Power-delay profile estimation for bundle-based channel estimation via learning

Abstract

A method for channel estimation of a precoded channel, comprising: generating an initial frequency autocorrelation of the precoded channel for a current bundle of received data transmissions; generating an extended frequency autocorrelation based on the initial frequency autocorrelation of the precoded channel; providing the extended frequency autocorrelation to a neural network; generating, by the neural network, an estimated frequency autocorrelation of an unprecoded channel based on the extended frequency autocorrelation; and generating an estimated power distribution profile of the unprecoded channel based on the estimated frequency autocorrelation.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to and the benefit of U.S. Provisional Application No. 63 / 004,918, filed April 3, 2020 (“PDP Estimation for Bundle-Based Channel Estimation by Reinforcement Learning Using an Actor-Critical Approach”), and U.S. Provisional Application No. 63 / 024,196, filed May 13, 2020 (“PDP Estimation for Bundle-Based Channel Estimation by Supervised Learning”), the entire contents of which are incorporated herein by reference. Technical Field

[0003] Aspects of the present disclosure relate to communication channel estimation. Background Art

[0004] The Physical Downlink Shared Channel (PDSCH) is a physical channel typically used to carry user data, dedicated control and user-specific higher-layer information, and downlink system information. Resource blocks (RBs) used for PDSCH can be allocated in bundles of two or more RBs. Resource blocks within a bundle can be precoded in the same manner, and other resource blocks mapped to different bundles can be precoded independently based on the decision of the radio node (e.g., 5G logical radio node, i.e., gNB).

[0005] Under the latest specifications released by the 3rd Generation Partnership Project (3GPP), there are two resource allocation scenarios for PDSCH, namely, precoding for all bundles (using the same or different precoding) or no precoding at all. The latter is equivalent to the same precoding in all bundles. The user equipment (UE) can take advantage of this configuration to improve channel estimation. When the same precoding is applied to all bundles, the frequency domain signal is homogeneous in terms of precoding and can therefore be converted to the time domain by performing an inverse fast Fourier transform (IFFT). This means that noise reduction in the time domain can be exploited based on the estimated power delay profile (PDP) of the user equipment, which recursively provides an advantage when measuring the channel power using an infinite impulse response (IIR) filter to estimate the PDP for the next time / frequency slot.

[0006] However, this approach is not applicable when each bundle is independently precoded with its own choice, because the user device does not know the precoding choice of the radio node at each bundle. Time domain conversion via IFFT is not feasible because the effective channel incorporating precoding is no longer homogeneous across bundles. The user device must then estimate the channel based on information in the frequency domain.

[0007] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute prior art. Summary of the Invention

[0008] Various aspects of the example embodiments of the present disclosure are directed to a system and method for performing channel estimation by utilizing a neural network (e.g., a deep learning neural network). According to some embodiments, the neural network is trained with samples having precoded channel correlations with labels having uncoded channel correlations, and thus the channel correlations can be converted to be homogeneous. In some embodiments, a channel estimator performs post-processing on the output of the neural network to further refine the estimate and allow it to be denoised in the time domain via an inverse fast Fourier transform (IFFT). This allows the channel estimator to estimate the channel correlations in the frequency domain even if random precoding is configured across the bundles.

[0009] According to some embodiments of the present disclosure, a method for channel estimation of a precoded channel is provided, the method comprising: generating an initial frequency autocorrelation of the precoded channel for a current bundle of received data transmissions; generating an extended frequency autocorrelation based on the initial frequency autocorrelation of the precoded channel; providing the extended frequency autocorrelation to a neural network; generating, by the neural network, an estimated frequency autocorrelation of an unprecoded channel based on the extended frequency autocorrelation; and generating an estimated power distribution profile of the unprecoded channel based on the estimated frequency autocorrelation.

[0010] In some embodiments, the current bundle includes a plurality of resource blocks, each resource block including a plurality of subcarriers.

[0011] In some embodiments, the non-precoded channel is an estimate of the precoded channel without precoding.

[0012] In some embodiments, generating the extended frequency autocorrelation includes performing edge extension on the initial frequency autocorrelation to expand the size of the estimated frequency autocorrelation to a fast Fourier transform (FFT) size, wherein the FFT size is an input size of the neural network.

[0013] In some embodiments, edge extension comprises linearly interpolating the values ​​of the initial frequency autocorrelations via an extension matrix.

[0014] In some embodiments, providing the spread frequency autocorrelation to the neural network includes providing a first half of the value of the spread frequency autocorrelation to the neural network, wherein a second half of the value of the spread frequency autocorrelation is a complex conjugate of the first half of the value of the spread frequency autocorrelation.

[0015] In some embodiments, generating the estimated frequency autocorrelation by the neural network includes: generating, by the neural network, a first half of the value of the estimated frequency autocorrelation of the unprecoded channel based on the expanded frequency autocorrelation; and calculating a second half of the value of the estimated frequency autocorrelation of the unprecoded channel as the complex conjugate of the value of the first half.

[0016] In some embodiments, generating the estimated power distribution profile includes: filtering the estimated frequency autocorrelation of the neural network output through a low-pass filter to generate a fine autocorrelation of the unprecoded channel; and performing an inverse FFT (IFFT) operation on the fine autocorrelation to generate the estimated power distribution profile.

[0017] In some embodiments, the low pass filter is a moving average filter.

[0018] In some embodiments, generating the initial frequency autocorrelation of the precoding channel for the current bundling bundle includes: generating a time autocorrelation for a previous bundling bundle of received data transmissions; generating a previous frequency autocorrelation for the previous bundling bundle based on a previously estimated power distribution profile; generating an estimated channel input response based on the time autocorrelation and the previous frequency autocorrelation; and generating the initial frequency autocorrelation of the precoding channel for the current bundling bundle based on the estimated channel input response.

[0019] In some embodiments, the method further comprises generating a truncated estimated power profile by truncating a size of the estimated power profile to match a size of an initial frequency autocorrelation of the precoded channel.

[0020] In some embodiments, the method further comprises normalizing the truncated estimated power distribution profile to unit power to generate a normalized estimated power distribution profile.

[0021] In some embodiments, the truncated estimated power profile has a length of a maximum delay spread of the precoded channel.

[0022] According to some embodiments of the present disclosure, a system for channel estimation of a precoded channel is provided, the system comprising: a processor; and a memory storing instructions that, when executed on the processor, cause the processor to: generate an initial frequency autocorrelation of the precoded channel for a current bundle of received data transmissions; generate an extended frequency autocorrelation based on the initial frequency autocorrelation of the precoded channel; provide the extended frequency autocorrelation to a neural network; generate, by the neural network, an estimated frequency autocorrelation of an unprecoded channel based on the extended frequency autocorrelation; and generate an estimated power distribution profile of the unprecoded channel based on the estimated frequency autocorrelation.

[0023] According to some embodiments of the present disclosure, a method for channel estimation of a precoded channel is provided, the method comprising: generating an initial frequency autocorrelation of the precoded channel for a current bundle of received data transmissions; providing the initial frequency autocorrelation to a policy network; the policy network generating an estimated frequency autocorrelation of an unprecoded channel based on the initial frequency autocorrelation; determining, by a value network, an instantaneous reward based on the estimated frequency autocorrelation; determining an advantage based on the instantaneous reward and a predicted total reward forward-propagated on the value network; and updating, based on the advantage, a policy of the policy network through back-propagation to reduce a block error rate.

[0024] In some embodiments, updating the policy of the policy network includes: determining a policy gradient according to the advantage; and updating coefficients of the policy network according to the policy gradient.

[0025] In some embodiments, the policy network and the value network are multilayer perceptrons.

[0026] In some embodiments, the method further comprises adding Gaussian noise to the estimated frequency autocorrelation to convert the discrete action space of the policy network into a continuous action space.

[0027] In some embodiments, the method further comprises generating an extended frequency autocorrelation based on the initial frequency autocorrelation of the precoded channel, wherein providing the initial frequency autocorrelation to the policy network comprises providing the extended frequency autocorrelation to the policy network, and wherein generating an estimated frequency autocorrelation of the non-precoded channel based on the extended frequency autocorrelation.

[0028] In some embodiments, the method further comprises filtering the estimated frequency autocorrelation via a low pass filter to generate a fine autocorrelation of the unprecoded channel; and performing an inverse FFT (IFFT) operation on the fine autocorrelation to generate the estimated power distribution profile. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] These and other features of some example embodiments of the present disclosure will be understood and appreciated with reference to the specification, claims, and drawings, in which:

[0030] Figure 1 is a block diagram of a communication system utilizing a channel estimator according to some embodiments of the present disclosure;

[0031] Figure 2A shows a mapping of bundles to time slots at a receiver according to some examples;

[0032] Figure 2B shows the allocation of bundles over frequency by radio at a receiver of a communication system according to some embodiments of the present disclosure;

[0033] Figure 3Ais a block diagram of a channel estimator utilizing supervised learning according to some embodiments of the present disclosure;

[0034] Figure 3B shows the effects of various components of a channel estimator according to some embodiments of the present disclosure;

[0035] Figures 4A to 4B A graph showing estimated autocorrelations output from a neural network and refined autocorrelations produced by a filter according to some embodiments of the present disclosure;

[0036] Figures 5A to 5B According to some embodiments of the present disclosure, a sprite power delay profile (PDP) value is compared to an estimated value;

[0037] Figures 6A to 6B shows the Block Error Rate (BLER) and Signal-to-Noise Ratio (SNR) performance gains of the channel estimator 100 for a rank-2 extended vehicle A-model (EVA) channel according to some example embodiments of the present disclosure;

[0038] Figure 7 is a block diagram of a channel estimator utilizing a policy network trained by an advantage actor-critic (A2C) method according to some embodiments of the present disclosure;

[0039] Figure 8A 、 8B and 8C illustrate block error rate (BLER) and signal-to-noise ratio (SNR) performance gains of channel estimators for rank-2 extended pedestrian A-model (EPA) channel, EVA channel, and extended typical urban model (ETA) channel according to some example embodiments of the present disclosure; and

[0040] Figure 9A 、 Figure 9B and Figure 9C The block error rate (BLER) and signal-to-noise ratio (SNR) performance gains of the channel estimators for a rank-4 EPA channel, an EVA channel, and an ETA channel according to some example embodiments of the present disclosure are respectively shown. DETAILED DESCRIPTION

[0041] The detailed description set forth below in conjunction with the accompanying drawings is intended as a description of some example embodiments of the systems and methods for channel estimation provided according to the present disclosure, and is not intended to represent the only form in which the disclosure in the accompanying drawings can be constructed or utilized. This description sets forth the features of the present disclosure in conjunction with the embodiments shown. However, it should be understood that the same or equivalent functions and structures can be implemented by different embodiments, and these different embodiments are also intended to be included within the scope of the present disclosure. As indicated elsewhere herein, similar element numbers are intended to indicate similar elements or features.

[0042] The latest technical standards for communication networks support configurations based on bundles, where each bundle can be precoded with its own selection from a precoding matrix. In order to facilitate communication in such a system, the user equipment (UE) must estimate the channel in the frequency domain (e.g., estimate the power delay profile (PDP) of the channel). A channel estimation technique, namely minimum mean square error (MMSE), utilizes second-order statistics of the channel, which consists of frequency and time correlations (e.g., frequency and time autocorrelation). The time correlation (e.g., time autocorrelation) can be determined based on known techniques. However, the frequency correlation (e.g., frequency autocorrelation) involves an accurate estimation of the PDP information. Assuming a uniform PDP when deriving the frequency correlation may lead to performance degradation, especially when the channel delay is relatively long.

[0043] Therefore, according to some embodiments, the channel estimator utilizes a neural network that receives the frequency correlation of the precoded channel from the previous time slot / bundle and outputs the PDP information for the current time slot / bundle. In some embodiments, the channel estimator also performs filtering, truncation, and normalization to refine the output of the neural network, which is used to estimate the channel for the current bundle / time slot.

[0044] Figure 1 is a block diagram of a communication system 1 utilizing a channel estimator 100 according to some embodiments of the present disclosure.

[0045] The communication system 1 may include a transmitter 10, a communication channel (e.g., a wireless multipath channel) 20, and a receiver 30. The transmitter 10 may include: a source 12 of input data; a channel encoder 14 configured to encode the input data to enable error correction at the receiver 30; a modulator 16 configured to generate a transmit signal based on the encoded input data; and a precoder 18 for precoding one or more bundles prior to transmission over the communication channel 20.

[0046] The receiver 30 includes a receiver filter 32 for filtering out noise that may have been added to the transmitted signal in the multipath channel 20; a detector 34 configured to reconstruct the encoded data from the received signal; and a channel decoder 36 configured to decode the reconstructed data to retrieve the input data generated by the source 12.

[0047] The transmitter 10 may be a radio node and the receiver 30 may be part of a user equipment, which may be mobile. For example, because the transmitter 10 and / or the receiver 30 are in motion, the communication channel 20 may not be constant and may vary over time. Mobile wireless communications may be adversely affected by multipath interference caused by reflections from the surrounding environment (e.g., hills, buildings, and other obstacles). Accurate estimation of the time-varying channel is key to providing reliability and high data rates at the receiver 30. Therefore, according to some embodiments, the receiver 30 further includes a channel estimator 100 that utilizes a neural network to estimate the channel, for each bundled transmitted signal, i.e., a channel impulse response (CIR), and provides it to the detector 34.

[0048] The signal y received by the receiver 30 can be expressed as:

[0049] y=p+n (Formula 1)

[0050] Where p is the reference signal (RS) channel vector of the demodulation reference signal (DMRS), n is the background noise, which has zero mean and covariance σ 2 I (where I is the identity matrix). Channel input response The estimate of can be expressed as:

[0051]

[0052] where R hp represents the correlation matrix between h and p. Similarly, R pp Represents the autocorrelation matrix of p. Autocorrelation R pp It may be a function of p alone, where p is the DMRS channel vector known to the receiver 30 .

[0053] Here, it is assumed that the channel distribution follows the generalized stationary uncorrelated scattering (WSSUS) model. In other words, the second-order moment of the channel is fixed and depends only on the amount of time or frequency difference, rather than every instantaneous value. Under the WSSUS model, the channel autocorrelation can be decomposed into a frequency domain part and a time domain part as follows:

[0054]

[0055] where h i,j is the complex channel gain at the i-th subcarrier of the j-th symbol, and r f () and r t () are the frequency and time autocorrelation functions, respectively. Appropriate choice of subcarrier numbers i and k and symbol values ​​j and l allows R hp and R pp based on Calculation.

[0056] Time autocorrelation function r t () can be calculated in a variety of ways. For example, the temporal autocorrelation function can rely on linear interpolation to obtain the correlation value between two symbols, which is given by the following formula:

[0057]

[0058] Where TC(x) is the correlation value for interval x, T s In other examples, the Jakes model can be used to generate:

[0059] r t (l) = J0(2πT s f D l) (Formula 5)

[0060] where J0 is the zero-order Bessel function of the first kind, f D represents the Doppler spread corresponding to the maximum Doppler shift.

[0061] Given the power delay profile (PDP) of channel 20, the frequency autocorrelation function can be expressed using the Fast Fourier Transform (FFT) of the channel power as

[0062]

[0063] Where L is the number of channel taps in the time domain (also called the maximum delay spread), and Δf is the subcarrier spacing. i and τ i are the power and delay of the i-th channel tap, respectively. The maximum delay spread L can be measured using a quasi-collocated (QCL) reference signal. Here, the total power in the profile is normalized to unit power, i.e.

[0064]

[0065] According to some embodiments, the channel estimator 100 estimates the value P by utilizing a neural network. i (e.g., as close to the ideal value as possible). The PDP value can be used to determine the frequency autocorrelation function r using Equation 6. f (). Frequency autocorrelation function r f () and the time autocorrelation function r t () For example, it is determined by Formula 6 or Formula 4 or 5 respectively. The receiver 30 can determine the channel correlation matrix R by Formula 3 h,h , from which the channel autocorrelation R can be calculated hp and R pp Receiver 30 may then determine the estimated channel input response via Equation 2 In some embodiments, receiver 30 estimates the channel response for each bundle of transmissions individually.

[0066] Figure 2A A mapping of bundles to time slots at the receiver 30 is shown according to some examples. Figure 2B Bundling bundle allocation over frequency on a radio (e.g., 5G New Radio (NR)) at a receiver 30 is shown in accordance with some embodiments of the present disclosure.

[0067] refer to Figure 2A In some examples, each received bundle 50 includes multiple resource blocks / elements 52 (e.g., 2 or 4 RBs) and corresponds to a separate time slot at the receiver. Each bundle 50 can have a different precoding. Each resource block 52 can include multiple subcarriers (SCs). In some examples, each resource block 52 includes 12 subcarriers, so each bundle corresponds to 24 subcarriers.

[0068] refer to Figure 2B , a total of N bundles (N is an integer greater than 1) are allocated in frequency (as shown on the vertical axis), where the maximum number of resource blocks N is determined based on both the subcarrier spacing and the channel bandwidth max (CHBW). Table 1 lists the maximum number N of resource blocks in the first frequency range (FR1) max In the example of Table 1, under a channel bandwidth of 100 MHz and a subcarrier spacing of 30 kHz, the maximum number of resource blocks N is max The highest (i.e. 273).

[0069] Table 1:

[0070]

[0071] like Figure 2B As shown, the size of the resource allocation (the 1st to Nth bundles 50) is smaller than the entire FFT size (e.g., 2048), which can accommodate different combinations of channel bandwidth and subcarrier spacing. Since the receiver 30 may not know the frequency resource allocation at the transmitter 10, in some embodiments, the receiver 30 is designed for a large FFT size, and when the resource allocation is smaller than the FFT size, the input is expanded. That is, even if the receiver 30 uses frequency autocorrelation for PDP estimation (which may be smaller than the FFT size), embodiments of the present disclosure also consider areas outside the full RB up to the entire FFT size.

[0072] According to some embodiments, to determine the PDP (a time-domain characteristic) for each bundle transmitted, the frequency correlation of the physical downlink shared channel (PDSCH; see Equation 20 below) combined with the DMRS in the previous time slot is fed into a neural network. The output of the frequency autocorrelation is post-processed to estimate the PDP for the current time slot. As a result, the channel estimator improves the block error rate (BLER) compared to prior art techniques that assume the channel has a uniform PDP.

[0073] Figure 3A is a block diagram of a channel estimator 100 using supervised learning according to some embodiments of the present disclosure. Figure 3B The effects of various components of the channel estimator 100 according to some embodiments of the present disclosure are shown.

[0074] According to some embodiments, channel estimator 100 includes edge extender 110, neural network 120, post-processor 125, and narrowband channel estimator (NBCE) 160. In some embodiments, post-processor 125 includes filter 130, inverse fast Fourier transform (IFFT) converter 140, and truncation and normalization block 150.

[0075] Reference Figures 3A to 3B In some embodiments, the neural network 120 receives the frequency autocorrelation of the precoded channel at its input and produces an estimated frequency autocorrelation at its output, but without the precoding. In other words, the neural network 120 effectively strips the precoding from the channel PDP for a particular bundle. After post-processing performed by the filter 130, the IFFT converter 140, and the truncation and normalization module 150, the estimated frequency autocorrelation of one bundle / time slot is used as input to the neural network 120 for a subsequent bundle / time slot. As the neural network 120 sequentially processes bundles in the received signal, the estimated channel PDP can become more accurate and precise (e.g., closer to the actual PDP). This may be due to the fact that, in reality, the PDP does not vary much from one time slot to the next. In this way, the channel estimator 100 can use information from previous time slots / bundles to enhance the calculations for the current time slot / bundle.

[0076] The input size of the neural network 120 is fixed to be the same as the Fast Fourier Transform (FFT) size. Here, the FFT size can represent the number of pins in the spectrum analysis window. This allows a single network to cover all resource block configurations allocated for PDSCH and DMRS, for example, up to 273 resource blocks (as shown in the example of Table 1). Sizing the neural network input to be the same as the FFT size eliminates the need to design multiple networks, each corresponding to a resource block size. This may be particularly desirable because the channel estimator 100 (e.g., channel estimator 100) may not be aware of the frequency resource allocation at the transmitter 10 and is therefore sized to accommodate different frequency resource allocations at the transmitter 10.

[0077] Therefore, according to some embodiments, to maintain the same size of the input features, the edge expander 110 expands the measured autocorrelation to the FFT size by using edge expansion. In some embodiments, the edge expander 110 interpolates the signal (eg, via linear interpolation) using an expansion matrix as follows.

[0078]

[0079] in

[0080]

[0081] where N f is the size of the FFT, N d is the size of the measured / calculated autocorrelation (also called initial frequency autocorrelation) The extended frequency autocorrelation can then be expressed as

[0082]

[0083] in It is defined in Equation 22 below.

[0084] However, embodiments of the present invention are not limited to the above interpolation, and any suitable spreading / interpolation technique may be employed to obtain the spread frequency autocorrelation based on the measured autocorrelation.

[0085] According to some embodiments, the channel estimator 100 utilizes the symmetric property of autocorrelation to remove repeated information from the neural network 120. Therefore, in some embodiments, the channel estimator 100 converts the extended frequency autocorrelation value Half of the Figure 3B As shown), only the corresponding value of the channel autocorrelation is calculated by the neural network. The other half of the estimated frequency autocorrelation can be calculated using the following formula:

[0086]

[0087] That is, one half of the estimated channel autocorrelation can be calculated as the complex conjugate of the other half. Therefore, the output of the neural network 120 can be restored to the full size of the FFT based on the half-size FFT at the input of the neural network 120. Performing the inference of the frequency autocorrelation value significantly reduces the computational load on the neural network 120 and improves the inference performance.

[0088] According to some embodiments, filter 130, IFFT converter 140, and truncation and normalization block 150 apply post-processing to the output of neural network 120 for further stabilization. In some embodiments, filter 130 applies a low-pass filter to the output of the neural network. It is the estimated autocorrelation of the unprecoded channel to generate a fine frequency autocorrelation The low-pass filter can be a moving average over frequency, expressed as

[0089]

[0090] Where 2n+1 is the order of the moving average.

[0091] Figures 4A to 4B shows the estimated autocorrelation output from the neural network 120 according to some embodiments of the present disclosure and the fine autocorrelation generated by filter 130 , where n = 2. Figure 4B In this figure, relative to Figure 4A be magnified. Figure 4A Better showing the fine autocorrelation Smoothness. Figures 4A to 4B In the example, curve 200 represents the estimated autocorrelation Curve 202 represents the fine autocorrelation

[0092] According to some embodiments, the IFFT converter 140 converts the refined frequency autocorrelation into an estimated PDP (ie, the estimated PDP in Equation 6) by performing an IFFT operation. i value). The truncation and normalization block 150 further refines the estimated PDP in the time domain. In some embodiments, the truncation and normalization block 150 truncates / punctures the estimated PDP to the length of the maximum delay spread L and normalizes the estimated PDP to unit power to satisfy the conditions of Equation 7. The PDP estimate is then given by the following formula:

[0093]

[0094] in, is the power value at each tap k derived from the output of the neural network 120. Thus, the truncation and normalization block 150 stabilizes the PDP estimate

[0095] According to some embodiments of the present disclosure, Figures 5A to 5B Compare the sprite (i.e., ideal) PDP value to the estimated For comparison. Figures 5A to 5B In FIG. 3 , curve 300 represents the ideal PDP value, and curve 302 represents the estimated PDP value. value. Figure 5A and Figure 5B The scenarios with 106 resource blocks and 2 resource blocks are shown respectively. Therefore, as the number of resource blocks in the received signal increases, the estimated PDP value becomes a better approximation of the ideal PDP value.

[0096] In some embodiments, NBCE 160 generates the frequency autocorrelation function r by performing an FFT operation on the fine PDP estimate according to Equation 6. f ().

[0097] According to some embodiments, the NBCE 160 also uses the frequency autocorrelation function r generated for the current bundle / time slot. f () to calculate the neural network input generated for the subsequent / next bundle / time slot. In some embodiments, the NBCE 160 uses the calculated frequency autocorrelation function r f () and the time autocorrelation function r t () (eg, determined by Formula 4 or 5) to determine the channel autocorrelation R by Formula 3 h,h , Formula 3 is used to calculate the channel autocorrelation R hp and R pp The NBCE 160 then calculates the estimated channel input response using Equation 2

[0098] According to some embodiments, the NBCE 160 then proceeds to calculate the frequency autocorrelations of subsequent time slots / bundles by using the following.

[0099]

[0100] Where s is the symbol index within the slot, r is the antenna index of the receiver 30, l is the layer index assigned to the PDSCH and DMRS ports, and n is the resource element (RE) index. is normalized (Equation 21) so that the maximum value is set in the middle of the vector. In some examples, The length of may vary depending on the number of resource blocks (RBs) allocated to the channel estimator 100 in each time slot.

[0101] According to some embodiments, channel estimator 100 (eg, NBCE 160) calculates the channel autocorrelation r for the first time slot being analyzed in the data transmission by neural network 120. f () Use a unified PDP.

[0102] In some embodiments, rather than relying solely on the channel autocorrelation r of the previous time slot f (), the channel estimator 100 calculates autocorrelations over multiple past time slots and averages the averaged autocorrelations before injecting them into the neural network 120.

[0103] According to some embodiments, the neural network 120 utilizes a model that correlates the multiple frequency autocorrelations of the precoded channels across the bundle with the multiple frequency autocorrelations without precoding. By utilizing the model and a supervised machine learning algorithm, such as one of various known regression or backpropagation algorithms, the neural network 120 estimates the autocorrelations It is the estimated frequency autocorrelation of the unprecoded channel for a given bundle.Here, an unprecoded channel refers to an estimate of a precoded channel in the absence of (eg, removal of or without) precoding.

[0104] According to some embodiments, the neural network 120 (e.g., a deep neural network) can be a special-purpose AI or a general-purpose AI and is trained using training data (e.g., pre-coded and non-coded frequency autocorrelations) and algorithms (e.g., a back-propagation algorithm).

[0105] The neural network 120 may include a set of weights for each parameter of a linear regression model, or the neural network 120 may include a set of weights for the connections between neurons of a trained neural network. In some embodiments, the frequency autocorrelation function r of the precoded channel across the bundle is f () is provided to the neural network 120 as a value to the input layer of the neural network 120, and the value (or a set of intermediate values) is forward propagated through the neural network 120 to produce an output, where the output is the estimated autocorrelation of the channel without precoding

[0106] In the training example, 3 different types of precoding for a bundle with 2 resource blocks are bypass (i.e. identity), random, and PMI (precoding matrix indicator) based precoding. The sprite PDP for each channel can also be used to calculate the R in the bundle, given the configuration used to generate the samples. pp and R hpTherefore, a pair of samples (i.e., the frequency autocorrelation of the precoded channel) and a label (i.e., the frequency autocorrelation of the unprecoded channel calculated backward from the sprite PDP) can be collected through simulation.

[0107] The maximum number of resource blocks (RBs) per subcarrier can be allocated to generate data samples instead of all the RBs per RB subcarrier. As described above, edge extension is used to keep the input of the neural network 120 at the size of the FFT. For example, when using a 15kHz subcarrier spacing on a 20MHz channel bandwidth, the maximum configurable number of RBs in the radio is 106. Similarly, with a 30kHz subcarrier spacing, 273 RBs can be allocated on a channel with a bandwidth of 100MHz, as shown in Table 1.

[0108] Figures 6A to 6B The block error rate (BLER) and signal-to-noise ratio (SNR) performance gains of the channel estimator 100 for a rank-2 extended vehicle A-model (EVA) channel according to some example embodiments of the present disclosure are shown. Figures 6A to 6B The NBCE with supervised learning is compared with three reference figures corresponding to ideal CE, uniform PDP and ideal PDP. Here, the legend "mlp-PDP" represents the NBCE with PDP estimation. As shown in the figure, the BLER of the channel estimator 100 is Figure 6A ) or full allocation of resources (e.g. Figure 6B ) is very close to the BLER of the ideal PDP.

[0109] As described above, the channel estimator 100 estimates the power delay profile (PDP) in Formula 6, ie, P , by using the frequency autocorrelation of the PDSCH combined with the DMRS in the previous slot via a neural network. i value, and ultimately enhances the error of channel estimation in the current time slot.

[0110] As described above, the channel estimator 100 according to some embodiments aims to estimate the PDP as close to the ideal value as possible. However, embodiments of the present invention are not limited thereto.

[0111] Assuming that the channel distribution follows a generalized stationary uncorrelated scattering (WSSUS) model, and due to imperfections in NBCE (e.g., due to estimation errors and background noise), an ideal PDP value may not guarantee optimization (e.g., minimization) of the black box error rate (BLER). As a result, according to some embodiments, channel estimation is performed in a manner that minimizes the mean square error (MSE) of the channel estimate, which may result in a reduction (e.g., minimization) of the BLER.

[0112] According to some examples, the NBCE PDP estimation is formulated as a one-step Markov decision process (MDP). That is, the action at time slot i does not affect the state at time slot i+1. The action is the receiver's PDP estimate for each time slot, and the state is only associated with the channel. Therefore, the one-step MDP is modeled as a trajectory that terminates after a single time step through the reward.

[0113] The MDP framework includes states, actions, and rewards. According to some embodiments, the states represent the frequency autocorrelation of the channels, and each channel can be precoded per bundle. Since the precoding matrix used by transmitter 10 is transparent to receiver 30, the frequency autocorrelation of each time slot can be calculated by combining the estimated channel with the precoding of the previous time slot according to Equations 20 to 22.

[0114] Figure 7 is a block diagram of a channel estimator 200 utilizing a policy network trained by an dominant actor-critic (A2C) method, according to some embodiments of the present disclosure.

[0115] According to some embodiments, the edge extender 110, post-processor 125, and narrowband channel estimator (NBCE) 160 of the channel estimator 200 are connected to Figures 3A to 3B Thus, the policy network 122 receives the same input as the neural network 120. and generates a frequency autocorrelation function at the output This is similar to the output of neural network 120. However, in some embodiments, the strategy network 122 is trained to produce the frequency autocorrelation of the unprecoded channel. The unprecoded channel may generate a PDP that may not be ideal but reduces (eg, minimizes) the BLER of the receiver 30 .

[0116] In some embodiments, the channel estimator 200 includes: a Gaussian noise generator 170 for adding Gaussian noise to the output of the policy network 122; and a value network 180 for evaluating the output of the policy network 122 and correcting the coefficients or neural weights of the policy network 122 to reduce (e.g., minimize) the overall BLER of the receiver 30. In some examples, the Gaussian noise can have a zero mean and a preset variance (e.g., a small fixed variance) and can convert the discrete action space of the policy network into a continuous action space. The policy network 122 also utilizes The real and imaginary parts are used to generate an action with Gaussian noise induced from the Gaussian noise generator 170, which is the frequency autocorrelation of the estimated non-precoded channel.

[0117] In some embodiments, the value network 180 receives the state (ie, the estimated frequency autocorrelation output by the NBCE 160) ) and generate corresponding rewards In some embodiments, the reward is the negative mean square error of the channel estimate relative to the ideal channel value in dB. It can be the reward of the forward propagation at the value network 180.

[0118] Here, the policy network 122 is referred to as an actor, and the value network 180 is referred to as a critic, which evaluates the goodness or badness of the actions taken by the actor.

[0119] In some examples, a pair of states and rewards is sampled for training in a value network with multiple random seeds (e.g., 20 random seeds). The network with the lowest loss function is selected as value network 180. Value network 180 can be a multilayer perceptron. According to some examples, value network 180 has a single hidden layer with 128 nodes. An S function can be used at the activation layer of value network 180, and the output layer of value network 180 can be bypassed without a specific function. The loss function can be designed to reduce (e.g., minimize) the mean square error (MSE).

[0120] In some embodiments, after the value network 180 calculates the reward, the channel estimator 200 calculates the advantage, which can be expressed as:

[0121]

[0122] where r(s i ,a i ) is in state s i Action a i The instantaneous reward caused by is the predicted forward total reward at the output of the value network 180, and i is the time slot index. The advantage shows that if the policy network 122 is in state s i Took action a i , then the expected reward (relative to the average state) is increased. In other words, if the advantage is positive, the gradient is moved in that direction, and if the advantage is negative, the gradient is moved in the opposite direction. The channel estimator 200 then calculates the target gradient.

[0123]

[0124] in is the gradient of the target J(θ), θ represents the coefficient of the policy network 122, t represents the time index from 0 to T, T represents the number of time steps, and π θ (a t |st ) represents the policy network 122 in a given state s t When used to determine the operation a t The probability function π of the policy network 122 can be trained by supervised learning θ (a t |s t ). In some examples, a one-step MDP is modeled to terminate the trajectory after a single time step, ie, T=1.

[0125] According to some embodiments, the channel estimator 200 then calculates the by The policy coefficients (or network coefficients) θ are replaced, and the policy is updated (eg, the coefficients of the policy network 170 are updated) via back propagation, where α is a small coefficient, such as 0.05.

[0126] According to some embodiments, during the training phase of the policy network 122, the channel estimator repeatedly performs the following steps: determining an action by the policy network 122 for a given state, determining the reward for that action and state by the value network 180, evaluating the advantage of learning through a one-step reinforcement learning, calculating the target gradient, and updating the policy coefficients through backpropagation. This cycle can continue until the improvement converges or until a predetermined threshold is met.

[0127] According to some embodiments, the value network 180 may be a specialized AI or a general AI and is trained using training data and algorithms (eg, a back-propagation algorithm).

[0128] The value network 180 may include a set of weights for each parameter of a linear regression model, or the value network 180 may include a set of weights for the connections between neurons of a trained neural network. In some embodiments, the frequency autocorrelation function of the unprecoded channel across the bundle is provided to the value network 180 as the values ​​of the input layer to the value network 180, and these values ​​(or a set of intermediate values) are propagated forward through the value network 180 to generate an output, where the output is the instantaneous reward resulting from the action taken by the policy network 122.

[0129] According to some embodiments, while value network 180 is present in channel estimator 200 for the purpose of training policy network 122, value network 180 may be omitted from channel estimator 200 during the inference phase when channel estimator 200 is used to perform channel estimation on an input signal.

[0130] In some examples, the receiver 30 may be equipped with 2 or 4 receiving antennas, and the transmitter 10 transmits signals with the same rank as the receiving antennas. Here, rank refers to the channel input response The number of resource blocks can be set to 106 on a channel bandwidth of 20 MHz.

[0131] In some examples, training can be performed with samples from all Extended Pedestrian A-Model (EPA), Extended Vehicle A-Model (EVA), and Extended Typical Urban Model (ETA) channels, with each value and policy network covering all channels. The initial policy network 122 can be initially trained using supervised learning, where all precoding options are sampled, such as identity, random, and PMI-based precoding. The neural network can have batch normalization enabled for its training so that the inputs to the hidden layers are normalized with zero mean and unit variance.

[0132] Figure 8A 、 8B 8C respectively show the block error rate (BLER) and signal-to-noise ratio (SNR) performance gains of the channel estimator 200 for a rank-2 EPA channel, an EVA channel, and an ETA channel according to some example embodiments of the present disclosure. Figure 9A 、 9B 9C respectively show the block error rate (BLER) and signal-to-noise ratio (SNR) performance gains of the channel estimator 200 for a rank-4 EPA channel, an EVA channel, and an ETA channel according to some example embodiments of the present disclosure.

[0133] EPA, EVA, and ETU are multipath fading channel model delay profiles, representing low, medium, and high delay spread environments, respectively. Assuming an EPA channel has a relatively short delay spread (e.g., only 410ns at most), there may not be much room to improve the PDP estimate compared to a uniform or ideal PDP assumption. However, as the maximum delay spread in EVA and ETU channels increases, the PDP estimate varies significantly across the delay spread range, and performance is significantly impacted. Therefore, it can also be improved rather than degraded.

[0134] like Figure 8A 、 Figure 9A As shown, the BLER v. SNR performance in the EPA channel does not change much across iterations (i.e., iterations of actor-evaluator training of the policy network 122). Figure 8B 、 Figure 9B and Figure 8C 、 Figure 9C As shown, A2C training of the policy network 122 does effectively improve performance by maximizing the reward (i.e., minimizing the MSE of the channel estimate over multiple interactions). Figures 8B to 8C and Figures 9B to 9CIn the example, we can observe that the A2C training of the policy network can even outperform the performance of the ideal PDP. In addition, as shown in the figure, a single network can also be applied to multiple different channel environments.

[0135] Table 2 provides the performance gain of using A2C for channel estimation under the unified PDP assumption for NBCE. Channel estimation using A2C may outperform channel estimation using supervised learning and may also outperform the scheme using ideal PDP. In other words, under the assumption of WSSUS, the ideal PDP may not be optimal for NBCE.

[0136] Table 2:

[0137] EPA ETU EVA Rank 2 0.5dB >2.0dB 0.6dB Rank 4 0.3dB >1.7dB 0.5dB

[0138] As described above, according to some embodiments, channel estimator 200 uses A2C to improve NBCE performance. While supervised learning is effective in estimating near-ideal PDPs, even ideal PDPs cannot guarantee low block error rates in NBCE under the WSSUS model. Therefore, channel estimator 200 evaluates its strategy by comparing it with the results of the value network, thereby using A2C to train the strategy network. Thus, the channel estimator using A2C reduces (e.g., minimizes) the MSE of the channel estimate, which can lead to improved performance in terms of BLER.

[0139] The operations performed by the components of the transmitter 100 and the receiver 200 (e.g., performed by the channel estimators 100 and 200) can be performed by a "processing circuit", which can include any combination of hardware, firmware, and software for processing data or digital signals. The processing circuit hardware can include, for example, an application-specific integrated circuit (ASIC), a general-purpose or dedicated central processing unit (CPU), a digital signal processor (DSP), a graphics processing unit (GPU), and a programmable logic device such as a field programmable gate array (FPGA). In the processing circuit, as used herein, each function is performed either by hardware configured to perform the function (i.e., hard-wired) or by more general hardware (e.g., a CPU) configured to execute instructions stored in a non-temporary storage medium. The processing circuit can be manufactured on a single printed wiring board (PWB) or distributed on multiple interconnected PWBs. The processing circuit can contain other processing circuits. For example, the processing circuit can include two processing circuits, an FPGA, and a CPU, interconnected on a PWB.

[0140] As used herein, the singular forms "one" and "an" are intended to also include plural forms, unless the context clearly indicates otherwise. It will also be understood that, when used in this specification, the terms "include" and / or "comprise" specify the presence of the features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements and / or components and / or their combinations. As used herein, the term "and / or" includes any and all combinations of one or more associated listed items. When expressions such as "at least one" are before an element list, the entire element list is modified and the individual elements in the list are not modified. In addition, when describing an embodiment of the inventive concept, the use of "can" refers to "one or more embodiments of the present disclosure." Similarly, the term "exemplary" is intended to indicate an example or explanation. As used herein, the terms "use," "in use," and "used" can be considered to be synonymous with the terms "utilize," "utilize," and "utilize," respectively.

[0141] For the purpose of this disclosure, “at least one of X, Y, and Z” and “at least one selected from the group consisting of X, Y, and Z” may be interpreted as only X; only Y; only Z; or any combination of two or more of X, Y, and Z, such as XYZ, XYY, YZ, and ZZ.

[0142] Additionally, the use of “may” when describing embodiments of the inventive concept refers to “one or more embodiments of the inventive concept.” Likewise, the term “exemplary” is intended to indicate an example or illustration.

[0143] While the present invention has been shown and described with reference to embodiments thereof, it will be apparent to those skilled in the art that various suitable changes in form and details may be made therein without departing from the spirit and scope of the invention as defined in the following claims and their equivalents.

Claims

1. A method for channel estimation for a precoded channel, the method comprising: generating an initial frequency autocorrelation of the precoded channel for a current bundle of received data transmissions; Generate an extended frequency autocorrelation based on an initial frequency autocorrelation of the precoded channel, wherein the size of the extended frequency autocorrelation is a Fast Fourier Transform (FFT) size; Providing extended frequency autocorrelations to neural networks; generating an estimated frequency autocorrelation of a non-precoded channel based on the spread frequency autocorrelation by a neural network; and Based on the estimated frequency autocorrelation, an estimated power distribution profile of the non-precoded channel is generated.

2. The method according to claim 1, wherein The current bundle includes a plurality of resource blocks, each of the resource blocks including a plurality of subcarriers.

3. The method according to claim 1, wherein The non-precoded channel is an estimate of the precoded channel without precoding.

4. The method according to claim 1, wherein Generating a spread-frequency autocorrelation involves: performing edge extension on the initial frequency autocorrelation to extend the size of the initial frequency autocorrelation to the FFT size, Here, the FFT size is the input size of the neural network.

5. The method according to claim 4, wherein The edge extension includes linearly interpolating the values ​​of the initial frequency autocorrelation via an extension matrix.

6. The method according to claim 1, wherein Providing extended frequency autocorrelation to a neural network involves: The neural network is fed with the first half of the values ​​of the extended frequency autocorrelations, Here, the second half of the value of the spread frequency autocorrelation is the complex conjugate of the first half of the value of the spread frequency autocorrelation.

7. The method according to claim 1, wherein Generating the estimated frequency autocorrelation by the neural network includes: At least some values ​​of estimated frequency autocorrelations of the non-precoded channel are generated by a neural network based on the spread frequency autocorrelations.

8. The method according to claim 1, wherein Generating the estimated power distribution profile comprises: filtering the estimated frequency autocorrelation output by the neural network through a low-pass filter to generate a fine autocorrelation of the unprecoded channel; and An inverse FFT operation is performed on the fine autocorrelations to generate an estimated power distribution profile.

9. The method according to claim 8, wherein The low-pass filter is a moving average filter.

10. The method according to claim 1, wherein Generating an initial frequency autocorrelation of the precoded channel for the current bundle includes: generating a temporal autocorrelation for a previously bundled bundle of received data transmissions; generating a previous frequency autocorrelation for a previous bundle based on a previously estimated power distribution profile; generating an estimated channel input response based on the time autocorrelation and the previous frequency autocorrelation; and An initial frequency autocorrelation of the precoded channel is generated for the current bundling bundle based on the estimated channel input response.

11. The method according to claim 1 , further comprising: A truncated estimated power profile is generated by truncating the size of the estimated power profile to match the size of an initial frequency autocorrelation of the precoded channel.

12. The method according to claim 11, further comprising: The truncated estimated power distribution profile is normalized to unit power to generate a normalized estimated power distribution profile.

13. The method according to claim 11, wherein The truncated estimated power profile has a length of the maximum delay spread of the precoded channel.

14. A system for channel estimation of a precoded channel, the system comprising: processor; and Memory, which stores instructions that, when executed on a processor, enable the processor to: generating an initial frequency autocorrelation of the precoded channel for a current bundle of received data transmissions; Generate an extended frequency autocorrelation based on an initial frequency autocorrelation of the precoded channel, wherein the size of the extended frequency autocorrelation is a Fast Fourier Transform (FFT) size; Providing extended frequency autocorrelations to neural networks; generating an estimated frequency autocorrelation of a non-precoded channel based on the spread frequency autocorrelation by a neural network; and Based on the estimated frequency autocorrelation, an estimated power distribution profile of the non-precoded channel is generated.

15. A method for channel estimation for a precoded channel, the method comprising: generating an initial frequency autocorrelation of the precoded channel for a current bundle of received data transmissions; Generate an extended frequency autocorrelation according to an initial frequency autocorrelation of a precoded channel, wherein a size of the extended frequency autocorrelation is a Fast Fourier Transform (FFT) size; Provide extended frequency autocorrelation to the policy network; generating, by the strategy network, an estimated frequency autocorrelation of a non-precoded channel based on the extended frequency autocorrelation; The instantaneous reward is determined by the value network based on the estimated frequency autocorrelation; Determine advantage based on the instantaneous reward and the predicted total reward forward propagated on the value network; and Based on the advantage, the policy of the policy network is updated through back-propagation to reduce the block error rate.

16. The method according to claim 15, wherein The strategies for updating the strategy network include: Determine policy gradients based on advantages; and Update the coefficients of the policy network based on the policy gradient.

17. The method according to claim 15, wherein: The policy network and the value network are multi-layer perceptrons.

18. The method according to claim 15, further comprising: Gaussian noise is added to the estimated frequency autocorrelations to transform the discrete action space of the policy network into a continuous action space.

19. The method according to claim 15, further comprising: filtering the estimated frequency autocorrelation through a low-pass filter to generate a fine autocorrelation of the non-precoded channel; and An inverse FFT operation is performed on the fine autocorrelations to generate an estimated power distribution profile.

Images