPs derived from standardized profiles and artificially generated exponential profiles.

[0152] The preceding list of hazards and imperfections is provided for illustrative purposes only and should not be interpreted as limiting. Other hazards and imperfections may be added or substituted depending on the relevance and importance of these effects for the system using the solution. For example, other effects include a CFO also at the receiver, a non-linear amplification model, an I / Q mismatch model, in general any type of imperfection in the radio frequency (RF) blocks at the transmit and receive ends, imperfect antenna calibration models, phase noise models, propagation delays or channel time spread that exceed the limits that can be adequately managed by the system, jammers or interfering systems, and in general any type of impulsive or non-impulsive noise that is not optimally managed by the system, etc.

[0153] The construction of different databases (i.e., datasets) is recommended for training with a few tens of thousands of received frames distributed according to specific SNR ranges. For example: - a dataset 'L': consisting only of data from frames received with a signal-to-noise ratio (SNR) between 0 and 20 dB, - a 'H' dataset: consisting only of data from frames received with a signal-to-noise ratio (SNR) between 30 and 40 dB, - an 'MLH' dataset: corresponding to 20% of SNRs from 0 to 20 dB and 80% of SNRs from 30 to 40 dB, and - a dataset 'M': corresponding to the set of frames received over a wide SNR range from 0 to 40 dB.

[0154] Among the datasets mentioned, in each use, in one embodiment, a portion of the data (for example, 10%) is reserved exclusively for validation, and the remainder (90% in the example) for training. All these percentages and definitions of dataset allocation are not binding, in the sense that a person skilled in supervised learning knows how to adjust these percentages to obtain performance as close to optimal as possible.

[0155] In addition, completely independent test datasets are created for each standardized PDP, and with discrete-valued SNRs to evaluate MSE performance by calculating the median and quantiles of the quadratic errors.

[0156] Training strategy for the sparsity detection network:

[0157] The training strategy plays an important role in finding neural structure weights well-suited to solving the algorithmic problem posed, in this case identifying the channel's impulse response support. A typical optimizer such as the ADAM algorithm or another variant of the stochastic gradient descent (SGD) algorithm can be used with standard values ​​of leaming rate and mini-batch size.

[0158] The cost function (or "loss function"), as is known, determines the weight values ​​used at the end of the training phase. The goal of training is to decrease the cost calculated by the cost function with each iteration of updating the neural network's weight values, until the outputs of the neural network are sufficiently close (according to a predefined threshold) to the respective desired outputs, here the ideal masks m', or until a maximum number of iterations is reached. The cost function is chosen in an embodiment equal to a function based on the binary cross-entropy between the ideal mask and the log-probabilistic output of the neural network.

[0159] For this, we can give the example of "MultiLabelSoftMargin"

[0160] loss(^m^ = or the example of “BinaryCrossEntropyWithLogitsLoss”

[0161] lossfY^xn^p) = -2^[pmyog(a(7a^)) + (1-^ where j : xh is the sigmoid function and p> 0 is a weight that strengthens (p > 1) or attenuates (p< 1) the impact of accurately predicting useful paths (ie = 1).

[0162] In one embodiment, a curriculum learning strategy is further employed, where the training and validation datasets can evolve as the learning epochs progress. For example, learning could start with the 'H' dataset to thoroughly learn the high SNR case, then subsequently include the low SNR cases with the 'MLH' dataset, and finally with the 'M' dataset to cover a broad range of SNRs, allowing the network to generalize across the entire range. Another example involves creating a dataset for each phase of learning by selecting a variable percentage of elements from a pre-existing set of datasets, for example, 'L', 'H', 'M'. Other methods known to the person skilled in the art for creating datasets relevant to a curriculum learning strategy can be used.

[0163] Finally, in one embodiment, the weights of the best learned network are saved based on the minimum of the cost function (loss criterion), or on the accuracy criterion per path (i.e., n [ 1 v^rsi ( ™ \ 1 ), or on BL kts ^k=l1 N the more restrictive criterion of accuracy of having correctly detected all the support of the realization of the impulse response of the channel (ie r? TTT^rsi / \ ), [11À-=11 ( mk ~ mk) with the indicator operator 1(P) which is 1 if the condition P is true, and 0 otherwise. It is also possible to run the learning procedure several times with different initializations of the network weights, according to known state-of-the-art methods, and to retain at the end the weights resulting from the trial that minimized the cost function and / or the validation error, and / or the other criteria previously described in this document or known to those skilled in the art.

[0164] Architecture _ for the sparsity detection network:

[0165] It has been established by the inventors that the residual layers (ResNet - "Residual A "network," that is, a set of neural network layers with residual connections as defined in K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016, with convolutional layers, exhibits very interesting behaviors and allows for rapid convergence in just a few epochs. The use of normalization layers has also contributed to convergence toward particularly interesting results.

[0166] A particularly efficient ResNet-based sparsity detection network architecture embodiment is illustrated in [Fig. 6], comprising:

[0167] - one or more input layers that allow the aggregation of observations retained (previous channel estimates, synchronization metrics, or any other function derived from this information) are used to construct a second-order tensor (i.e., a matrix) of size Njn X Kyg, where Nest is the number of input aspects of the network. An aspect (or feature) here corresponds to a dimension of the tensors processed by the neural structures and is also known as a channel in the jargon of neural structures (this does not, of course, refer to the propagation channel). The terms "depth" or "volume" are sometimes also used to refer to the number of aspects. Furthermore, using absolute values ​​of the impulse response inputs can have the dual benefit of using real inputs instead of complex inputs (reducing dimensionality and therefore the memory used and number of network weights) and evaluating the power of the inputs, especially when the inputs These quantities are scaled relative to the variance (or covariance matrix) of the noise and interference. Finally, using these quantities in the logarithmic domain facilitates learning thanks to a higher dynamic range of observations.

[0168] - one or more neuronal layers to allow the passage of Njn aspects to Nc aspects, where Nc is the number of intermediate aspects in the network. The larger the number of aspects, the easier it will be for subsequent layers to identify an interesting feature of the data to achieve the objective (accompanied by increased computational complexity, which requires a trade-off).

[0169] - NL residual connection layers that work on these Nc aspects, in summing their inputs to their outputs through residual connections. In particular, ResNet structures such as in [Fig.7], where the ResNet has one (case a) or several (case b) non-linearities and where the order of placement of the layers can vary (case a, b, c), and where some inputs can be directly connected to the output without being affected by the processing (case d).

[0170] - one or more neuronal layers to allow the passage of Nc aspects to a single output aspect which will correspond to the log-probabilities of presence or absence of paths on the different indices over the length of the TS, KTS. Subsequently the desired mask is obtained by in = 1(1^ > 0 ).

[0171] The refinement network of the estimated

[0172] The neural network in block EST_CNL_RAFF 16 is called the estimate refinement network and receives as input a pre-existing estimate of the channel impulse response, here r Nts, potentially denoised thanks to the [h [n]j . mask of the sparsity detection network, and aims to provide at output an estimate of the impulse response of the rni channel with even less noise.

[0173] Optionally, one or more of the following inputs may be used: • the estimated noise variance, for example from the block 11, or another noise variance estimator, for example based on knowledge of the mask m which would use the IRC inputs corresponding to noise to obtain a new estimate of its variance; • the (log-)probabilistic estimates of the knowledge of useful paths and noise (which could for example come from the previously mentioned sparsity detection network).

[0174] In embodiments of this CIR estimate refinement network, one or more of the provisions detailed below are implemented for the following three aspects:

[0175] - strategy for constructing training, validation and databases test ;

[0176] - training strategy:

[0177] - architecture.

[0178] The first two steps in building the databases and training strategy include: • The use of a channel impulse response database (as perceived by the telecommunications system receiver at the input of the channel 11 pre-estimation block in the absence of noise, or with a noise level negligible compared to the minimum noise level in the telecommunications system's operating range). This database can be generated by a simulator or obtained from electromagnetic channel surveys. The database contains the training database targets for the channel estimation refinement neural network. • The use of block 11 to provide the preliminary channel estimates and, optionally, the noise variance estimate. This block performs the additional operations described in the section "Strategy for Building Training, Validating, and Testing the Parsimony Detection Network" to obtain a training dataset for learning the support detection network, but only if the latter is trained in conjunction with the refinement network. If only the refinement network is trained, these additional operations to obtain the training dataset for learning the support detection network are not performed. • The use of block 12, i.e., including the support detection neural network, whose structure and parameters are used during the training phase described in the "Training Strategy for the Parsimony Detection Network" section, if the support detection neural network is also trained. In the case of training only the refinement network, block 12 is used solely for inference; i.e., the weights of the neural network in block 12 are fixed, and this network is used only to calculate the outputs. Block 14 is used to calculate denoised channel estimates from the mask provided by block 12 and the estimates provided by block 11. In the presence of noise in the received signal derived from the database channels, the outputs of block 14 serve as inputs to the training set of the refinement neural network. Further details are available in [reference to relevant section]. the section "Strategy for building training, validation and testing databases for the refinement network". The use of block 16, i.e. the refinement neural network, of which the structure and parameters are used during the training phase described in the section "Refinement Network Training Strategy".

[0179] Step 200 outputs the final parameters of block 16, obtained through training. In the case of joint training of block 12 and block 16, step 200 also outputs the final parameters of block 12.

[0180] In one embodiment, one and / or the other of these provisions are implemented in a prerequisite step 200 to steps 101-105 by an electronic construction and training platform.

[0181] They are less restrictive than those described with reference to the parsimony detection network.

[0182] Strategy for constructing training, validation and databases refinery network test:

[0183] The elements indicated above relating to the construction of the database The rules for the parsimony detection network also apply here. However, it will be necessary to add the impulse response of the ideal channel and the variance of noise (j^ (or the covariance matrix Rw in the multi-antenna case) as additional targets in the databases. Indeed, since propagation channels are simulated on computers, or emulated by hardware channel emulators, or come from channel measurement campaigns in realistic deployments (also called channel surveys), it is possible to access, in a laboratory environment, the exact realization of the CIR and SINR. Therefore, in By exploiting this opportunity, these databases can be built.

[0184] Refinement network training strategy:

[0185] The major difference in the training strategy compared to the support detection network is the cost function: here both the input and output are estimates of the impulse response of the channel {}nts, which are complex quantities: cost functions that are a function of the squared error are used for example, the training aims to minimize this cost function (at least to make it less than a predefined threshold).

[0186] The mean squared error function ("MeanSquareError") can in particular be used: 101871 loss({£[«] This allows for satisfactory results over a given SNR range. However, when the range is too wide, this cost function will give more weight to low SNR samples than to high SNR samples (which is potentially attractive for channel estimators, depending on the system's operating point). For better generalization of the network to any SNR, the use of the SNR-normalized mean square error ("SNRNormalizedMeanSquareError") is recommended. 101881 loss( 1X {htm = This allows the learning process to give equal weight to high SNR samples and low SNR samples. In the case of a multi-antenna receptor, these two loss functions are expressed respectively as loss({Â[n]}^ And lossj {*[n]}^=r {i[n] -*t[n])

[0189] In one embodiment of the refinement network training, the support detection network of channel 12 is kept fixed, for example because the parameters (neuron weights) of the block 12 neural network have been pre-trained according to one of the methods described previously or in the section "Training Strategy for the Parsimony Detection Network" and are already available. In this case, the cost functions described in that section can be used for training the refinement network 16.

[0190] In another embodiment of training the refinement network 16, the channel support detection network of block 12 is trained jointly, according to strategies known to those skilled in the art. For example, a global cost function can be used, formed by the weighted sum of one of the cost functions used for the network of block 12 and one of the cost functions used for the network of network 16. In this case, at each step of the chosen SDG algorithm, the parameters of both networks are updated. Another joint training strategy consists of optimizing the parameters of one network (e.g., network 16) during one or more steps of the chosen SDG algorithm while keeping the parameters of the other network (e.g., block 12) fixed. Then, the network that has undergone one or more steps of The parameters of one network (e.g., network 16) are kept fixed, while the other network (e.g., block 12) is updated during one or more steps of the SDG algorithm. This alternating optimization of the two networks is conducted over a number of iterations until convergence is reached or until one of the optimization stopping criteria known to a person skilled in the art is met. During these alternating training phases, one of the relevant cost functions for the network being optimized is chosen. The cost functions can be changed at each iteration of the alternating optimization.

[0191] Refinement network architecture:

[0192] Similar findings to those of the support detection network were made during experiments with the refinement network: therefore, an architecture similar to that of Figure 6 is recommended, but with the main difference being that it deals with complex numbers this time. Thus, as illustrated in Figure 8 representing ResNet-based refinement network architectures, it is possible either to work with these quantities, considering the real and imaginary parts as two aspects ("features") of KTS elements, and work with one-dimensional convolutions, or to concatenate them in the last dimension to consider them as a single aspect of 2KTS elements, or to consider it as a single aspect of 2 × KTg elements and work with two-dimensional convolutions.

[0193] Example of implementation

[0194] Consider an application case where single-carrier frames are received in the receiving device 10, with four pilot sequences = 4) of KTS = 32 symbols each. A preliminary channel estimation is first performed with the LS algorithm on each of these four TS by the EST_CNL_0 block 11 (step 101).

[0195] Regarding the sparsity detection network in the EAST_MASK 12 block implemented in the application considered here (frames each with four pilot sequences of 32 symbols): - only the estimated LS if rn already are provided as input available, and an estimate of the variance of the channel noise; - the aggregation layer is implemented so as to have a single aspect, i.e. AL = 1, by calculating the quantity i P'2 for I 11--1 i k=l, ...,KTS; - Nc = 3 aspects for the intermediate residual layers, and the transition from 1 to 3 aspects is done with a separable convolutional layer (for lower computational cost) in one dimension, with a kernel of size 1, followed by a layer normalization of kernel (3, Kps) (i.e., normalization is performed by calculating the mean and variance over all elements of the 2nd order tensor of size 3 X Kpg); - the detection network includes N p = 1 residual layer consisting respectively of a ReLU (Rectified Linear Unit) nonlinearity, a one-dimensional separable convolutional layer with a kernel of size 3, a "mini-batch" normalization layer, another ReLU nonlinearity, a second one-dimensional separable convolutional layer with a kernel of size 3; - to finish the single-channel transition layer, is done with another one-dimensional separable convolutional layer, with a kernel of size 1.

[0196] This structure is illustrated in [Fig.9].

[0197] The refinement network considered in this application, with reference to [Fig. 10], includes, for example, 3-dimensional tensors to handle the representation of complex numbers and two-dimensional convolutions. Furthermore, the four channel estimates corresponding to the different TS sequences are associated with an aspect containing only the estimated variance of the channel noise, and an aspect that replicates the log-probabilistic estimates of the sparsity detection network, to which a sigmoid function is applied.

[0198] This input tensor, containing 6 aspects of size 2 X Kpg, is processed by a 2D separable convolution with a kernel size of (1,1) to increase the number of aspects to 8. This is followed by a ResNet consisting of ReLU nonlinearities, 2D separable convolutions with a kernel size of (3,2), and mini-batch normalization. Finally, a last layer of 2D separable convolution with a kernel size of (1,1) aggregates the 8 aspects into 4, attempting to correctly recover the estimates from the 4 pilot sequences.

[0199] Thus according to the invention, the neural network does not directly denoise the channel estimates, but determines the useful and not useful paths, by means of detecting the channel support by neural network, in the form of a binary mask and then denoising is performed according to this binary mask.

[0200] The invention proposed in this document thus performs additional processing following an initial channel estimation, for example using the LS algorithm, to bring the quality of the CIR estimates closer to the optimal CIR in the MSE sense. These processes are carried out through the use of neural networks within a hybrid artificial intelligence architecture (hybrid AI, referring to algorithmic solutions that inject data (such as neural networks) and business knowledge based on AI into AI solutions). (using traditional methods, or the occasional use of data-driven AI solutions). This architecture is designed to maintain the affordable channel estimation complexity while improving the LS channel estimation performance through a sparsity detection technique (or CIR support detection) followed by possible non-linear filtering.

[0201] Steps 101 to 105 of the process, or at least one or some of them, are implemented, for example, by executing software instructions on a processor. Alternatively, they are implemented by dedicated hardware, typically a digital integrated circuit, either specific (ASIC) or based on programmable logic (e.g., FPGA / Field Programmable Gate Array).

Claims

Demands

1. Method for improving the determination of the estimate of a telecommunication channel between a transmitting device and a receiving device (10), as a function of at least one signal frame comprising predefined pilot sequences, each of length KTS, transmitted via said channel to the receiving telecommunication device and a first estimate of the telecommunication channel having been made, following the reception of said frame by said receiving device (10), as a function of said received frame and said predefined pilot sequences and having provided at least first estimates of the channel impulse response r > i; with h'[n] = n], h2[nh ...,h Krs [n] °û h [il] corresponds to the path of the impulse response of the teletransmission channel relating to the nth pilot sequence of the frame; said process being characterized in that it comprises the following steps implemented by the receiving device: - providing input to a neural network of the receiving device with initial input data, and obtaining at the output of said network an estimate of a binary mask fhk. k = 1 to Kts, determined as a function of said initial input data provided, such that: I^=°, if the path with index k is determined to consist only of noise; 01^= 1, otherwise; said first input data being a function of the correlation between the received frame and said predefined pilot sequences; - determination of second estimates of the channel, r ' i Nts, as a function of the first estimates of the LJJ Z3=X channel, r • i ^rs, and of the determined binary mask, LLJJ n=l according to the following rule: î / 'r , ÂJn], simk=l .1 0, simk = 0 TS according to which the neural network has been previously trained during a learning phase using a set of the first input data of training frames such that, for first input data corresponding to a considered training frame, the desired output Hï^ of the neural network, to which is compared the output provided by the neural network in the learning phase, is fixed at: - 1 if and only if a predefined criterion Cfitest greater than a predefined threshold T - 0 otherwise; where the criterion Cl'it^ is a function of the realizations of the channel with little or no noise on the pilot sequences and at least one element, among the criterion C ritet the threshold Test determined as a function of the SNR or the SINR on said telecommunication channel for said considered training frame.

2. A method for improving the determination of the estimate of a telecommunications channel according to claim 1, wherein the first input data of the neural network comprises at least one of the following data sets relating to said frame: - the first channel estimates ri Nts, th[n]| ! LLJJ 22=1 - synchronization criterion(a) rrii NSYCh, characterizing the channel impulse response; - the first estimates of the channel r • ï NTS and the estimates of the {h [n] j . LLJJ 22=1 SINR through the noise variance fif2 relative to said frame or the covariance matrix Rw; - the first estimates of the channel r । NTS and the estimates of the {h [n] J . LJJ 21=1 SINR through the noise variances r ,2 r 11^$ or the covariance matrices । J- TS m^dves at each pilot sequence; - synchronization criterion(a) rr -1 i Nsych , characterizing the 1 vinj impulse response of the channel and the estimates of the SINR through the noise variance '2 or the covariance matrix

3. A method for improving the determination of the estimate of a telecommunication channel according to claim 1 or 2, wherein the threshold is chosen to be equal to where is the variance of the noise on the training frame considered.

4. A method for improving the determination of the estimate of a telecommunications channel according to any one of claims 1 to 2, wherein the threshold is chosen to be equal to p^, where is the estimate of the noise variance on the frame considered and Fc is a constant which, over a range of SNRs of interest for the profile of the channel considered, minimizes the root mean square error expressed in the form: minE «] h W ) ] 1 h where E is the statistical expectation, on stochastic realizations of the channel within the profile and the range of SNRs considered, and Â* [ U ] is a channel es^m^ obtained through denoising the previously estimated j with the ideal mask m* via h [12] = [22]

5. A method for improving the determination of the estimated value of a telecommunications channel according to any one of the preceding claims, wherein the criterion critk = n ] HRw[ n ] lh^ n ] , with R^fn] the covariance matrix of the noise affecting the thjth pilot sequence of the considered learning frame.

6. A method for improving the determination of the estimate of a telecommunication channel according to any one of claims 1 to 4, wherein the criterion Crit^ will be calculated from the formula , with RjJiî] the covariance matrix of the noise affecting the uth pilot sequence of the training frame considered; but replacing in said formula: Rw[n ] by the average value Rw of the covariance matrix of the noise affecting the training frame considered and / or the impulse responses of the channel by their time average over the frame.

7. A method for improving the determination of the estimate of a telecommunication channel according to any one of the preceding claims, wherein during the learning phase, the weights of the neural network are obtained as a function of a cost function based on crossed binary entropy.

8. A method for improving the determination of the estimate of a telecommunication channel according to any one of the preceding claims, wherein the neural network comprises: - one or more layers for aggregating the first input data to have N aspects of tensors of size IxKt^ - one or more layers for going from Njn aspects to Nc aspects of tensors of size 1X °where Nc is the number of intermediate aspects of the network; - NL ResNet type layers, and - one or more layers for going from Nc to NTS aspects of tensors of size 1XK^g.

9. A method for improving the determination of the estimate of a telecommunications channel according to any one of the preceding claims, comprising a step of: - providing, as input to another neural network of the receiving device, second input data from said other network neurons comprising at least the second estimated of the channel, ri Nts, and obtaining at the output of said other network of u J n=l neurons of third estimated of the channel, rri NTS; according to which said other neural network results from a learning phase of the other network aimed at minimizing a cost function, said cost function being a function of the mean squared error between the outputs obtained from said other network in the learning phase and the desired values ​​of the true impulse response of the channel, obtained by simulation.

10. A method for improving the determination of the estimate of a telecommunications channel according to claim 9, comprising at least one of the following provisions i, j: i / the second input data of said other neural network further comprise at least one element among probabilistic information of the channel support delivered at the output of the channel support estimation neural network and an estimate of the noise variance; j / said other neural network comprises: - one or more layers for going from aspects to Nc tensor aspects of size 2 X KT^, where Nc is the number of intermediate aspects of the network and greater than NTS; - Nl ResNet type layers, and - one or more layers for going from Nc to NTS tensor aspects of size 2 X KTS.

11. Telecommunications receiving device (10) adapted to determine the estimate of a telecommunications channel between said receiving device and a transmitting device, based on at least one signal frame comprising N-pg predefined pilot sequences, each of length KTS, transmitted via said channel to the telecommunications receiving device, wherein a first estimate of the telecommunications channel has been made, following the reception of said frame by said receiving device, based on said received frame and said predefined pilot sequences and having provided at least first estimates of the impulse response of channel r » 1 ^ts; with h'[n] = pjnhhjn].....h Krs [n] where j corresponds to the k'e!oe path of the impulse response of the teletransmission channel relative to the ijth pilot sequence of the frame; said receiving device (10) being characterized in that it comprises a neural network adapted to receive first input data as input, and to obtain at the output of said network an estimate of a binary mask fh^, k= 1 to K^g, determined as a function of said first input data provided, such that: ihk= 0, if the path of index k is determined as containing only noise; mk= 1, otherwise; said first input data being a function of the correlation between the received frame and said predefined pilot sequences; said receiving device (10) being adapted to determine estimated channel seconds, r NTS, as a function of the first estimated values ​​of the channel, r • i ^ts, and of the binary mask determined, according to the following rule: hk[n], simk=l 0, simk=0 in which the neural network has been previously trained during a learning phase using a set of initial input data from training frames such that, for initial input data corresponding to a given training frame, the desired output 1¾ of the neural network, to which the output provided by the neural network during the learning phase is compared, is fixed at: - 1 if and only if a predefined criterion crit^ is greater than a predefined threshold T - 0 otherwise; where the crit^ criterion is a function of the channel realizations with little or no noise on the pilot sequences and at least one element, among the crit^ criterion and the T& threshold, is determined as a function of the SNR or the SINR on said telecommunication channel for said training frame considered.

Images