Interference mitigation

Abstract

Disclosed is a method comprising receiving a data signal and one or more reference signals; obtaining a raw channel estimate for the received data signal based on the one or more reference signals; obtaining a first filtered channel estimate by denoising the raw channel estimate with a first filter; obtaining a set of disturbance vectors based at least on the first filtered channel estimate; estimating one or more covariance matrices based on the set of disturbance vectors; interpolating the first filtered channel estimate into a first channel matrix; inputting the first channel matrix and the one or more covariance matrices into an interference rejection combining equalizer configured to mitigate interference; and processing an output of the interference rejection combining equalizer with at least one machine learning model trained to detect signals, or with a combination of the at least one machine learning model and a pre-defined receiver.

Description

[0001]INTERFERENCE^MITIGATION^ FIELD The following example embodiments relate to wireless communication. BACKGROUND Interference in wireless communication occurs when unwanted signals disrupt the intended transmission. These disruptions can lead to temporary signal loss, degraded receiver performance, or compromised output quality in electronic devices. Mitigating interference is desirable because it enhances wireless network functionality and improves the reliability of wireless communication. SUMMARY The scope of protection sought for various example embodiments is set out by the claims. The example embodiments and features, if any, described in this specification that do not fall under the scope of the claims are to be interpreted as examples useful for understanding various embodiments. According to a first aspect, there is provided an apparatus comprising: means for receiving a data signal and one or more reference signals; means for obtaining a raw channel estimate for the received data signal based on the one or more reference signals; means for obtaining a first filtered channel estimate by denoising the raw channel estimate with a first filter; means for obtaining a set of disturbance vectors based at least on the first filtered channel estimate; means for estimating one or more covariance matrices based on the set of disturbance vectors; means for interpolating the first filtered channel estimate into a first channel matrix; means for inputting the first channel matrix and the one or more covariance matrices into an interference rejection combining equalizer configured to mitigate interference; and means for processing an output of the interference rejection combining equalizer with at least one machine learning model trained to detect signals, or with a combination of the at least one machine learning model and a pre-defined receiver. According to a second aspect, there is provided the apparatus of the first aspect, wherein the processing of the output of the interference rejection combining equalizer comprises: processing the output of the interference rejection combining equalizer with a deep-learning-based detector trained to detect signals; and processing an output of the deep-learning-based detector with a deep- learning-based demapper trained to demap the detected signals. According to a third aspect, there is provided the apparatus of the second aspect, further comprising: means for obtaining a second filtered channel estimate by denoising the raw channel estimate with a second filter different from the first filter; means for interpolating the second filtered channel estimate into a second channel matrix; and means for inputting the second channel matrix into another equalizer that is different from the interference rejection combining equalizer and utilized in parallel with the interference rejection combining equalizer, wherein the deep-learning-based detector is configured to combine the output of the interference rejection combining equalizer and an output of the other equalizer. According to a fourth aspect, there is provided the apparatus of the third aspect, wherein the other equalizer comprises a linear minimum mean square error equalizer. According to a fifth aspect, there is provided the apparatus of the third or fourth aspect, wherein the first filter comprises a trainable linear convolutional filter of a set of at least two trainable linear convolutional filters used for denoising the raw channel estimate, and wherein the second filter comprises another trainable linear convolutional filter of the set of at least two trainable linear convolutional filters. According to a sixth aspect, there is provided the apparatus of the third or fourth aspect, wherein the first filter and the second filter are pre-defined non- learned filters. According to a seventh aspect, there is provided the apparatus of the third or fourth aspect, wherein the first filter comprises a trainable linear convolutional filter, and wherein the second filter comprises a pre-defined non- learned filter. According to an eighth aspect, there is provided the apparatus of the third or fourth aspect, wherein the first filter comprises a pre-defined non-learned filter, and wherein the second filter comprises a trainable linear convolutional filter. According to a ninth aspect, there is provided the apparatus of any of the second to eighth aspects, further comprising: means for switching the deep- learning-based demapper to a non-deep-learning-based demapper. According to a tenth aspect, there is provided the apparatus of any of the first to ninth aspects, wherein the means for obtaining the set of disturbance vectors are configured to utilize an artificial neural network to obtain the set of disturbance vectors in a sliding window fashion. According to an eleventh aspect, there is provided the apparatus of the tenth aspect, wherein the artificial neural network is configured to determine the set of disturbance vectors based on the first filtered channel estimate, the received data signal, and the one or more reference signals. According to a twelfth aspect, there is provided the apparatus of the tenth or eleventh aspect, wherein the one or more covariance matrices estimated based on the set of disturbance vectors are used as a replacement of one or more interference-plus-noise covariance matrices in the interference rejection combining equalizer. According to a thirteenth aspect, there is provided the apparatus of any of the first to ninth aspects, wherein the means for obtaining the set of disturbance vectors are configured to utilize a pre-defined interference-plus-noise vector collection algorithm to obtain the set of disturbance vectors, wherein the one or more covariance matrices estimated based on the set of disturbance vectors comprise one or more interference-plus-noise covariance matrices. According to a fourteenth aspect, there is provided the apparatus of any of the first to thirteenth aspects, wherein the apparatus comprises a radio receiver. According to a fifteenth aspect, there is provided a method comprising: receiving a data signal and one or more reference signals; obtaining a raw channel estimate for the received data signal based on the one or more reference signals; obtaining a first filtered channel estimate by denoising the raw channel estimate with a first filter; obtaining a set of disturbance vectors based at least on the first filtered channel estimate; estimating one or more covariance matrices based on the set of disturbance vectors; interpolating the first filtered channel estimate into a first channel matrix; inputting the first channel matrix and the one or more covariance matrices into an interference rejection combining equalizer configured to mitigate interference; and processing an output of the interference rejection combining equalizer with at least one machine learning model trained to detect signals, or with a combination of the at least one machine learning model and a pre- defined receiver. According to a sixteenth aspect, there is provided a computer program comprising instructions which, when executed by an apparatus, cause the apparatus to perform at least the following: receiving a data signal and one or more reference signals; obtaining a raw channel estimate for the received data signal based on the one or more reference signals; obtaining a first filtered channel estimate by denoising the raw channel estimate with a first filter; obtaining a set of disturbance vectors based at least on the first filtered channel estimate; estimating one or more covariance matrices based on the set of disturbance vectors; interpolating the first filtered channel estimate into a first channel matrix; inputting the first channel matrix and the one or more covariance matrices into an interference rejection combining equalizer configured to mitigate interference; and processing an output of the interference rejection combining equalizer with at least one machine learning model trained to detect signals, or with a combination of the at least one machine learning model and a pre-defined receiver. According to a seventeenth aspect, there is provided a non-transitory computer readable medium comprising program instructions which, when executed by an apparatus, cause the apparatus to perform at least the following: receiving a data signal and one or more reference signals; obtaining a raw channel estimate for the received data signal based on the one or more reference signals; obtaining a first filtered channel estimate by denoising the raw channel estimate with a first filter; obtaining a set of disturbance vectors based at least on the first filtered channel estimate; estimating one or more covariance matrices based on the set of disturbance vectors; interpolating the first filtered channel estimate into a first channel matrix; inputting the first channel matrix and the one or more covariance matrices into an interference rejection combining equalizer configured to mitigate interference; and processing an output of the interference rejection combining equalizer with at least one machine learning model trained to detect signals, or with a combination of the at least one machine learning model and a pre- defined receiver. According to an eighteenth aspect, there is provided a computer readable medium comprising program instructions which, when executed by an apparatus, cause the apparatus to perform at least the following: receiving a data signal and one or more reference signals; obtaining a raw channel estimate for the received data signal based on the one or more reference signals; obtaining a first filtered channel estimate by denoising the raw channel estimate with a first filter; obtaining a set of disturbance vectors based at least on the first filtered channel estimate; estimating one or more covariance matrices based on the set of disturbance vectors; interpolating the first filtered channel estimate into a first channel matrix; inputting the first channel matrix and the one or more covariance matrices into an interference rejection combining equalizer configured to mitigate interference; and processing an output of the interference rejection combining equalizer with at least one machine learning model trained to detect signals, or with a combination of the at least one machine learning model and a pre-defined receiver. According to a nineteenth aspect, there is provided an apparatus comprising at least one processor, and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to: receive a data signal and one or more reference signals; obtain a raw channel estimate for the received data signal based on the one or more reference signals; obtain a first filtered channel estimate by denoising the raw channel estimate with a first filter; obtain a set of disturbance vectors based at least on the first filtered channel estimate; estimate one or more covariance matrices based on the set of disturbance vectors; interpolate the first filtered channel estimate into a first channel matrix; input the first channel matrix and the one or more covariance matrices into an interference rejection combining equalizer configured to mitigate interference; and process an output of the interference rejection combining equalizer with at least one machine learning model trained to detect signals, or with a combination of the at least one machine learning model and a pre-defined receiver. BRIEF DESCRIPTION OF THE DRAWINGS In the following, various example embodiments will be described in greater detail with reference to the accompanying drawings, in which FIG.1A illustrates an example embodiment of a radio receiver; FIG.1B illustrates an example embodiment of a radio receiver; FIG.1C illustrates an example embodiment of a radio receiver; FIG.2 illustrates a flow chart; FIG.3 illustrates a flow chart; FIG.4 illustrates a flow chart; FIG.5 illustrates a flow chart; FIG.6 illustrates an example of a wireless communication network; FIG.7 illustrates an example of an apparatus; FIG.8 illustrates an example of an artificial neural network; and FIG.9 illustrates an example of a computational node. DETAILED DESCRIPTION The following embodiments are exemplifying. Although the specification may refer to “an”, “one”, or “some” embodiment(s) in several locations of the text, this does not necessarily mean that each reference is made to the same embodiment(s), or that a particular feature only applies to a single embodiment. Single features of different embodiments may also be combined to provide other embodiments within the scope of the claims. Furthermore, the words "comprising" and "including" should be understood as not limiting the described embodiments to consist of only those features that have been mentioned, and such embodiments may also contain features that have not been specifically mentioned. Reference numbers, in the description and / or in the claims, serve to illustrate the embodiments with reference to the drawings, without limiting the embodiments to these examples only. Artificial intelligence (AI) or machine learning (ML) based solutions may be utilized for physical layer processing of radio systems. One such algorithm is called DeepRx, which refers to a convolutional neural network (CNN) based fully learned receiver. Simulation results indicate that such a fully learned receiver may achieve high performance even with considerably sparser demodulation reference signal (DMRS) patterns than required by legacy receivers. This is made possible by the fact that DeepRx learns to carry out data-aided detection, where it utilizes the unknown received data symbols to make the detection more accurate. The DeepRx receiver may also be used in multiple-input and multiple- output (MIMO) systems. Similar performance benefits may be obtained with spatial multiplexing by adapting the ML model architecture accordingly. However, currently, there are no DeepRx variants available that are tailored to scenarios with high levels of inter-cell interference. Inter-cell interference refers to the interference that occurs between neighboring cells in a cellular communication network. When neighboring cells reuse the same frequency channels, interference can occur between them. This interference affects the quality of communication for users in both cells. Mitigating inter-cell interference is desirable in order to provide better spectral efficiency, reliable communication, and improve network performance. The minimum mean square error interference rejection combining (MMSE-IRC) receiver, which is a non-ML based receiver design, is one possible option for mitigating inter-cell interference. The MMSE-IRC receiver is described in the following. For generating a transmitted signal, a radio transmitter may use a modulation scheme, such as orthogonal frequency division multiplexing (OFDM), where groups of bits may be mapped to complex-valued modulation symbols carried on subcarriers of an OFDM symbol. A subcarrier of an OFDM symbol is an example of a resource element (RE) used to carry data. In an OFDM transmission, the signal at a particular OFDM symbol and subcarrier index (i.e., resource element) may be defined as: ^= ^^ + ^ + ^ = ^^ + ^ ∈ ℂ^^ ,where ^^ is the number of receive antennas, ^ ∈ ℂ^^ × ^^ is the channel matrix, ^ ∈ℂ^^ is the transmitted data, ^ ∈ ℂ^^ is the ^ ∈ ℂ^^ is circular Gaussian noise, and ^ = ^ + ^ denotes the interference-plus-noise vector. Herein it isassumed that ^, ^, ^, and thus also ^ and ^ have zero mean. Furthermore, E^^^^^ =^, where ^ denotes the identity matrix. The linear MMSE-IRC equalizer may be givenby the minimizer of E^‖^ − ^ ^‖ ^ ^^^, where ^^ = ^ ^, yielding:^ = " ^ $%&^ ! ^^ + #^ ^where the interference-plus-noise covariance matrix (INCM) may be defined as: #^ = E^"^ + ^$"^ + ^$^^ = '^^^^^Since the channel and INCM are in practice unknown, they should be estimated. To this end, one or more DMRS pilot symbols carrying OFDM symbols may be transmitted, and raw channel estimates may be computed. Channelestimates ^( ∈ ℂ^^ ×^^ may then be obtained by smoothening and interpolating theraw channel estimates. From the smoothened channel estimates, the interference-plus-noise vectors may be obtained by ^() = ^) − ^( )^) , where ^) denotes the *thDMRS pilot symbol. The INCM, #^, may then be approximated by an interference- plus-noise sample covariance matrix (INSCM) estimate computed from theestimated interference-plus-noise vectors via + = &,∑,) / & ^()^(^ &) =,.(.( ^ , where .(is a matrix such that its *th column is ^()and 0 is plus- noise vectors over which the INSCM is computed. The INSCM may be computed on a resource block basis, where each resource block has its own estimate computed from the interference-plus-noise vectors in that particular resource block. To further reduce the variance of the INSCM, it is possible to compute the INSCM averaged over multiple resource blocks and / or over multiple (pilot-containing) OFDM symbols. Some example embodiments may provide a machine-learning-based receiver, which is capable of mitigating interference (e.g., inter-cell interference or any other type of interference) spatially, while also learning data-aided detection. The machine-learning-based receiver may be capable of mitigating interference from any source, without knowing any prior information about the interferer or the interference. The example embodiments described below may enable improved detection accuracy and / or reduced block error rate (BLER) under interference with just a modest increase in complexity compared to a non-ML interference rejection combining (IRC) receiver. Furthermore, some example embodiments may provide an option to switch to a simpler detector and demapper, for example if there is a need to reduce energy consumption, or even fall back to a non-ML IRC by using a non-ML demapper. FIG. 1A illustrates an example embodiment of an ML-based radio receiver 100 configured to mitigate interference spatially and learn data-aided detection. Referring to FIG. 1A, in block 101, a raw channel estimate (denoted as^( 1) is obtained for a received data signal (denoted as ^1) based on one or morereceived reference signals (denoted as ^1). The data signal may comprise, for example, voice, video, or any other type of data. The data may be arranged as an array of modulation symbols (e.g., real or complex-valued constellation points). In wireless communication, the channel refers to the medium through which wireless signals (e.g., radio waves) travel. The channel may introduce noise, distortion, and other effects. The received data signal may comprise a noisy version of the channel-transformed signal. The raw channel estimate refers to the initial channel estimate obtained directly from the received (noisy) data signal samples without any additional processing. The one or more reference signals may be used for comparison to help understand how the received data signal (after passing through the channel) differs from the original transmitted data signal. Herein a reference signal may also be referred to as a pilot. The one or more reference signals may comprise received symbols at predetermined reference resource elements (e.g., pilot subcarriers). The receiver 100 may be preconfigured with information about the transmitted reference signal(s). For example, the receiver 100 may determine the raw channel estimate based on multiplying the received one or more reference signals and expected values of the one or more reference signals. In block 102, at least two filtered channel estimates are obtained by denoising the raw channel estimate with a set of at least two filters configured to do at least two different smoothenings to the channel. In other words, a first filteredchannel estimate (denoted as ^( 213) is obtained by denoising the raw channelestimate with a first filter, and a second filtered channel estimate (denoted as ^( 21&)is obtained by denoising the raw channel estimate with a second filter different from the first filter. Since the raw channel estimate may contain noise due to environmental factors or imperfections in the receiver, applying denoising techniques can improve its accuracy. For example, the set of at least two filters may comprise a set of at least two trainable linear convolutional filters (i.e., ML-based filters) used for denoising the raw channel estimate. The first filter may comprise one trainable linear convolutional filter of the set of at least two trainable linear convolutional filters, and the second filter may comprise another trainable linear convolutional filter of the set of at least two trainable linear convolutional filters. The purpose of the trainable linear convolutional filters is to denoise or smoothen the raw channel estimate. This may be done by introducing two one- dimensional filters (denoted as 45,&and 45,^) in the frequency direction of the resource grid (one filter per equalizer 107, 109). The same filters may be applied for all antennas and layers. The length of the filters may be predetermined (e.g., up to 11 coefficients), however, the exact length can be considered as ahyperparameter. For * = 0, 1, the filtered (smoothened) channel estimates may begiven by: ^( 21,) = ^( 1 ∗ 9"45,)$ where ∗ denotes the convolution operation, and 9: 4 ↦ 9"4$ is a function appliedto 4. Herein 9"4) = |w| / ‖4‖&, where | ⋅ | denotes an element-wise absolute value,and ‖4‖& = ∑@ |?@| denotes the ℓ&-norm. Alternatively, other functions, such as theidentity function 9"4$ = 4 or 9"4$ = 4 / ‖4‖&, may be used. another example, the first filter and / or the second filter may be a pre-defined non-learned denoising filter (i.e., a non-machine-learning-based filter), such as a simple moving average or a weighted moving average. The simple moving average refers to a filter that computes the average of neighboring data points over a sliding window. The weighted moving average is similar to the simple moving average, but assigns different weights to each data point within the window. As another example, the first filter and / or the second filter may be learned deep neural networks, or even a single deep neural network that outputs one or two different filtered channel estimates. The deep neural network(s) may be fully connected neural network(s), convolutional neural network(s) or transformer(s), or any other artificial neural network architecture. Herein the term “transformer” refers to a type of deep learning model. Alternatively, any other suitable filter can be used. Block 103 depicts the inputs of an artificial neural network 104A (denoted as “DisturbNN” herein), wherein the inputs comprise the data (^1)corresponding to the pilot symbols and the matrix-vector product (^( 213^1)between the first filtered channel estimate and the pilot symbols. In block 104A, a set of disturbance vectors is obtained based on the inputs of block 103 by utilizing the DisturbNN, which is able to learn to output a disturbance matrix BC in a sliding window fashion based on the input terms 103,such that a disturbance covariance matrix #( = BCBC^ (i.e., a covariance matrixcomputed from the disturbance matrix BC) can be effectively used in place of the interference-plus-noise covariance matrix (INCM) in the IRC equalization block 107. For example, the DisturbNN 104A may comprise a convolutional neural network, or a transformer, or a fully connected neural network (FCNN). In other words, the disturbance matrix BCis the output matrix of the DisturbNN 104A, and the disturbance vectors refer to the columns of the disturbance matrix. Herein the disturbance refers to interference plus noise. The disturbance matrix contains features of the interference and noise, so that thedisturbance covariance matrix #( = BCBC^ can be used as the interference-plus-noisecovariance matrix in the IRC equalization. The purpose of the DisturbNN is to imitate or improve the computation of the interference-plus-noise vectors in a convolutional sliding window fashion. For example, a complex-valued CNN may be used to compute: D( = EFF"GH 1, G1$where the output D( is a tensor comprising a complex-valued disturbance matrix BCof size ^^ × I for each subcarrier. The input tensor G1 contains the received dataat the pilot locations, and the input tensor GH 1 is such that it contains the matrix-vector products ^( 213^1 for each pilot symbol, and thus, estimates the portion ofthe received signal contributed by the transmitted pilots. In block 105, one or more disturbance covariance matrices are estimated based on the set of disturbance vectors outputted by the DisturbNN 104A. A distinct disturbance covariance matrix may be estimated for a given resource element, or a subcarrier, or a set of subcarriers (e.g., one disturbance covariance matrix for each physical resource block, or one disturbance covariance matrix for a set of physical resource blocks). In other words, multiple disturbance covariance matrices may be estimated for the whole transmission interval or slot, wherein each disturbance covariance matrix may be specific to a certain resource element, or subcarrier, or set of subcarriers (e.g., physical resource blocks). For each resource element or subcarrier or set of subcarriers, the disturbance covariance matrix may be computed as: #( = BCBC^where BCBC^means that the disturbance matrix BCis multiplied by its own conjugate transpose (Hermitian transpose). In other words, herein the disturbance covariance matrix is a Hermitian positive semidefinite matrix. This disturbance covariance matrix #( is used in place of the INCM in theIRC equalizer 107. In other words, the one or more disturbance covariance matrices estimated based on the set of disturbance vectors are used as a replacement or estimate of one or more interference-plus-noise covariance matrices in the interference rejection combining equalizer 107. The DisturbNN 104A can thus be thought of estimating a set of interference-plus-noise-like vectors or disturbance vectors (i.e., columns of the disturbance matrix BC) in an efficient manner. However, it should be noted that the number of columns I in BC can be different from the number of interference-plus- noise vectors used to compute the INSCM in an MMSE-IRC receiver, and thus I can be considered as a hyperparameter. Furthermore, due to the flexibility of the setup,it is possible to, for example, select the size of the output D( (e.g., by usingsubsampling or CNN strides) so that it learns only one BC per resource block as in legacy IRC, thus reducing the overall complexity. In block 106, the first filtered channel estimate is interpolated into afirst channel matrix (denoted as ^( 3), for example by using nearest-neighborinterpolation. Interpolation involves estimating values between known data points. In this context, the interpolation means that the filtered channel estimate is used to fill in missing or intermediate values. The channel matrix represents the entire channel response, including multiple paths, delays, and gains. By interpolating the filtered estimate into the channel matrix, a more complete picture of the channel characteristics can be created. In other words, the channel matrix models the radio channel that the data signal propagates through. In block 107, the first channel matrix ^( 3 and the one or moredisturbance covariance matrices #( are inputted into an interference rejectioncombining (IRC) equalizer configured to mitigate interference. After obtaining the disturbance matrix BC from the DisturbNN 104A, theIRC block 107 computes, for each resource element, the data symbol estimates: ^^^ ! = J( ^^! K where J( ^ ! = L^( 3^(^3 + #( + M^N%& ^( 3, where ^( 3 denotes the first channel matrix corresponding to the resource elementafter interpolation. The parameter M can be chosen to a small positive value to ensure invertibility, or it can be set to zero. ^ denotes the identity matrix. Thedisturbance covariance matrix #( = BCBC^ in the above equalizer equation may beunique for a particular resource element, or subcarrier, or set of subcarriers (e.g., physical resource blocks). It should be noted that the above formulation of the IRC equalizer is just one example, and there may be alternative formulations of the IRC equalizer and potential approximations or optimizations to the IRC, which can alternatively be used. In block 108, the second filtered channel estimate is interpolated into asecond channel matrix (denoted as ^( &).In block 109, the channel matrix is inputted into another equalizer that is different from the interference rejection combining equalizer 107 and utilized in parallel with the interference rejection combining equalizer 107. The received data signal may also be provided as an input to the other equalizer 109. For example, the other equalizer 109 may comprise a linear minimum mean square error (LMMSE) equalizer. Block 109 may also be referred to as an LMMSE equalizer shortcut. The LMMSE equalizer shortcut may be computed byinterpolating the second filtered channel estimate ^( 21& to the full resource grid.Then, for each resource element, LMMSE equalization may be applied by computing: ^^ (^( ^U%&(^OPPQR = L^ &^ & + σT ^N ^ & ^where ^( & denotes to the resourceelement after the interpolation 108, and σ^UTis the noise variance or its estimate. When this is a part of an ML for example, be estimated via the INSCMas VC^T = "Tr"+$ / ^^$^, or some based on the estimated signal-to-noise ratio, or set to a small fixed value. In block 110, the output of the IRC equalizer 107 and the output of the other equalizer (e.g., LMMSE equalizer) 109 are processed with (inputted to) a deep-learning-based detector 110 (denoted as “DetectorNN” herein) trained to detect signals. In other words, the deep-learning-based detector 110 is configured to combine the output of the IRC equalizer 107 and the output of the other equalizer 109. In addition to the outputs (i.e., constellation symbol estimates) of the equalizers 107, 109, the input of the deep-learning-based detector 110 may also include modulation information (mod_info) and pilot density information (pilot_density). The modulation information may indicate, for example, how many bits there are per symbol. The pilot density information may indicate how many pilots there are in the transmission time interval (TTI), for example. The purpose of the deep-learning-based detector 110 is to preprocess the data for a deep- learning-based demapper 111 (denoted as “DemapperNN” herein). The deep- learning-based detector 110 may comprise, for example, a convolutional neural network, or a transformer, or a fully connected neural network (different from the DisturbNN 104A and the DemapperNN 111). In block 111, an output of the deep-learning-based detector 110 is processed with (inputted to) the deep-learning-based demapper 111 trained to demap the detected signals. In addition to the output of the deep-learning-based detector 110, the input of the deep-learning-based demapper 111 may also include the modulation information and the pilot density information. The output of the deep-learning-based demapper 111 comprises bit log-likelihood ratios (LLRs) of the bits transmitted in the received data signal. Bit LLRs represent the confidence levels for each bit in the received data signal, aiding error correction decoding in wireless communication systems, allowing the recovery of information from noisy communication channels. The deep-learning-based demapper 111 may comprise, for example, a convolutional neural network, or a transformer, or a fully connected neural network (different from the DisturbNN 104A and the DetectorNN 110). In other words, the receiver 100 utilizes two parallel equalizer blocks 107, 109, one of which implements IRC-like equalization, while the other one may implement LMMSE equalization, for example. The outputs of these two parallel equalizer blocks 107, 109 are then combined by the trained DetectorNN 110 and fed to the trained DemapperNN 111. Thus, the receiver 100 may be considered as one large deep learning model. The trained DetectorNN 110 and the trained DemapperNN 111 together perform data-aided detection. The model (e.g., the DisturbNN 104A, the DetectorNN 110, and the DemapperNN 111) may be trained end-to-end using supervised training and stochastic gradient descent. The frequency-domain antenna data and raw channel estimates are the input and bit log-likelihood ratios are the outputs. The outputs may be included in the loss function for training the neural network(s) of the reception chain of the radio receiver 100. One example of such a loss term may comprise the binary cross-entropy between the transmitted and received bits (e.g., LLRs). The ML receiver 100 may be constructed in such a way that it can be simplified to a non-ML IRC receiver by switching the DemapperNN 111 to a non- ML-based demapper. FIG.1B illustrates another example embodiment of an ML-based radio receiver 100 configured to mitigate interference spatially and learn data-aided detection. A difference between FIG. 1A and FIG. 1B is that, in FIG. 1B, the DisturbNN 104A of FIG. 1A is replaced with a non-machine-learning-based algorithm 104B used for collecting the set of disturbance vectors (i.e., interference- plus-noise vectors in this case). In other words, in the example embodiment of FIG. 1B, a pre-defined interference-plus-noise vector collection algorithm 104B is utilized to obtain the set of disturbance vectors (interference-plus-noise vectors). In this example embodiment, the one or more covariance matrices estimated in block 105 based on the set of disturbance vectors (interference-plus-noise vectors) comprise one or more interference-plus-noise covariance matrices (i.e., the #^matrix is estimated using the INSCM). FIG.1C illustrates another example embodiment of an ML-based radio receiver 100 configured to mitigate interference spatially and learn data-aided detection. Referring to FIG.1C, in block 101, a raw channel estimate is obtained for a received data signal based on one or more received reference signals. Herein a reference signal may also be referred to as a pilot. In block 102C, a first filtered channel estimate is obtained by denoising the raw channel estimate with at least a first filter (i.e., with at least one filter). For example, the first filter 102C may comprise a trainable linear convolutional filter, or a pre-defined non-learned filter. In block 104C, a set of disturbance vectors is obtained based at least on the first filtered channel estimate. For example, an artificial neural network may be used to obtain the set of disturbance vectors in a sliding window fashion. The artificial neural network may be configured to determine the set of disturbance vectors based on the first filtered channel estimate, the received data signal, and the one or more reference signals. The artificial neural network may comprise, for example, a convolutional neural network, or a transformer, or a fully connected neural network. The one or more covariance matrices estimated based on the set of disturbance vectors may be used as a replacement of one or more interference-plus-noise covariance matrices in the interference rejection combining equalizer. As another example, a pre-defined interference-plus-noise vector collection algorithm may be used to obtain the set of disturbance vectors, in which case the one or more covariance matrices estimated based on the set of disturbance vectors may comprise one or more interference-plus-noise covariance matrices, wherein each interference-plus-noise covariance matrix may correspond to a different resource element or subcarrier or set of subcarriers. In block 105, one or more covariance matrices are estimated based on the set of disturbance vectors. In block 106, the first filtered channel estimate is interpolated into a first channel matrix. In block 107, the first channel matrix and the one or more covariance matrices are inputted into an interference rejection combining equalizer configured to mitigate interference. In block 110C, an output of the interference rejection combining equalizer is processed with at least one machine learning model 110C trained to detect signals, or with a combination of the at least one machine learning model 110C and a pre-defined receiver. For example, the at least one machine learning model 110C may comprise the deep-learning-based detector 110 and / or the deep- learning-based demapper 111. FIG.2 illustrates a flow chart according to an example embodiment of a method for mitigating interference. The method may be performed by an apparatus 700 depicted in FIG.7. The apparatus 700 may comprise or include the radio receiver 100. For example, the apparatus 700 may be a wireless communication device such as a user equipment or an access node of a wireless communication network. Referring to FIG.2, in block 201, a data signal and one or more reference signals are received. Herein a reference signal may also be referred to as a pilot. In block 202, a raw channel estimate is obtained for the received data signal based on the one or more reference signals. In block 203, a first filtered channel estimate is obtained by denoising the raw channel estimate with a first filter. In block 204, a set of disturbance vectors is obtained based at least on the first filtered channel estimate. As an example, an artificial neural network may be used to obtain the set of disturbance vectors in a sliding window fashion. The artificial neural network may be configured to determine the set of disturbance vectors based on the first filtered channel estimate, the received data signal, and the one or more reference signals. The artificial neural network may comprise, for example, a convolutional neural network, or a transformer, or a fully connected neural network. The one or more covariance matrices estimated based on the set of disturbance vectors may be used as a replacement of one or more interference-plus-noise covariance matrices in the interference rejection combining equalizer. As another example, a pre-defined interference-plus-noise vector collection algorithm may be used to obtain the set of disturbance vectors, in which case the one or more covariance matrices estimated based on the set of disturbance vectors may comprise one or more interference-plus-noise covariance matrices. In block 205, one or more covariance matrices are estimated based on the set of disturbance vectors. In block 206, the first filtered channel estimate is interpolated into a first channel matrix. In block 207, the first channel matrix and the one or more covariance matrices are inputted into an interference rejection combining equalizer configured to mitigate interference. In block 208, an output of the interference rejection combining equalizer is processed with at least one machine learning model trained to detect signals, or with a combination of the at least one machine learning model and a pre- defined receiver. FIG.3 illustrates a flow chart according to an example embodiment of a method for mitigating interference. The method may be performed by an apparatus 700 depicted in FIG.7. The apparatus 700 may comprise or include the radio receiver 100. For example, the apparatus 700 may be a wireless communication device such as a user equipment or an access node of a wireless communication network. Referring to FIG.3, in block 301, a data signal and one or more reference signals are received. Herein a reference signal may also be referred to as a pilot. In block 302, a raw channel estimate is obtained for the received data signal based on the one or more reference signals. In block 303, a first filtered channel estimate is obtained by denoising the raw channel estimate with a first filter. In block 304, a set of disturbance vectors is obtained based at least on the first filtered channel estimate. For example, an artificial neural network may be used to obtain the set of disturbance vectors in a sliding window fashion. The artificial neural network may be configured to determine the set of disturbance vectors based on the first filtered channel estimate, the received data signal, and the one or more reference signals. The artificial neural network may comprise, for example, a convolutional neural network, or a transformer, or a fully connected neural network. The one or more covariance matrices estimated based on the set of disturbance vectors may be used as a replacement of one or more interference-plus-noise covariance matrices in the interference rejection combining equalizer. As another example, a pre-defined interference-plus-noise vector collection algorithm may be used to obtain the set of disturbance vectors, in which case the one or more covariance matrices estimated based on the set of disturbance vectors may comprise one or more interference-plus-noise covariance matrices. In block 305, one or more covariance matrices are estimated based on the set of disturbance vectors. In block 306, the first filtered channel estimate is interpolated into a first channel matrix. In block 307, the first channel matrix and the one or more covariance matrices are inputted into an interference rejection combining equalizer configured to mitigate interference. In block 308, an output of the interference rejection combining equalizer is processed with a deep-learning-based detector trained to detect signals. In block 309, an output of the deep-learning-based detector is processed with a deep-learning-based demapper trained to demap the detected signals. FIG.4 illustrates a flow chart according to an example embodiment of a method for mitigating interference. The method may be performed by an apparatus 700 depicted in FIG.7. The apparatus 700 may comprise or include the radio receiver 100. For example, the apparatus 700 may be a wireless communication device such as a user equipment or an access node of a wireless communication network. Referring to FIG. 4, in block 401, a deep-learning-based demapper is switched to a non-deep-learning-based demapper (or to a non-machine-learning- based demapper). The switching may be done at runtime, for example by using an artificial neural network trained to determine when to switch between the deep- learning-based demapper and the non-deep-learning-based demapper. Alternatively, the deep-learning-based demapper may be switched to the non- deep-learning-based demapper (or the non-machine-learning-based demapper) before runtime, for example already when designing the receiver and its hardware. In block 402, a data signal and one or more reference signals are received. Herein a reference signal may also be referred to as a pilot. In block 403, a raw channel estimate is obtained for the received data signal based on the one or more reference signals. In block 404, a first filtered channel estimate is obtained by denoising the raw channel estimate with a first filter. In block 405, a set of disturbance vectors is obtained based at least on the first filtered channel estimate. For example, an artificial neural network may be used to obtain the set of disturbance vectors in a sliding window fashion. The artificial neural network may be configured to determine the set of disturbance vectors based on the first filtered channel estimate, the received data signal, and the one or more reference signals. The artificial neural network may comprise, for example, a convolutional neural network, or a transformer, or a fully connected neural network. The one or more covariance matrices estimated based on the set of disturbance vectors may be used as a replacement of one or more interference-plus-noise covariance matrices in the interference rejection combining equalizer. As another example, a pre-defined interference-plus-noise vector collection algorithm may be used to obtain the set of disturbance vectors, in which case the one or more covariance matrices estimated based on the set of disturbance vectors may comprise one or more interference-plus-noise covariance matrices. In block 406, one or more covariance matrices are estimated based on the set of disturbance vectors. In block 407, the first filtered channel estimate is interpolated into a first channel matrix. In block 408, the first channel matrix and the one or more covariance matrices are inputted into an interference rejection combining equalizer configured to mitigate interference. In block 409, an output of the interference rejection combining equalizer is processed with a deep-learning-based detector trained to detect signals. In block 410, an output of the deep-learning-based detector is processed with the non-deep-learning-based demapper (or the non-machine- learning-based demapper). FIG.5 illustrates a flow chart according to an example embodiment of a method for mitigating interference. The method may be performed by an apparatus 700 depicted in FIG.7. The apparatus 700 may comprise or include the radio receiver 100. For example, the apparatus 700 may be a wireless communication device such as a user equipment or an access node of a wireless communication network. Referring to FIG.5, in block 501, a data signal and one or more reference signals are received. Herein a reference signal may also be referred to as a pilot. In block 502, a raw channel estimate is obtained for the received data signal based on the one or more reference signals. In block 503, a first filtered channel estimate is obtained by denoising the raw channel estimate with a first filter. In block 504, a set of disturbance vectors is obtained based at least on the first filtered channel estimate. As an example, an artificial neural network may be used to obtain the set of disturbance vectors in a sliding window fashion. The artificial neural network may be configured to determine the set of disturbance vectors based on the first filtered channel estimate, the received data signal, and the one or more reference signals. The artificial neural network may comprise, for example, a convolutional neural network, or a transformer, or a fully connected neural network. The one or more covariance matrices estimated based on the set of disturbance vectors may be used as a replacement of one or more interference-plus-noise covariance matrices in the interference rejection combining equalizer. As another example, a pre-defined interference-plus-noise vector collection algorithm may be used to obtain the set of disturbance vectors, in which case the one or more covariance matrices estimated based on the set of disturbance vectors may comprise one or more interference-plus-noise covariance matrices. In block 505, one or more covariance matrices are estimated based on the set of disturbance vectors. In block 506, the first filtered channel estimate is interpolated into a first channel matrix. In block 507, the first channel matrix and the one or more covariance matrices are inputted into an interference rejection combining equalizer configured to mitigate interference. In block 508, a second filtered channel estimate is obtained by denoising the raw channel estimate with a second filter different from the first filter. For example, the other equalizer may comprise a linear minimum mean square error equalizer, or any other suitable equalizer. As an example, the first filter may comprise a trainable linear convolutional filter of a set of at least two trainable linear convolutional filters used for denoising the raw channel estimate, and the second filter may comprises another trainable linear convolutional filter of the set of at least two trainable linear convolutional filters. Herein the term “trainable linear convolutional filter” may refer to a machine-learning-based linear convolutional filter. As another example, the first filter and the second filter may be pre- defined non-learned filters (i.e., non-machine-learning-based filters). As another example, the first filter may comprises a trainable linear convolutional filter, and the second filter may comprise a pre-defined non-learned filter. As another example, the first filter may comprise a pre-defined non- learned filter, and the second filter may comprise a trainable linear convolutional filter. In block 509, the second filtered channel estimate is interpolated into a second channel matrix. In block 510, the second channel matrix is inputted into another equalizer that is different from the interference rejection combining equalizer and utilized in parallel with the interference rejection combining equalizer. The other equalizer may also be configured to mitigate interference, but differently from the interference rejection combining equalizer. In block 511, an output of the interference rejection combining equalizer and an output of the other equalizer are processed with a deep-learning- based detector trained to detect signals. The deep-learning-based detector may be configured to combine the output of the interference rejection combining equalizer and the output of the other equalizer. In block 512, an output of the deep-learning-based detector is processed with a deep-learning-based demapper trained to demap the detected signals. The blocks, related functions, and information exchanges (messages) described above by means of FIGS. 2 to 5 are in no absolute chronological order, and some of them may be performed simultaneously or in an order differing from the described one. Other functions can also be executed between them or within them, and other information may be sent, and / or other rules applied. Some of the blocks or part of the blocks or one or more pieces of information can also be left out or replaced by a corresponding block or part of the block or one or more pieces of information. Herein the terms “first filtered channel estimate” and “second filtered channel estimate” are used to distinguish these two filtered channel estimates, and they do not necessarily mean a specific order of the filtered channel estimates. Similarly, the terms “first channel matrix” and “second channel matrix” are used to distinguish these two different channel matrices, and they do not necessarily mean a specific order of the channel matrices. Similarly, the terms “first filter” and “second filter” are used to distinguish these two different filters, and they do not necessarily mean a specific order of the filters. As used herein, “at least one of the following: ” and “at least one of ” and similar wording, where the list of two or more elements are joined by “and” or “or”, mean at least any one of the elements, or at least any two or more of the elements, or at least all the elements. Some example embodiments described herein may be implemented in a wireless communication network comprising a radio access network based on one or more of the following radio access technologies (RATs): global system for mobile communications (GSM) or any other second generation (2G) radio access technology, universal mobile telecommunication system (UMTS, 3G) based on basic wideband-code division multiple access (W-CDMA), high-speed packet access (HSPA), long term evolution (LTE), LTE-Advanced, fourth generation (4G), fifth generation (5G), 5G new radio (NR), 5G-Advanced (i.e., 3GPP NR Rel-18 and beyond), or sixth generation (6G). Some examples of radio access networks include the universal mobile telecommunications system (UMTS) radio access network (UTRAN), the evolved universal terrestrial radio access network (E-UTRA), or the next generation radio access network (NG-RAN). The wireless communication network may further comprise a core network, and some example embodiments may also be applied to network functions of the core network. It should be noted that the embodiments are not restricted to the wireless communication network given as an example, but a person skilled in the art may also apply the solution to other wireless communication networks or systems provided with necessary properties. For example, some example embodiments may also be applied to a communication system based on Institute of Electrical and Electronics Engineers (IEEE) 802.11 specifications, or a communication system based on IEEE 802.15 specifications. FIG. 6 depicts an example of a simplified wireless communication network showing some physical and logical entities. The connections shown in FIG. 6 may be physical connections or logical connections. It is apparent to a person skilled in the art that the wireless communication network may also comprise other physical and logical entities than those shown in FIG.6. The example embodiments described herein are not, however, restricted to the wireless communication network given as an example but a person skilled in the art may apply the example embodiments described herein to other wireless communication networks provided with necessary properties. The example wireless communication network shown in FIG.6 includes a radio access network (RAN) and a core network 610. FIG. 6 shows user equipment (UE) 600, 602 configured to be in a wireless connection on one or more communication channels in a radio cell with an access node 604 of a radio access network. The access node 604 may comprise a computing device configured to control the radio resources of the access node 604 and to be in a wireless connection with one or more UEs 600, 602. The access node 604 may also be referred to as a base station, a base transceiver station (BTS), an access point, a cell site, a network node, a radio access network node, or a RAN node. The access node 604 may be, for example, an evolved NodeB (abbreviated as eNB or eNodeB), or a next generation evolved NodeB (abbreviated as ng-eNB), or a next generation NodeB (abbreviated as gNB or gNodeB), providing the radio cell. The access node 604 may include or be coupled to transceivers. From the transceivers of the access node 604, a connection may be provided to an antenna unit that establishes a bi-directional radio link to one or more UEs 600, 602. The antenna unit may comprise an antenna or antenna element, or a plurality of antennas or antenna elements. The wireless connection (e.g., radio link) from a UE 600, 602 to the access node 604 may be called uplink (UL) or reverse link, and the wireless connection (e.g., radio link) from the access node 604 to the UE 600, 602 may be called downlink (DL) or forward link. A UE 600 may also communicate directly with another UE 602, and vice versa, via a wireless connection generally referred to as a sidelink (SL). It should be appreciated that the access node 604 or its functionalities may be implemented by using any node, host, server, access point or other entity suitable for providing such functionalities. The radio access network may comprise more than one access node 604, in which case the access nodes may also be configured to communicate with one another over wired or wireless links. These links between access nodes may be used for sending and receiving control plane signaling and also for routing data from one access node to another access node. The access node 604 may further be connected to a core network (CN) 610. The core network 610 may comprise an evolved packet core (EPC) network and / or a 5thgeneration core network (5GC). The EPC may comprise network entities, such as a serving gateway (S-GW for routing and forwarding data packets), a packet data network gateway (P-GW) for providing connectivity of UEs to external packet data networks, and / or a mobility management entity (MME). The 5GC may comprise one or more network functions, such as at least one of: a user plane function (UPF), an access and mobility management function (AMF), a location management function (LMF), and / or a session management function (SMF). The core network 610 may also be able to communicate with one or more external networks 613, such as a public switched telephone network or the Internet, or utilize services provided by them. For example, in 5G wireless communication networks, the UPF of the core network 610 may be configured to communicate with an external data network via an N6 interface. In LTE wireless communication networks, the P-GW of the core network 610 may be configured to communicate with an external data network. It should also be understood that the distribution of functions between core network operations and access node operations may differ in future wireless communication networks compared to that of the LTE or 5G, or even be non- existent. The illustrated UE 600, 602 is one type of an apparatus to which resources on the air interface may be allocated and assigned. The UE 600, 602 may also be called a wireless communication device, a subscriber unit, a mobile station, a remote terminal, an access terminal, a user terminal, a terminal device, or a user device, just to mention but a few names. The UE 600, 602 may be a computing device operating with or without a subscriber identification module (SIM), including, but not limited to, the following types of computing devices: a mobile phone, a smartphone, a personal digital assistant (PDA), a handset, a computing device comprising a wireless modem (e.g., an alarm or measurement device, etc.), a laptop computer, a desktop computer, a tablet, a game console, a notebook, a multimedia device, a reduced capability (RedCap) device, a wearable device (e.g., a watch, earphones or eyeglasses) with radio parts, a sensor comprising a wireless modem, or a computing device comprising a wireless modem integrated in a vehicle. It should be appreciated that the UE 600, 602 may also be a nearly exclusive uplink-only device, of which an example may be a camera or video camera loading images or video clips to a network. The UE 600, 602 may also be a device having capability to operate in an Internet of Things (IoT) network, which is a scenario in which objects may be provided with the ability to transfer data over a network without requiring human-to-human or human-to-computer interaction. The wireless communication network may also be able to support the usage of cloud services. For example, at least part of core network operations may be carried out as a cloud service (this is depicted in FIG.6 by “cloud” 614). The UE 600, 602 may also utilize the cloud 614. In some applications, the computation for a given UE may be carried out in the cloud 614 or in another UE. The wireless communication network may also comprise a central control entity, such as a network management system (NMS), or the like. The NMS is a centralized suite of software and hardware used to monitor, control, and administer the network infrastructure. The NMS is responsible for a wide range of tasks such as fault management, configuration management, security management, performance management, and accounting management. The NMS enables network operators to efficiently manage and optimize network resources, ensuring that the network delivers high performance, reliability, and security. 5G enables using multiple-input and multiple-output (MIMO) antennas in the access node 604 and / or the UE 600, 602, many more base stations or access nodes than an LTE network (a so-called small cell concept), including macro sites operating in co-operation with smaller stations and employing a variety of radio technologies depending on service needs, use cases and / or spectrum available.5G wireless communication networks may support a wide range of use cases and related applications including video streaming, augmented reality, different ways of data sharing and various forms of machine-type applications, such as (massive) machine-type communications (mMTC), including vehicular safety, different sensors and real-time control. In 5G wireless communication networks, access nodes and / or UEs may have multiple radio interfaces, such as below 6 gigahertz (GHz), centimeter wave (cmWave) and millimeter wave (mmWave), and also being integrable with legacy radio access technologies, such as LTE. Integration with LTE may be implemented, for example, as a system, where macro coverage may be provided by LTE, and 5G radio interface access may come from small cells by aggregation to LTE. In other words, a 5G wireless communication network may support both inter-RAT operability (such as interoperability between LTE and 5G) and inter-RI operability (inter-radio interface operability, such as between below 6GHz, cmWave, and mmWave). 5G wireless communication networks may also apply network slicing, in which multiple independent and dedicated virtual sub-networks (network instances) may be created within the same physical infrastructure to run services that have different requirements on latency, reliability, throughput and mobility. In one embodiment, an access node 604 may comprise: a radio unit (RU) 603 comprising a radio transceiver (TRX), i.e., a transmitter (Tx) and a receiver (Rx); one or more distributed units (DUs) 605 that may be used for the so-called Layer 1 (L1) processing and real-time Layer 2 (L2) processing; and a central unit (CU) 608 (also known as a centralized unit) that may be used for non-real-time L2 and Layer 3 (L3) processing. The CU 608 may be connected to the one or more DUs 605 for example via an F1 interface. Such an embodiment of the access node 604 may enable the centralization of CUs relative to the cell sites and DUs, whereas DUs may be more distributed and may even remain at cell sites. The CU and DU together may also be referred to as baseband or a baseband unit (BBU). The CU and DU may also be comprised in a radio access point (RAP). The CU 608 may be a logical node hosting radio resource control (RRC), service data adaptation protocol (SDAP) and / or packet data convergence protocol (PDCP), of the NR protocol stack for an access node 604. The CU 608 may comprise a control plane (CU-CP), which may be a logical node hosting the RRC and the control plane part of the PDCP protocol of the NR protocol stack for the access node 604. The CU 608 may further comprise a user plane (CU-UP), which may be a logical node hosting the user plane part of the PDCP protocol and the SDAP protocol of the CU for the access node 604. The DU 605 may be a logical node hosting radio link control (RLC), medium access control (MAC) and / or physical (PHY) layers of the NR protocol stack for the access node 604. The operations of the DU 605 may be at least partly controlled by the CU 608. It should also be understood that the distribution of functions between the DU 605 and the CU 608 may vary depending on the implementation. Cloud computing systems may also be used to provide the CU 608 and / or DU 605. A CU provided by a cloud computing system may be referred to as a virtualized CU (vCU). In addition to the vCU, there may also be a virtualized DU (vDU) provided by a cloud computing system. Furthermore, there may also be a combination, where the DU may be implemented on so-called bare metal solutions, for example application-specific integrated circuit (ASIC) or customer-specific standard product (CSSP) system-on-a-chip (SoC). Edge cloud may be brought into the radio access network by utilizing network function virtualization (NFV) and software defined networking (SDN). Using edge cloud may mean access node operations to be carried out, at least partly, in a computing system operationally coupled to a remote radio head (RRH) or a radio unit (RU) 603 of an access node 604. It is also possible that access node operations may be performed on a distributed computing system or a cloud computing system located at the access node 604. Application of cloud RAN architecture enables RAN real-time functions being carried out at the radio access network (e.g., in a DU 605), and non-real-time functions being carried out in a centralized manner (e.g., in a CU 608). 5G (or new radio, NR) wireless communication networks may support multiple hierarchies, where multi-access edge computing (MEC) servers may be placed between the core network 610 and the access node 604. It should be appreciated that MEC may be applied in LTE wireless communication networks as well. A 5G wireless communication network (“5G network”) may also comprise a non-terrestrial communication network, such as a satellite communication network, to enhance or complement the coverage of the 5G radio access network. For example, satellite communication may support the transfer of data between the 5G radio access network and the core network 610, enabling more extensive network coverage. Possible use cases may include: providing service continuity for machine-to-machine (M2M) or Internet of Things (IoT) devices or for passengers on board of vehicles, or ensuring service availability for critical communications, and future railway, maritime, or aeronautical communications. Satellite communication may utilize geostationary earth orbit (GEO) satellite systems, but also low earth orbit (LEO) satellite systems, in particular mega-constellations (i.e., systems in which hundreds of (nano)satellites are deployed). A given satellite 606 in the mega-constellation may cover several satellite-enabled network entities that create on-ground cells. The on-ground cells may be created through an on-ground relay access node or by an access node located on-ground or in a satellite. It is obvious for a person skilled in the art that the access node 604 depicted in FIG. 6 is just an example of a part of a radio access network, and in practice the radio access network may comprise a plurality of access nodes 604, the UEs 600, 602 may have access to a plurality of radio cells, and the radio access network may also comprise other apparatuses, such as physical layer relay access nodes or other entities. At least one of the access nodes may be a Home eNodeB or a Home gNodeB. A Home gNodeB or a Home eNodeB is a type of access node that may be used to provide indoor coverage inside a home, office, or other indoor environment. Additionally, in a geographical area of a radio access network, a plurality of different kinds of radio cells as well as a plurality of radio cells may be provided. Radio cells may be macro cells (or umbrella cells) which may be large cells having a diameter of up to tens of kilometers, or smaller cells such as micro-, femto- or picocells. The access node(s) 604 of FIG.6 may provide any kind of these cells. A cellular radio network may be implemented as a multilayer access networks including several kinds of radio cells. In multilayer access networks, one access node may provide one kind of a radio cell or radio cells, and thus a plurality of access nodes may be needed to provide such a multilayer access network. For fulfilling the need for improving performance of radio access networks, the concept of “plug-and-play” access nodes may be introduced. A radio access network, which may be able to use “plug-and-play” access nodes, may include, in addition to Home eNodeBs or Home gNodeBs, a Home Node B gateway (HNB-GW) (not shown in FIG. 6). An HNB-GW, which may be installed within an operator’s radio access network, may aggregate traffic from a large number of Home eNodeBs or Home gNodeBs back to a core network 610 of the operator. 6G wireless communication networks are expected to adopt flexible decentralized and / or distributed computing systems and architecture and ubiquitous computing, with local spectrum licensing, spectrum sharing, infrastructure sharing, and intelligent automated management underpinned by mobile edge computing, artificial intelligence, short-packet communication and blockchain technologies. Key features of 6G may include intelligent connected management and control functions, programmability, integrated sensing and communication, reduction of energy footprint, trustworthy infrastructure, scalability and affordability. In addition to these, 6G is also targeting new use cases covering the integration of localization and sensing capabilities into system definition to unifying user experience across physical and digital worlds. FIG.7 illustrates an example of an apparatus 700 comprising means for performing one or more of the example embodiments (e.g., the method of any of FIGS. 2 to 5) described above. For example, the apparatus 700 may be a wireless communication device such as a user equipment 600, 602 or an access node 604 of a wireless communication network. The apparatus 700 may comprise, for example, a circuitry or a chipset applicable for realizing one or more of the example embodiments described above. The apparatus 700 may be an electronic device or computing system comprising one or more electronic circuitries. The apparatus 700 may comprise a control circuitry 710 such as at least one processor, and at least one memory 720 storing instructions 722 which, when executed by the at least one processor, cause the apparatus 700 to carry out one or more of the example embodiments described above. Such instructions 722 may, for example, include computer program code (software). The at least one processor and the at least one memory storing the instructions may provide the means for providing or causing the performance of any of the methods and / or blocks described above. In another embodiment, the means may be a network function of the core network 110, or the means may be network function virtualization infrastructure. The processor is coupled to the memory 720. The processor is configured to read and write data to and from the memory 720. The memory 720 may comprise one or more memory units. The memory units may be volatile or non-volatile. It is to be noted that there may be one or more units of non-volatile memory and one or more units of volatile memory or, alternatively, one or more units of non-volatile memory, or, alternatively, one or more units of volatile memory. Volatile memory may be for example random-access memory (RAM), dynamic random-access memory (DRAM) or synchronous dynamic random-access memory (SDRAM). Non-volatile memory may be for example read-only memory (ROM), programmable read-only memory (PROM), electronically erasable programmable read-only memory (EEPROM), flash memory, optical storage or magnetic storage. In general, memories may be referred to as non-transitory computer readable media. The term “non-transitory,” as used herein, is a limitation of the medium itself (i.e., tangible, not a signal) as opposed to a limitation on data storage persistency (e.g., RAM vs. ROM). The memory 720 stores computer readable instructions that are executed by the processor. For example, non-volatile memory stores the computer readable instructions, and the processor executes the instructions using volatile memory for temporary storage of data and / or instructions. The computer readable instructions may have been pre-stored to the memory 720 or, alternatively or additionally, they may be received, by the apparatus, via an electromagnetic carrier signal and / or may be copied from a physical entity such as a computer program product. Execution of the computer readable instructions causes the apparatus 700 to perform one or more of the functionalities described above. The memory 720 may be implemented using any suitable data storage technology, such as semiconductor-based memory devices, flash memory, magnetic memory devices and systems, optical memory devices and systems, fixed memory and / or removable memory. The apparatus 700 may further comprise or be connected to a communication interface 730 comprising hardware and / or software for realizing communication connectivity according to one or more communication protocols. The communication interface 730 may comprise at least one receiver (Rx) 100 that may be integrated to the apparatus 700 or that the apparatus 700 may be connected to. The communication interface 730 may further comprise at least one transmitter (Tx) that may be integrated to the apparatus 700 or that the apparatus 700 may be connected to. The communication interface 730 may provide means for performing some of the blocks and / or functions (e.g., receiving) for one or more example embodiments described above. The communication interface 730 may comprise one or more components, such as: power amplifier, digital front end (DFE), analog-to-digital converter (ADC), digital-to-analog converter (DAC), frequency converter, (de)modulator, and / or encoder / decoder circuitries, controlled by the corresponding controlling units. The communication interface 730 provides the apparatus with radio communication capabilities to communicate in the wireless communication network. The communication interface may, for example, provide a radio interface to one or more UEs 600, 602 and / or to one or more access nodes 604 of the wireless communication network. It is to be noted that the apparatus 700 may further comprise various components not illustrated in FIG. 7. The various components may be hardware components and / or software components. As used in this application, the term “circuitry” may refer to one or more or all of the following: a) hardware-only circuit implementations (such as implementations in only analog and / or digital circuitry); and b) combinations of hardware circuits and software, such as (as applicable): i) a combination of analog and / or digital hardware circuit(s) with software / firmware and ii) any portions of hardware processor(s) with software (including digital signal processor(s), software, and memory(ies) that work together to cause an apparatus, such as a mobile phone, to perform various functions); and c) hardware circuit(s) and / or processor(s), such as a microprocessor(s) or a portion of a microprocessor(s), that requires software (for example firmware) for operation, but the software may not be present when it is not needed for operation. This definition of circuitry applies to all uses of this term in this application, including in any claims. As a further example, as used in this application, the term circuitry also covers an implementation of merely a hardware circuit or processor (or multiple processors) or portion of a hardware circuit or processor and its (or their) accompanying software and / or firmware. The term circuitry also covers, for example and if applicable to the particular claim element, a baseband integrated circuit or processor integrated circuit for a mobile device or a similar integrated circuit in server, a cellular network device, or other computing or network device. FIG.8 illustrates an example of an artificial neural network 830 with one hidden layer 802, and FIG. 9 illustrates an example of a computational node 804. However, it should be noted that the artificial neural network 830 may also comprise more than one hidden layer 802. An artificial neural network (ANN) 830 comprises a set of rules that are designed to execute tasks such as regression, classification, clustering, and pattern recognition. The ANN may achieve such objectives with a learning / training procedure, where they are shown various examples of input data, along with the desired output. This way, the ANN learns to identify the proper output for any input within the training data manifold. Learning / training by using labels is called supervised learning and learning without labels is called unsupervised learning. Deep learning (also known as deep structured learning or hierarchical learning) is part of a broader family of machine learning methods based on the layers used in the artificial neural network. A deep neural network (DNN) 830 is an artificial neural network comprising multiple hidden layers 802 between the input layer 800 and the output layer 814. Training of DNN allows it to find the correct mathematical manipulation to transform the input into the proper output, even when the relationship is highly non-linear and / or complicated. Deep learning may require a large amount of input data. For example, the deep neural network 830 may refer to the DisturbNN 104, or the deep-learning-based detector 110, or the deep-learning-based demapper 111. A given hidden layer 802 comprises nodes 804, 806, 808, 810, 812, where the computation takes place. As shown in FIG.9, a given node 804 combines input data 800 with a set of coefficients, or weights 900, that either amplify or dampen that input 800, thereby assigning significance to inputs 800 with regard to the task that the algorithm is trying to learn. The input-weight products are added 902 and the sum is passed through an activation function 904, to determine whether and to what extent that signal should progress further through the neural network 830 to affect the ultimate outcome, such as an act of classification. In the process, the neural network learns to recognize correlations between certain relevant features and optimal results. In the case of classification, the output of a DNN 830 may be considered as a likelihood of a particular outcome. In this case, the number of layers 802 may vary proportional to the number of the used input data 800. However, when the number of input data 800 is high, the accuracy of the outcome 814 is more reliable. On the other hand, when there are fewer layers 802, the computation might take less time and thereby reduce the latency. However, this highly depends on the specific DNN architecture and / or the computational resources available. Initial weights 900 of the model can be set in various alternative ways. During the training phase, they may be adapted to improve the accuracy of the process based on analyzing errors in decision-making. Training a model is basically a trial-and-error activity. In principle, a given node 804, 806, 808, 810, 812 of the neural network 830 makes a decision (input*weight) and then compares this decision to collected data to find out the difference to the collected data. In other words, it determines the error, based on which the weights 900 are adjusted. Thus, the training of the model may be considered a corrective feedback loop. For example, a neural network model may be trained using a stochastic gradient descent optimization algorithm, for which the gradients are calculated using the backpropagation algorithm. The gradient descent algorithm seeks to change the weights 900, so that the next evaluation reduces the error, meaning that the optimization algorithm is navigating down the gradient (or slope) of error. It is also possible to use any other suitable optimization algorithm, if it provides sufficiently accurate weights 900. Consequently, the trained parameters of the neural network 830 may comprise the weights 900. In the context of an optimization algorithm, the function used to evaluate a candidate solution (i.e., a set of weights) is referred to as the objective function. With neural networks, where the target is to minimize the error, the objective function may be referred to as a cost function or a loss function. In adjusting weights 900, any suitable method may be used as a loss function. Some examples of a loss function are mean squared error (MSE), maximum likelihood estimation (MLE), and cross entropy. As for the activation function 904 of the node 804, it defines the output 814 of that node 804 given an input or set of inputs 800. The node 804 calculates a weighted sum of inputs, perhaps adds a bias, and then makes a decision as “activate” or “not activate” based on a decision threshold as a binary activation or using an activation function 904 that gives a nonlinear decision function. Any suitable activation function 904 may be used, for example sigmoid, rectified linear unit (ReLU), normalized exponential function (softmax), sotfplus, tanh, etc. In deep learning, the activation function 904 may be set at the layer level and applies to all neurons (nodes) in that layer. The output 814 is then used as input for the next node and so on until a desired solution to the original problem is found. The techniques and methods described herein may be implemented by various means. For example, these techniques may be implemented in hardware (one or more devices), firmware (one or more devices), software (one or more modules), or combinations thereof. For a hardware implementation, the apparatus(es) of example embodiments may be implemented within one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), graphics processing units (GPUs), processors, controllers, micro-controllers, microprocessors, other electronic units designed to perform the functions described herein, or a combination thereof. For firmware or software, the implementation can be carried out through modules of at least one chipset (for example procedures, functions, and so on) that perform the functions described herein. The software codes may be stored in a memory unit and executed by processors. The memory unit may be implemented within the processor or externally to the processor. In the latter case, it can be communicatively coupled to the processor via various means, as is known in the art. Additionally, the components of the systems described herein may be rearranged and / or complemented by additional components in order to facilitate the achievements of the various aspects, etc., described with regard thereto, and they are not limited to the precise configurations set forth in the given figures, as will be appreciated by one skilled in the art. It will be obvious to a person skilled in the art that, as technology advances, the inventive concept may be implemented in various ways within the scope of the claims. The embodiments are not limited to the example embodiments described above, but may vary within the scope of the claims. Therefore, all words and expressions should be interpreted broadly, and they are intended to illustrate, not to restrict, the embodiments.

Claims

CLAIMS 1. An apparatus comprising at least one processor, and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to: receive a data signal and one or more reference signals; obtain a raw channel estimate for the received data signal based on the one or more reference signals; obtain a first filtered channel estimate by denoising the raw channel estimate with a first filter; obtain a set of disturbance vectors based at least on the first filtered channel estimate; estimate one or more covariance matrices based on the set of disturbance vectors; interpolate the first filtered channel estimate into a first channel matrix; input the first channel matrix and the one or more covariance matrices into an interference rejection combining equalizer configured to mitigate interference; and process an output of the interference rejection combining equalizer with at least one machine learning model trained to detect signals, or with a combination of the at least one machine learning model and a pre-defined receiver.

2. The apparatus of claim 1, wherein the processing of the output of the interference rejection combining equalizer comprises: processing the output of the interference rejection combining equalizer with a deep-learning-based detector trained to detect signals; and processing an output of the deep-learning-based detector with a deep- learning-based demapper trained to demap the detected signals.

3. The apparatus of claim 2, further being caused to:obtain a second filtered channel estimate by denoising the raw channel estimate with a second filter different from the first filter; interpolate the second filtered channel estimate into a second channel matrix; and input the second channel matrix into another equalizer that is different from the interference rejection combining equalizer and utilized in parallel with the interference rejection combining equalizer, wherein the deep-learning-based detector is configured to combine the output of the interference rejection combining equalizer and an output of the other equalizer.

4. The apparatus of claim 3, wherein the other equalizer comprises a linear minimum mean square error equalizer.

5. The apparatus of any of claims 3 to 4, wherein the first filter comprises a trainable linear convolutional filter of a set of at least two trainable linear convolutional filters used for denoising the raw channel estimate, and wherein the second filter comprises another trainable linear convolutional filter of the set of at least two trainable linear convolutional filters.

6. The apparatus of any of claims 3 to 4, wherein the first filter and the second filter are pre-defined non-learned filters.

7. The apparatus of any of claims 3 to 4, wherein the first filter comprises a trainable linear convolutional filter, and wherein the second filter comprises a pre-defined non-learned filter.

8. The apparatus of any of claims 3 to 4, wherein the first filter comprises a pre-defined non-learned filter, and wherein the second filter comprises a trainable linear convolutional filter.

9. The apparatus of any of claims 2 to 8, further being caused to: switch the deep-learning-based demapper to a non-deep-learning- based demapper.

10. The apparatus of any preceding claim, wherein the apparatus is caused to utilize an artificial neural network to obtain the set of disturbance vectors in a sliding window fashion.

11. The apparatus of claim 10, wherein the artificial neural network is configured to determine the set of disturbance vectors based on the first filtered channel estimate, the received data signal, and the one or more reference signals.

12. The apparatus of any of claims 10 to 11, wherein the one or more covariance matrices estimated based on the set of disturbance vectors are used as a replacement of one or more interference-plus-noise covariance matrices in the interference rejection combining equalizer.

13. The apparatus of any of claims 1 to 9, wherein the apparatus is caused to utilize a pre-defined interference-plus-noise vector collection algorithm to obtain the set of disturbance vectors, wherein the one or more covariance matrices estimated based on the set of disturbance vectors comprise one or more interference-plus-noise covariance matrices.

14. The apparatus of any preceding claim, wherein the apparatus comprises a radio receiver.

15. An apparatus comprising: means for receiving a data signal and one or more reference signals;obtaining a raw channel estimate for the received data signal based on the one or more reference signals; means for obtaining a first filtered channel estimate by denoising the raw channel estimate with a first filter; means for obtaining a set of disturbance vectors based at least on the first filtered channel estimate; means for estimating one or more covariance matrices based on the set of disturbance vectors; means for interpolating the first filtered channel estimate into a first channel matrix; means for inputting the first channel matrix and the one or more covariance matrices into an interference rejection combining equalizer configured to mitigate interference; and means for processing an output of the interference rejection combining equalizer with at least one machine learning model trained to detect signals, or with a combination of the at least one machine learning model and a pre-defined receiver.

16. A method comprising: receiving a data signal and one or more reference signals; obtaining a raw channel estimate for the received data signal based on the one or more reference signals; obtaining a first filtered channel estimate by denoising the raw channel estimate with a first filter; obtaining a set of disturbance vectors based at least on the first filtered channel estimate; estimating one or more covariance matrices based on the set of disturbance vectors; interpolating the first filtered channel estimate into a first channel matrix;inputting the first channel matrix and the one or more covariance matrices into an interference rejection combining equalizer configured to mitigate interference; and processing an output of the interference rejection combining equalizer with at least one machine learning model trained to detect signals, or with a combination of the at least one machine learning model and a pre-defined receiver.

17. A non-transitory computer readable medium comprising program instructions which, when executed by an apparatus, cause the apparatus to perform at least the following: receiving a data signal and one or more reference signals; obtaining a raw channel estimate for the received data signal based on the one or more reference signals; obtaining a first filtered channel estimate by denoising the raw channel estimate with a first filter; obtaining a set of disturbance vectors based at least on the first filtered channel estimate; estimating one or more covariance matrices based on the set of disturbance vectors; interpolating the first filtered channel estimate into a first channel matrix; inputting the first channel matrix and the one or more covariance matrices into an interference rejection combining equalizer configured to mitigate interference; and processing an output of the interference rejection combining equalizer with at least one machine learning model trained to detect signals, or with a combination of the at least one machine learning model and a pre-defined receiver.

Images