Doppler extension estimation based on supervised learning

Abstract

This disclosure relates to supervised learning-based Doppler spread estimation, providing a radio receiver including channel estimator processing circuitry, wherein the channel estimator processing circuitry includes: a feature extractor configured to extract one or more features from a received signal, wherein the one or more features include channel correlation estimated based on a reference signal in a current time slot, the estimated channel correlation indicating the rate of change of the wireless channel over time; and a Doppler spread estimator configured to estimate the Doppler spread of the wireless channel by providing the one or more features to one or more Doppler shift predictors trained on training data spanning a training signal-to-noise ratio (SNR) range and a training Doppler shift range, wherein each Doppler shift predictor is trained on a portion of training data corresponding to a different portion of the training data.

Description

[0001] This application claims priority and benefit to U.S. Provisional Patent Application No. 63 / 004,894, filed April 3, 2020, with the United States Patent and Trademark Office, the entire disclosure of which is incorporated herein by reference. Technical Field

[0002] Various aspects of embodiments of this disclosure relate to wireless communication and supervised learning-based Doppler spread estimation. Background Technology

[0003] In the field of wireless communication, radio transceivers monitor the characteristics of the communication channel to adapt to changing signal propagation conditions in the electromagnetic environment. These characteristics are commonly referred to as channel state information (CSI) and can include effects such as scattering, fading, power attenuation, and Doppler spread. The channel estimation (CE) block of a radio transceiver can be used to estimate the CSI of various channels for received radio signals within the operating frequency range of the radio transceiver.

[0004] In the field of wireless communication, when a radio transmitter and a radio receiver move relative to each other, the radio receiver can receive a Doppler-shifted version of the transmitted signal. For example, in the case of a cellular terrestrial mobile radio system, the base station (e.g., a cell tower) is typically fixed, while one or more mobile stations (e.g., smartphones) communicating with the base station can be stationary or mobile. Generally, when a mobile station moves toward the base station, the frequency of the received signal will be shifted upwards (increased), and when the mobile station moves away from the base station, the frequency of the received signal will be shifted downwards (decreased). For example, when a mobile station is in a fast-moving car, the observed Doppler shift fo d The amplitude is typically greater than the Doppler shift amplitude when the mobile station is placed on an office desk. The broadening of the transmitted signal's spectrum due to the rate of change of the Doppler shift is called Doppler spread (D). s .

[0005] Doppler extension is commonly used for time interpolation in the channel estimation block of radio, as well as some software control, as part of a system for radio to adapt the transmission to the current channel conditions in order to achieve reliable communication. Summary of the Invention

[0006] Various aspects of embodiments of this disclosure relate to systems and methods for estimating Doppler spread based on supervised machine learning.

[0007] According to one embodiment of this disclosure, a method for estimating the Doppler spread of a wireless channel includes: extracting one or more features from a received signal by processing circuitry of a radio receiver, wherein the one or more features include an estimated channel correlation in a current time slot, the estimated channel correlation indicating the rate of change of the wireless channel over time; and calculating the Doppler spread of the wireless channel by the processing circuitry by providing the one or more features to one or more Doppler shift predictors trained on training data spanning a training signal-to-noise ratio (SNR) range and a training Doppler shift range, wherein each Doppler shift predictor is trained on a subset of training data corresponding to a different portion of the training data. The estimated channel correlation may include a single channel correlation filtered by an infinite impulse response.

[0008] The one or more features may include one or more estimated channel correlations based on one or more reference signals in one or more previous time slots.

[0009] The reference signal may be a tracking reference signal.

[0010] Each of the one or more Doppler frequency shift predictors may be trained based on different sub-ranges of the training SNR range, each sub-range having a lower limit and an upper limit, and the method may further include: determining the current SNR of the received signal; and selecting a Doppler frequency shift predictor from the one or more Doppler frequency shift predictors based on the current SNR, wherein the lower limit of the corresponding sub-range of the selected Doppler frequency shift predictor is higher than the current SNR.

[0011] The lower limit of the corresponding subrange of the selected Doppler frequency shift predictor is closest to the current SNR among the lower limits of the subranges above the current SNR.

[0012] Each of the one or more Doppler shift predictors may be trained based on different subranges of the training Doppler shift range, and the method may further include: the processing circuit calculating one or more classification probabilities by providing the one or more features to the Doppler shift classifier network, wherein each of the one or more classification probabilities corresponds to a different Doppler shift predictor among the one or more Doppler shift predictors, the one or more features may be provided to the one or more Doppler shift predictors to calculate one or more predicted Doppler shifts, and the step of calculating the Doppler spread may include: combining the one or more predicted Doppler shifts according to the one or more classification probabilities.

[0013] The step of combining the one or more predicted Doppler shifts may include summing the one or more predicted Doppler shifts multiplied by one or more products of the corresponding classification probabilities of the one or more classification probabilities.

[0014] The step of combining the one or more predicted Doppler shifts may include: outputting the predicted Doppler shift with the highest probability among the one or more predicted Doppler shifts that corresponds to the highest classification probability among the one or more classification probabilities.

[0015] The training SNR range of the training data can be greater than the operating SNR range of the radio receiver.

[0016] Each of the one or more Doppler shift predictors can be trained to calculate and predict the Doppler shift based on a regression model.

[0017] Each of the one or more Doppler shift predictors can be trained to classify the one or more features by calculating one or more probabilities corresponding to each of the one or more ranges of the Doppler shift.

[0018] Each of the one or more Doppler shift predictors can be a multilayer perceptron (MLP).

[0019] According to one embodiment of this disclosure, a radio receiver includes channel estimator processing circuitry, wherein the channel estimator processing circuitry includes: a feature extractor configured to extract one or more features from a received signal, wherein the one or more features include an estimated channel correlation in a current time slot, the estimated channel correlation indicating the rate of change of the wireless channel over time; and a Doppler spread estimator configured to estimate the Doppler spread of the wireless channel by providing the one or more features to one or more Doppler shift predictors trained on training data spanning a training signal-to-noise ratio (SNR) range and a training Doppler shift range, wherein each Doppler shift predictor is trained on a portion of training data corresponding to a different portion of the training data. The estimated channel correlation may include a single channel correlation filtered by an infinite impulse response.

[0020] The one or more features may include one or more estimated channel correlations based on one or more reference signals in one or more previous time slots.

[0021] The reference signal may be a tracking reference signal.

[0022] Each of the one or more Doppler frequency shift predictors may be trained based on different sub-ranges of the training SNR range, each sub-range having a lower limit and an upper limit, and the channel estimator processing circuit may further include: an SNR extractor configured to extract the current SNR of the received signal; and a predictor selector configured to select a Doppler frequency shift predictor from the one or more Doppler frequency shift predictors based on the current SNR, wherein the lower limit of the corresponding sub-range of the selected Doppler frequency shift predictor is higher than the current SNR.

[0023] The lower limit of the corresponding subrange of the selected Doppler frequency shift predictor is closest to the current SNR among the lower limits of the subranges above the current SNR.

[0024] Each of the one or more Doppler shift predictors may be trained based on different subranges of the training Doppler shift range, and the Doppler extension estimator may include a Doppler shift classifier network, wherein the Doppler shift classifier network is configured to calculate one or more classification probabilities that an input feature belongs to a category corresponding to the one or more Doppler shift predictors, the Doppler extension estimator may be configured to provide the one or more features to the one or more Doppler shift predictors to calculate one or more predicted Doppler shifts, and the Doppler extension estimator may be configured to calculate the Doppler extension by combining the one or more predicted Doppler shifts according to the one or more classification probabilities.

[0025] The Doppler extension estimator can be configured to combine the one or more predicted Doppler shifts by summing the products of the one or more predicted Doppler shifts multiplied by the corresponding classification probabilities of the one or more classification probabilities.

[0026] The Doppler extension estimator can be configured to combine the one or more predicted Doppler shifts by outputting the predicted Doppler shift with the highest probability among the one or more classification probabilities corresponding to the highest classification probability.

[0027] The training SNR range of the training data can be greater than the operating SNR range of the radio receiver.

[0028] Each of the one or more Doppler shift predictors can be trained to calculate and predict the Doppler shift based on a regression model.

[0029] Each of the one or more Doppler shift predictors can be trained to classify the one or more features by calculating one or more probabilities corresponding to each of the one or more ranges of the Doppler shift.

[0030] Each of the one or more Doppler shift predictors can be a multilayer perceptron. Attached Figure Description

[0031] The accompanying drawings, together with the specification, illustrate exemplary embodiments of the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0032] Figure 1 This is a schematic diagram of a mobile station.

[0033] Figure 2 This is a schematic block diagram of a wireless communication system in which a base station sends signals to a mobile station.

[0034] Figure 3 This is a block diagram of a Doppler extended predictor according to an embodiment of the present disclosure.

[0035] Figure 4 It is a graph comparing the relationship between channel correlation and Doppler shift when using a Bessel function or when using a multilayer perceptron according to an embodiment of the present disclosure as the mapping function.

[0036] Figure 5 This is a block diagram of a Doppler frequency shift estimator according to an embodiment of the present disclosure.

[0037] Figure 6 This is a flowchart illustrating a method for estimating Doppler spread using one or more Doppler frequency shift predictors according to an embodiment of the present disclosure.

[0038] Figure 7A This is a flowchart illustrating a method for selecting a Doppler shift predictor between two different Doppler shift predictors according to an embodiment of the present disclosure.

[0039] Figure 7B This is a flowchart illustrating a method for selecting at least one Doppler shift predictor among one or more different Doppler shift predictors according to an embodiment of the present disclosure.

[0040] Figure 8 This is a block diagram of a Doppler frequency shift estimator according to an embodiment of the present disclosure, wherein the Doppler frequency shift estimator is configured to estimate the Doppler frequency shift by combining predictions from one or more Doppler frequency shift predictors.

[0041] Figure 9 This is a block diagram of a Doppler frequency shift estimator according to an embodiment of the present disclosure, wherein the Doppler frequency shift estimator is configured to estimate the Doppler frequency shift by combining predictions from one or more Doppler frequency shift predictors based on confidence levels in the Doppler frequency shift predictors.

[0042] Figure 10 This is a schematic diagram of using an IIR filter to combine one or more channel-dependent Doppler shift predictors according to an embodiment of the present disclosure.

[0043] Figure 11 This is a schematic diagram of a Doppler frequency shift predictor according to an embodiment of the present disclosure, wherein one or more channel correlations from a window of a TRS time slot are provided as input to a multilayer perceptron. Detailed Implementation

[0044] In the following detailed description, only certain exemplary embodiments of this disclosure are shown and described by way of illustration. As those skilled in the art will recognize, this disclosure may be implemented in many different forms and should not be construed as limited to the embodiments set forth herein.

[0045] For clarity, aspects of embodiments of this disclosure will be described herein in the context of a cellular modem's radio transceiver. However, embodiments of this disclosure are not limited thereto, and those skilled in the art will understand prior to the effective filing date of this application that embodiments of this disclosure can also be applied to estimating Doppler spread in other contexts.

[0046] In some radio transmission standards (such as the 5G New Radio (NR) standard), cellular radio transceivers can use the tracking reference signal (TRS) received from the base station to estimate the Doppler spread D. s For example, one technique calculates an estimated channel correlation between the channels corresponding to two tracking reference signals, and then feeds the estimated channel correlation to an inverse Bessel function to obtain the Doppler spread D. s .

[0047] However, this method assumes that the relationship between channel correlation and Doppler spread follows an inverse Bessel function. In reality, the channel does not always exhibit an inverse Bessel function relationship between channel correlation and Doppler spread. Furthermore, when calculating channel correlation, noise power needs to be removed from the channel power, so the accuracy of the channel correlation estimate is sensitive to the estimation of noise variance. In addition, in some cases, there may only be one TRS slot per TRS cycle (two TRS symbols (symbo1)), so only one correlation value can be used for Doppler spread estimation, resulting in low resolution of the Doppler spread estimate. For example, under the 5G NR standard, when operating in FR2 (frequency range 2, encompassing bands in the millimeter wave range from 24 GHz to 100 GHz), only one TRS slot may be transmitted per TRS cycle.

[0048] Therefore, aspects of embodiments of this disclosure relate to using supervised machine learning to estimate Doppler spread. More specifically, some aspects of embodiments of this disclosure relate to using supervised learning (e.g., using a machine learning model such as a multilayer perceptron (MLP) neural network or other neural networks) to learn a mapping function (or “Doppler spread predictor”) from estimated channel correlation to Doppler spread, wherein the mapping function is trained on collected experimental data that correlates channel correlation with Doppler spread. During online prediction, the estimated correlation is fed to the learned mapping function (Doppler spread predictor) to generate the estimated Doppler spread.

[0049] The mapping function (or Doppler spread predictor) according to embodiments of this disclosure is trained based on data collected from operating radio communication systems and thus matches the actual behavior of these operating systems, rather than relying on specific assumptions about the operating environment and the behavior of those systems. Furthermore, given a sufficiently large and diverse training dataset, the trained Doppler spread predictor according to embodiments of this disclosure is able to summarize and produce robust (e.g., accurate) estimates of the Doppler spread over a range of different operating conditions (e.g., different SNRs spread over the training signal-to-noise ratio (SNR) range), thereby enabling embodiments of this disclosure to compensate for noise variance. According to some aspects of embodiments of this disclosure, the Doppler spread predictor also utilizes one or more previously estimated channel correlations (e.g., from previous TRS cycles) to improve the estimate of the Doppler spread at the current TRS cycle.

[0050] Figure 1 This is a schematic diagram of a mobile station. (For example...) Figure 1As shown, mobile station 10 may include an antenna 11 configured to receive (e.g., transmitted by a base station) electromagnetic signal 30. The received signal may be provided to a receive filter 12 (e.g., a bandpass filter), and the filtered signal may be provided to detector 14 and channel estimator 16. Channel estimator 16 may generate channel state information (CSI) for controlling detector 14 and other components of mobile station 10 to adapt to changing conditions in the environment, such as movement of mobile station 10 relative to base station and / or changes in the environment through which electromagnetic signal 30 propagates. According to some embodiments of this disclosure, channel estimator 16 communicates with or includes Doppler spread estimator 100. The output of channel estimator 16 (which may include the output of Doppler spread estimator 100 or may include information calculated based on the output of Doppler spread estimator 100) is provided to detector 14 that performs symbol detection using channel estimation. Decoder 18 can be configured to receive detected symbols from detector 14 and decode the detected symbols into data 50 (such as a digital bitstream) for consumption by applications in mobile station 10 (such as voice calls, data packets, etc.). In various embodiments of this disclosure, components of mobile station 10 (such as filter 12, detector 14, channel estimator 16, Doppler spread estimator 100, and decoder 18) can be implemented in one or more processing circuits of a digital radio (e.g., a radio baseband processor (BP or BPP), a central processing unit (CPU), a microcontroller, a digital signal processor (DSP), a field-programmable gate array (FPGA), or an application-specific integrated circuit (ASIC)). Various portions of these blocks can be implemented in the same circuit (e.g., on the same die or in the same package) or in different circuits (e.g., on different dies or in different packages, connected via a communication bus).

[0051] Figure 2 This is a schematic block diagram of a communication system in which base station 20 transmits signal 30 to mobile station 10, wherein mobile station 10 includes a Doppler spread estimator 100. The magnitude of the Doppler frequency shift in the signal received by mobile station 10 may depend on the relative motion (e.g., velocity) of mobile station 10 relative to base station 20, and therefore the Doppler spread in the signal received by base station 20 may also depend on the relative motion of mobile station 10 relative to base station 20.

[0052] As mentioned above, some comparison systems, based on the assumption that channel statistics follow the Jakes model, use the inverse Bessel function to calculate the Doppler extension from the estimated channel correlation. For example, based on the Jakes model, the channel autocorrelation function has the form of Equation 1:

[0053] R(τ,f d )=J0(2πf dτ) (1)

[0054] Where τ represents the time difference, f d The maximum Doppler frequency shift is represented as defined in Equation 2 below:

[0055]

[0056] Where c represents the speed of light, f represents the frequency of the transmitted signal 30, and v is the speed of the mobile station 10. Formally, the Doppler spread is defined as D... s =f d -(-f d )=2f d As used herein, the terms Doppler spread estimation and maximum Doppler shift estimation are used interchangeably, both referring to the aforementioned f d The estimate is given by J0(·), which represents the zeroth-order Bessel function of the first kind, as shown in Equation 3.

[0057]

[0058] Note the channel autocorrelation function R(τ, f) mentioned above. d This is derived based on the assumption of the Jakes channel model. Therefore, when using the inverse Bessel function to describe channel correlation and Doppler shift f... d When considering the relationship between channel correlation and Doppler spread, the Jakes channel model is implicitly assumed. This model assumes validity for some types of channel models (such as the Extended Pedestrian A Model (EPA), Extended Vehicle A Model (EVA), Extended Typical City Model (ETU), and Tapped Delay Line (TDL) models), but may be invalid for others (e.g., the Clustered Delay Line (CDL) model). In practice, due to non-ideal estimations of channel correlation, such as those for the inverse Bessel function, the estimated channel correlation may not accurately describe the relationship between the estimated channel correlation and Doppler spread. For example, the number of reference signal elements may be insufficient to calculate an accurate average of the channel correlation, or the noise variance calculation may be too inaccurate.

[0059] Figure 3 This is a block diagram of a Doppler extension estimator 100 according to an embodiment of the present disclosure. Figure 3As shown, according to one embodiment of this disclosure, a Doppler spread estimator 100 is configured to receive an input estimated channel. A feature extractor 110 is configured to extract features from the provided channel, wherein the features may include, for example, estimated channel correlations. The extracted features are provided to a trained Doppler frequency shift predictor 120, wherein the trained Doppler frequency shift predictor 120 is configured to compute an estimated Doppler frequency shift 300 value based on the extracted features. The trained Doppler frequency shift predictor 120 according to an embodiment of this disclosure is constructed based on training data collected from a real physical system or a practical link-level simulator, and can therefore be trained to model the behavior of the system more accurately than the inverse Bessel function.

[0060] According to some embodiments of this disclosure, feature extractor 110 is configured to extract features from the input estimated channel. Specific features include information decoded or computed from various characteristics of the input estimated channel. In some embodiments of this disclosure, feature extractor 110 is configured to extract channel correlation C(T) from the input estimated channel. These channel correlations can be computed based on the TRS signals of one or more TRS time slots in the received signal. Typically, the estimated channel... The channel correlation C(T) is given by Equation 4:

[0061]

[0062] in, σ represents the channel estimated at symbol time t and subcarrier k. 2 γ represents the noise variance contained in the estimated channel, and γ∈[0,1] is a configurable parameter used to adjust how much noise is subtracted. T represents the time difference between two TRS symbols within a TRS slot or across two consecutive TRS slots. If two TRS slots are allocated to each TRS period, channel correlation can be calculated between pairs of TRS symbols within a given TRS slot or across two consecutive TRS slots.

[0063] In practice, it may be difficult to accurately estimate the noise variance σ. 2 Furthermore, it may be difficult to choose an appropriate value for the noise reduction parameter γ. According to some embodiments of this disclosure, γ is set to 0 (zero) to eliminate the influence of noise variance estimation. Additionally, in some cases, using the sample mean to implement the expectation operation E{·} is more practical. Therefore, in some embodiments of this disclosure, the channel correlation C(T) is implemented according to Equation 5:

[0064]

[0065] Where K is the number of subcarriers.

[0066] In addition to calculating the current channel correlation based on the TRS symbols of the current TRS slot, feature extractor 110 may also calculate additional features. For example, in some embodiments of this disclosure, feature extractor 110 may further calculate features based on channel correlations estimated from previous TRS periods (e.g., feature extractor 110 may include a memory storing a window of channel correlations previously calculated for earlier TRS periods).

[0067] According to some embodiments of this disclosure, the Doppler spread predictor is trained based on actual measurement data from a running wireless communication system or from a real link-level simulator. Therefore, in some embodiments of this disclosure, the training data may include data from various parameters such as signal-to-noise ratio (SNR), channel (e.g., EPA, EVA, TDL-A, TDL-D), digital port configuration (e.g., 1x2, 1x4, 1x8), analog antenna configuration (e.g., 2, 4, 8), and Doppler frequency shift f. d The transmission data captured between the transmitter and receiver is based on the modulation and coding scheme (MCS), subcarrier spacing (SCS), and Fast Fourier Transform (FFT) size. For each specific combination of these parameters (e.g., for each specific set of transmission settings), training data is recorded during transmission. Each sample in the training data includes an input and an output, where the input includes features extracted by the feature extractor 110, and the output includes the Doppler frequency shift fd that occurred during transmission (e.g., the Doppler frequency shift fd measured in the case of the operating system). d Or, in the case of a simulator, the Doppler frequency shift f can be configured. d The final training dataset contains a large amount of training data collected from the transmission under various combinations of different parameters.

[0068] The Doppler shift predictor 120 may include a neural network (e.g., a multilayer perceptron (MLP), a recurrent neural network (RNN), a long short-term memory (LSTM) network, or other forms of machine learning model. For illustrative purposes, embodiments of the present disclosure will be described in more detail with respect to the use of a multilayer perceptron (MLP) as a machine learning model, but embodiments of the present disclosure are not limited thereto.

[0069] When applying supervised learning, the machine learning model is typically trained by adjusting multiple parameters of the model (e.g., the weights of connections between neurons in a neural network) to minimize the channel estimation by the model based on the input (e.g., the channel estimated by the feature extractor 110 from the input). The extracted features) are used to calculate the value and compare it with the ground truth (e.g., the Doppler frequency shift f associated with channel h or configuration). dThe cost function is defined as the sum of the parameters of a given model and the parameters of a given model. In the case of neural networks, the training process may include applying a backpropagation algorithm with gradient descent to iteratively update the model's parameters to minimize the cost function.

[0070] The collected training data consists of many pairs of data samples, where each pair contains some input (e.g., C(T) (or some other input features)) and output (ground true Doppler spread f). d During offline training of the Doppler frequency shift predictor 120 (e.g., a multilayer perceptron), the Doppler frequency shift predictor 120 is trained to map the input (e.g., the channel correlation C(T) extracted from the signal by the feature extractor 110) to the output (the ground true Doppler spread f). d During online prediction (e.g., when the Doppler shift predictor 120 is deployed in the mobile station 10 to calculate channel state information), the current channel correlation C(T) is calculated and C(T) is input into the Doppler shift predictor 120 to calculate the estimated Doppler spread.

[0071] More specifically, when considered within the learning framework, at least due to the estimated Doppler frequency shift f d Since it is a single continuous value, the Doppler extension estimation can be formulated as a regression problem. Therefore, when the training of the Doppler frequency shift predictor 100 is considered as a regression problem, the optimization process of training a machine learning model based on a regression model can be performed by minimizing the sum of squared errors with respect to the parameter θ, as shown in Equation 6:

[0072]

[0073] The Doppler extended prediction function can be expressed as F θ θ represents the learning parameters of the Doppler spread prediction function, the input features are represented as g, and the true Doppler spread is represented as f. d In the above text, it is assumed that the training dataset contains N input / output pairs {g}. n f d，n}, n = 1, ..., N.

[0074] However, the Doppler spread can span a wide range, potentially reaching several kilohertz at FR2 (e.g., millimeter-wave frequencies). Therefore, the estimated Doppler frequency shift f d It can also span a wide range, for example, f d ∈[0, 2000]. If the training of the machine learning model is based on solving the above sum of squares error minimization function, then the training samples corresponding to the small Doppler expansion will be weakened, because for small f d，n error (F) θ (g n )-fd，n ) 2 Usually smaller than for large f d，n error (F) θ (g n )-f d，n ) 2 When the actual Doppler extension is small, this will result in a very inaccurate Doppler extension estimate.

[0075] Therefore, in some embodiments of this disclosure, the process of training a machine learning model based on a regression model can be performed by minimizing the normalized sum of squares error with respect to the parameter θ, as shown in Equation 7:

[0076]

[0077] By using the normalized sum of squared errors described above as the cost function, the training samples corresponding to the Doppler expansion will be weakened because the error (F) θ (g n )-f d，n ) 2 Divide by Therefore, targeting large f d，n The cost is smaller (e.g., the cost function is smaller relative to f). d，n (Normalized)

[0078] In some embodiments of this disclosure, Doppler expansion prediction is treated as a classification problem rather than a regression problem by quantizing the Doppler expansion range into multiple small regions or ranges. Each region of the Doppler expansion is considered a class and is represented by a single Doppler expansion (e.g., the median Doppler expansion of that region). A Doppler expansion predictor is then trained to map the input feature g to the probability or confidence of each class (e.g., compute multiple probabilities, where each probability represents the confidence of mapping the input feature to a corresponding region among the multiple regions of the Doppler expansion). In some embodiments, a final Doppler expansion estimate is subsequently computed by representing the Doppler expansion of each class based on a combination of probabilities for each class, as discussed in more detail below.

[0079] Table 1 below presents an example of dividing the Doppler frequency shift range into M different categories. More specifically, assume the entire range of the Doppler frequency shift in the training data is f. d ∈[r0, r M If the entire range can be divided into M non-overlapping continuous regions for a total of M categories, then the m-th region can be divided into M non-overlapping continuous regions for a total of M categories. For example, the m-th region corresponds to the Doppler shift range [r]. m-1 r m Correspondingly. Each category is determined by the corresponding Doppler frequency shift. Let m = 1, ..., M, where m = 1, ..., M. When preparing the training dataset, the Doppler shift (e.g., f) associated with each training data sample is... d，n The bins are assigned to the corresponding region among the M regions (e.g., finding the m-th region such that f d,n ∈[r m-1 r m This allows the training sample to be classified into class m. During training, cross-entropy is used as the cost function.

[0080] Table 1

[0081]

[0082]

[0083] Given input features, the Doppler shift predictor trained as a classifier will output a function that satisfies... The M-dimensional vector [c1, ..., c M ] T , where c m This represents the probability (or confidence level) that the input feature belongs to category m. The cost function can be expressed as follows, where the mapping function or Doppler shift predictor F is characterized by the coefficient θ. θ The input feature g of the nth training data sample n Mapping to an M-dimensional vector [c n,1 c n,2 c n,M The nth training data belongs to category v. n , making As shown in Equation 8:

[0084] stF θ (g n )=[c n,1 ,c n，2 c n,M ] T ,

[0085] When performing inference or prediction, the Doppler shift predictor trained as a classifier generates predictions for M classes c1, ..., c2. M The predicted probability for each category. Given the representative Doppler shift for each category. The final Doppler shift estimate can be obtained by summing the products of each representative Doppler shift multiplied by its corresponding prediction probability, as shown in Equation 9:

[0086]

[0087] Alternatively, it can be represented by selecting the “maximum combination” of representative Doppler extensions corresponding to the highest predicted probability, as shown in Equation 10:

[0088]

[0089] Where I(·) represents the indicator function as shown in Equation 11:

[0090]

[0091] However, as described above, in some embodiments of this disclosure, γ is set to 0 when the feature extractor 110 calculates the estimated channel correlation C(T). Therefore, the estimated channel correlation calculated by the feature extractor 110 may be smaller than the true channel correlation (because setting γ to zero in Equation 4 makes the denominator larger). The degree to which the estimated channel correlation C(T) is smaller than the true channel correlation is more pronounced at lower SNRs (e.g., at lower SNRs, there is a greater difference between the estimated channel correlation and the true channel correlation). This is because lower SNRs are associated with larger noise variance σ. 2 Accordingly, since the term -γ·σ 2 Setting it to zero, the lower the SNR, the greater the increase in the denominator (e.g., more noise variance σ should be subtracted from the estimated channel power in the denominator). 2 However, setting γ=0 leads to an even larger denominator. As a result, if the working SNR during online prediction differs from the SNR of the training dataset, then even for the same true Doppler extension f... d Furthermore, the estimated channel correlation C(T) during online prediction will differ from the C(T) for the training set. In other words, due to the mismatch between the SNR during operation and the SNR of the offline training dataset, the mapping function learned based on the offline training data may not be suitable for online prediction.

[0092] Figure 4 This is a graph showing the relationship between channel correlation and Doppler shift when using a Bessel function or when using a multilayer perceptron according to an embodiment of this disclosure as the mapping function. (Refer to...) Figure 4 Suppose we learn an MLP from a set of training data such that it is related to the Doppler frequency shift f d The corresponding channel correlation is C′. Then, during online estimation, if the working SNR is higher than the SNR range of the training dataset, the estimated channel correlation (denoted as C′) is... The estimated Doppler frequency shift (expressed as) will be greater than C′. The frequency shift will be less than the true Doppler frequency shift f. d On the other hand, if the working SNR is lower than the SNR range of the training dataset, the estimated channel correlation (denoted as...) The value will be less than C′, and the corresponding estimated Doppler frequency shift (expressed as...) This will be greater than the true Doppler frequency shift f. d .like Figure 4 As shown, due to the mismatch between the SNR of the training dataset and the online prediction, the estimated Doppler shift (e.g., and It may deviate from the true value f d Therefore, when the actual SNR is close to the SNR of the data used to train the Doppler shift predictor 120, the predicted estimated Doppler shift f calculated by the Doppler shift predictor 120... d More accurate.

[0093] In practice, the operating SNR range of a wireless transceiver can be wide, and it can be difficult to train a single Doppler shift predictor that operates across the entire operating SNR. Therefore, in some embodiments of this disclosure, the training data is divided into multiple subsets (e.g., R subsets or subranges), each subset corresponding to a different portion of the entire SNR range (or training SNR range) of the training data, and each subset or subrange of the training data is used (e.g., in a manner substantially similar to that discussed above) to train a separate Doppler shift predictor P for the corresponding r-th portion of the operating SNR range. r (Each Doppler shift predictor may have the same architecture or a different architecture).

[0094] In some embodiments of this disclosure, each subset has the same size, or may otherwise be uniformly spaced along the training SNR range of the training data on a linear or logarithmic scale. In other embodiments of this disclosure, the subsets have different sizes (e.g., non-uniformly spaced along the training SNR range). For example, the training data may be partitioned into subsets such that there are more predictors trained to provide predictions of Doppler shift for practically more frequently observed portions of the working SNR range (e.g., causing the Doppler shift estimator to produce more accurate results more frequently). As another example, the training data may be partitioned into subsets such that there are more predictors in regions of the working SNR range that are more sensitive to inaccuracies in the estimated Doppler shift or estimated Doppler spread.

[0095] Figure 5 This is a block diagram of a Doppler extension estimator 100 according to an embodiment of the present disclosure. Figure 5 The Doppler extended estimator 100 is basically similar to Figure 3 The Doppler frequency shift estimator shown also includes a signal-to-noise ratio (SNR) extractor 130 and a predictor selector 140. Figure 6This is a flowchart illustrating a method for estimating Doppler spread using multiple Doppler frequency shift predictors according to an embodiment of the present disclosure. (Refer to...) Figure 6 In operation 610, feature extractor 110 estimates the channel from the input. Feature extraction is performed as described above and in more detail below. In operation 630, the SNR extractor 130 extracts the SNR of the received signal and provides the extracted SNR to the predictor selector 140, wherein the predictor selector 140 is configured to select from R trained Doppler frequency shift predictors P (e.g., Doppler frequency shift predictors P1, P2, ..., P...). R Select a specific Doppler frequency shift predictor P) r For example, the training SNR range for each Doppler frequency shift predictor P can be set as shown in Table 2 below:

[0096] Table 2

[0097]

[0098] Generally, more accurate channel state information (CIS) leads to higher performance in radio receivers because it allows radios to tune their parameters to match the actual conditions of the channel. However, due to various environmental conditions, channel estimators may overestimate or underestimate various parameters of the CIS, including Doppler spread.

[0099] Based on experimental observations, overestimation of the Doppler spread leads to better block error rate (BLER) performance (e.g., lower error rate) than underestimation of the Doppler spread. In some experiments, setting the estimated Doppler shift to 25% higher (1.25 * 900 Hz = 1,125 Hz) yielded better BLER performance than the actual Doppler shift when the true Doppler shift was 900 Hz. d Slightly better performance, while setting the estimated Doppler frequency shift to 25% lower (0.75*900Hz=675Hz) results in a higher error rate.

[0100] In addition, such as Figure 4 As shown, when the working SNR is lower than the SNR range of the training dataset (related to the estimated channel correlation)... Correspondingly, the estimated Doppler spread This will be greater than the true Doppler extension f. d In other words, a Doppler spread predictor trained on data with a higher SNR than the provided input channel will lead to an overestimation of the Doppler spread.

[0101] Therefore, in some embodiments of this disclosure, the predictor selector 140 of the Doppler spread estimator 100 selects a Doppler shift predictor 122, wherein the Doppler shift predictor 122 is biased towards overestimation (or over-estimation) of the Doppler spread and away from underestimation (or under-estimation) of the Doppler spread in order to improve the BLER performance of the radio. In some embodiments, the bias is achieved by selecting one Doppler shift predictor from R Doppler shift predictors, wherein the Doppler shift predictor is trained on a subset of a dataset having an SNR range that is adjacent to and higher than the current working SNR as determined by the SNR extractor 130 (e.g., the next SNR range having a lower bound higher than the currently estimated SNR).

[0102] Figure 7A This is a flowchart illustrating a method for selecting a Doppler shift predictor between two different Doppler shift predictors according to an embodiment of the present disclosure. Figure 7A Corresponding to the case of R=2, where the first Doppler frequency shift predictor P1 ow The second Doppler frequency shift predictor P is trained on low SNR data (e.g., a subset of data collected using low SNR signals in the range [snr1, snr2]). high It is trained on high SNR data (e.g., a subset of data collected using high SNR signals in the range [snr2, snr3]). It is assumed that the operating SNR range for the radio is smaller than that for the second Doppler shift predictor P trained on the high SNR range. high The lower end of the SNR range (e.g., snr2). Therefore, refer to Figure 7A In operation 651, predictor selector 140 determines the current estimate. Is it less than the threshold SNR (SNR) thr Based on the example range given above, in some embodiments, SNR thr =snr1, in other words, the threshold SNR is the lower end of the SNR range of a predictor trained with low SNR signals. When the current estimate... Less than the threshold SNR (SNR) thr During operation 652, predictor selector 140 selects the first Doppler shift predictor P trained on low SNR data. low (This is because the predictor is trained on data that is closest to the current estimated SNR, and also on data with an SNR higher than the current estimated SNR). When the current estimated SNR... Not less than the threshold SNR (SNR) thr (For example, greater than or equal to the threshold SNR) thrDuring operation 653, predictor selector 140 selects the second Doppler shift predictor P trained on high SNR. high .

[0103] Figure 7B This is a flowchart illustrating a method for selecting at least one Doppler frequency shift predictor among a plurality of different Doppler frequency shift predictors (e.g., R different Doppler frequency shift predictors) according to an embodiment of the present disclosure. As described above, it is assumed that the operating SNR range of the system (e.g., a radio receiver) is from snr0 to snr R Furthermore, the entire working SNR range is divided into R regions. In operation 654, predictor selector 140 determines the current estimate. Is it in the first SNR range ([snr0, snr1), or is it...? If so, then in operation 655, predictor selector 140 selects a first predictor P1 trained based on training data from a second SNR range [snr1, snr2) that is adjacent to and higher than the first SNR range [snr0, snr1). If the current estimate is within the range... If the value is not within the first SNR range ([snr0, snr1)), then in operation 656, predictor selector 140 determines the current estimate. Does it fall within the second SNR range [snr1, snr2)? If so, in operation 657, predictor selector 140 selects a second predictor P2 trained based on training data from a third SNR range [snr2, snr3) that is adjacent to and higher than the second SNR range [snr1, snr2). In summary, if the current estimate... If the value is not within the first SNR range ([snr0, snr1)), then the predictor selector 140 will select the current estimated value. The process continues in a similar manner, comparing each of the remaining regions of the R regions within the working SNR range to select the appropriate predictor from the predictors. In operation 658, the final SNR range is [snr]. R-1 ,snr R ), and in operation 659 selection (based on the last SNR range [snr R The final Doppler frequency shift predictor P was trained using training data from (∞). R As mentioned above, assume the working range is from snr0 to snr. R And therefore greater than snr R The SNR will not be observed or will exceed the operating specifications.

[0104] Return to reference Figure 5 and Figure 6 After predictor selector 140 has selected a predictor in operation 650, the selected Doppler frequency shift predictor P is used in operation 670. r To estimate the Doppler frequency shift in operation 670 to calculate the estimated Doppler frequency shift As described above, in some embodiments, the estimated Doppler spread according to And compared to the largest estimated Doppler shift Related. Therefore, aspects of embodiments of this disclosure relate to the ability to estimate the channel based on the input. The Doppler spread estimator 100 is used to estimate the Doppler frequency shift by extracting the input features.

[0105] In some embodiments of this disclosure, training is performed by combining training data based on different portions of the Doppler frequency shift range (as opposed to training on different portions of the SNR range, as described above regarding...). Figure 5 , 6 In the cases described in embodiments 7A and 7B, predictions made by multiple Doppler frequency shift predictors are used to estimate a single estimated Doppler frequency shift (or Doppler spread). For example, the training data described above can be divided into R distinct subsets by dividing the entire Doppler frequency shift range into R distinct subranges, wherein each subset of the training data includes data from a corresponding subrange of the distinct subranges of the Doppler frequency shift range. For example, the training Doppler frequency shift range for each Doppler frequency shift predictor P can be set as shown in Table 3 below:

[0106] Table 3:

[0107]

[0108]

[0109] As shown in Table 3, each of the R Doppler frequency shift predictors is configured to calculate the corresponding output Doppler frequency shift f′ based on the input features. For example, the r-th predictor P r The predicted Doppler frequency shift f′ is calculated from its corresponding training Doppler frequency shift range. r (For example, where f′ r ∈[f dr f d(r+1) While Table 3 shows embodiments where the Doppler frequency shifter ranges do not overlap, the embodiments disclosed herein are not limited thereto. For example, in some embodiments, adjacent Doppler frequency shifter ranges corresponding to adjacent predictors have some overlap (e.g., from range [f]). d1a f d1bThe predictor P1 is trained on data from the range [f] and can be trained on data from the range [f] d2a f d2b Train a predictor P2 on the data, where f d2a <f d1b ).

[0110] Figure 8 This is a block diagram of a Doppler frequency shift estimator according to an embodiment of the present disclosure, wherein the Doppler frequency shift estimator is configured to estimate the Doppler frequency shift by combining predictions from multiple Doppler frequency shift predictors trained on different portions of the Doppler frequency shift range. Figure 8 In the illustrated embodiment, the Doppler frequency shift predictor 120 includes R trained Doppler frequency shift predictors. Input features are provided to each of the R trained Doppler frequency shift predictors to compute R predicted Doppler frequency shifts f′1, ..., f′. R Additionally, the input features are provided to a Doppler shift predictor classifier network (or Doppler shift classifier network) P0, which is trained to calculate the probability that the input features belong to each of R categories (e.g., the probability that the input features fall within each of R sub-ranges of the Doppler shift, or to predict which of the R Doppler shift predictors will predict the most accurate Doppler shift for a given input feature). The output of the classification-based network P0 is an R-dimensional vector [c1,...,c...]. R ], where each value c r This represents the probability or confidence level of the input feature corresponding to the r-th category (e.g., corresponding to the r-th Doppler shift predictor). The outputs of the R Doppler shift predictors are then combined by the combiner 810 using, for example, an average combination calculated based on the sum of the products of each predicted Doppler shift f′ multiplied by its corresponding predicted probability (as shown in Equation 12), or by selecting the largest combination of predicted Doppler shifts corresponding to the highest predicted probability (as shown in Equation 13).

[0111]

[0112] Where I(·) represents the indicator function.

[0113] Figure 9 This is a block diagram of a Doppler frequency shift estimator according to an embodiment of the present disclosure, wherein the Doppler frequency shift estimator is configured to estimate the Doppler frequency shift by combining predictions from a plurality of Doppler frequency shift predictors through an average combination. Figure 9 As shown, R predicted Doppler frequency shifts f′1, f′2, ..., f′ are used. R Multiply by their corresponding probabilities or confidence levels c1, c2, ..., c RAnd sum the products to calculate the estimated Doppler frequency shift.

[0114] As described above, the input features extracted from the input estimation channel may include, for example, channel features estimated from the received TRS symbols. Calculate the current channel correlation C(T).

[0115] As described above, in the comparison system used to estimate Doppler spread, the estimated channel correlation is fed to the inverse Bessel function to obtain the estimated Doppler spread. In practice, to obtain a more stable estimate of the channel correlation C(T), an infinite impulse response (IIR) filter is applied to the estimated channel correlation in each TRS cycle, resulting in an IIR-filtered channel correlation. Because the channel correlation directly measures the change in channel h, in some embodiments of this disclosure, the IIR-filtered channel correlation... The input features are provided to the Doppler frequency shift predictor P (e.g., the currently selected Doppler frequency shift predictor among a plurality of Doppler frequency shift predictors).

[0116] Similarly, an IIR filter can be applied to stabilize the estimated channel correlation, so the final input feature of the Doppler shift predictor is only one value – the estimated channel correlation after IIR filtering. In some embodiments of this disclosure, the Doppler shift predictor is a multilayer perceptron (MLP). Figure 10 This is a schematic diagram illustrating the use of IIR filters to combine multiple channel-correlated Doppler frequency shift predictors according to an embodiment of this disclosure. For example, as... Figure 10 As shown, assuming the current TRS period is the nth TRS period, there exist n estimated channel correlations C1(T), C2(T), ..., C1010, which are provided as inputs to the infinite impulse response (IIR) filter 1010. n (T), and then the IIR-filtered channel correlation of these n inputs can be expressed as: exist Figure 10 In one embodiment, the Doppler frequency shift predictor is implemented as a multilayer perceptron configured to perform regression, wherein the MLP has an input layer 123 having a single node, the input layer 123 being configured to receive the input IIR-filtered channel correlation. The filtered channel correlation is then provided to hidden layer 125, which has multiple nodes associated with multiple weights (or parameters). At each node, the input IIR-filtered channel correlation is... The inputs are multiplied by the corresponding weights, and the product is passed through an activation function (e.g., a sigmoid function or a rectified linear unit (ReLU)). The output layer 127, with a single node, is configured to receive inputs from multiple nodes in the hidden layer and combine them (e.g., multiply the outputs of the activation functions of the hidden layer nodes by the weights, sum the results, and pass them through an activation function to compute the predicted Doppler shift f). d，n ).

[0117] Because the IIR filter coefficients are set to fixed values, the way feature extractor 110 combines previously estimated channel correlations with currently estimated channel correlations is fixed in its design and may not adapt to changing conditions or other factors. Furthermore, although channel correlations are estimated in each TRS cycle, in this setup, the final input to the Doppler shift predictor is only an IIR-filtered channel correlation, thus losing some information that might be included in previously estimated channel correlations.

[0118] Therefore, some aspects of embodiments of this disclosure involve using the currently estimated channel correlation C n (T) and channel correlations from windows of multiple previous TRS cycles are provided as input features to the Doppler frequency shift predictor.

[0119] Figure 11 This is a schematic diagram of a Doppler frequency shift predictor according to an embodiment of the present disclosure, wherein multiple channel correlations from a window of TRS time slots are provided as input to a multilayer perceptron. Figure 11 In the illustrated embodiment, the C from the previous TRS cycle n-Δ (T), ..., C n-2 (T), C n-1 The multiple channel correlations of a causal window (T) (where Δ is the number of previous TRS periods, or equivalently, the size of the window in TRS periods) and in the current TRS period C nThe estimated channel correlations in (T) are combined to obtain a total of Δ+1 channel correlations included in the input features provided to the input layer 123 of the multilayer perceptron, wherein the input layer 123 includes a separate node for each of the Δ+1 channel correlations. Each channel correlation in C(T) is multiplied by a corresponding weight (e.g., weights learned during the training process) and provided from the nodes of the input layer 123 to each node of the hidden layer 125. At each node of the hidden layer 125, all incoming products (e.g., the product of channel correlations and weights) are summed and passed through an activation function (e.g., the sigmoid function). Each node in the hidden layer 125 provides the output from the activation function to the output layer 127, whereby the output layer 127 combines the outputs of the activation functions in the hidden layer 125 to compute the predicted Doppler frequency shift fd n. This combination operation multiplies the output of the activation function of each node in the hidden layer with their respective weights (which are learned through the training process), sums the weighted products, and passes the sum through an activation function (e.g., a sigmoid function or a rectified linear unit (ReLU)).

[0120] By incorporating estimated channel correlations from previous ΔTRS periods, these embodiments of the present disclosure provide the Doppler shift predictor with more information about how the channel changes over time. Furthermore, the training process trains the Doppler shift predictor to combine these estimated channel correlations using a learned set of parameters or coefficients, rather than fixing those coefficients according to an IIR filter. Because the learned parameters can compute the same results as an IIR filter (e.g., the learned coefficients might lead to an IIR filter), it is expected that a trained Doppler shift predictor using multiple channel coefficients from a causal window of previous TRS periods will perform no worse than an IIR filter.

[0121] Those skilled in the art will understand that modifications are possible. Figure 10 and Figure 11 The architecture shown in the embodiment for performing regression is to perform classification into one of M categories of the Doppler shift range as described above by using M nodes in the output layer 127 and encoding the correct category of the training data using one-hot encoding.

[0122] As described above, some embodiments of this disclosure involve using a multilayer perceptron as a Doppler shift predictor for a neural network to predict based on provided input features (such as current channel correlation C). n (T) and C from multiple previous TRS cycles n-Δ (T), ..., C n-2 (T), C n-1The channel correlation of the window (T) predicts the Doppler frequency shift. However, embodiments of this disclosure are not limited thereto. For example, in some embodiments of this disclosure, the current channel correlation C n (T) and C from multiple previous TRS cycles n-Δ (T), ..., C n-2 (T), C n-1 The channel correlation of the window (T) is used as an input feature and is provided to a recurrent neural network (RNN) or a long short-term memory (LSTM) neural network.

[0123] Therefore, aspects of embodiments of this disclosure relate to systems and methods for calculating an estimated Doppler spread based on information from an input estimated channel, including channel correlations calculated based on a reference signal. According to some embodiments, the estimated Doppler spread is calculated based on one or more trained Doppler shift predictors, wherein the Doppler shift predictors are trained based on measurements collected from an actual physical radio receiver or from an actual link-level simulator. Some aspects of embodiments of this disclosure relate to selecting a Doppler shift predictor from a plurality of Doppler shift predictors based on a currently estimated SNR, wherein each of the plurality of Doppler shift predictors is trained on training data on different portions of training data, such as grouped by multiple portions of the SNR range of the training data. Some aspects of embodiments of this disclosure relate to combining the outputs of a plurality of Doppler shift predictors based on calculated input features and one or more probabilities corresponding to each Doppler shift predictor trained on data from different portions of the Doppler shift range in the training data.

[0124] The term "processing circuitry" is used herein to refer to any combination of hardware, firmware, and software for processing data or digital signals. Processing circuitry hardware may include, for example, a radio baseband processor (BP or BBP), an application-specific integrated circuit (ASIC), a general-purpose or special-purpose central processing unit (CPU), a digital signal processor (DSP), a graphics processing unit (GPU), or a programmable logic device such as a field-programmable gate array (FPGA). In processing circuitry, as used herein, each function is executed by hardware configured (i.e., hardwired) to perform said function, or by more general-purpose hardware (such as a CPU) configured to execute instructions stored in a non-transitory storage medium. Processing circuitry may be fabricated on a single printed circuit board (PCB) or distributed across several interconnected PCBs. Processing circuitry may include other processing circuitry; for example, processing circuitry may include two processing circuits, an FPGA and a CPU, interconnected on a PCB.

[0125] It should be understood that although the terms “first,” “second,” “third,” etc., may be used herein to describe various elements, components, regions, layers, and / or portions, these elements, components, regions, layers, and / or portions should not be limited by these terms. These terms are used only to distinguish one element, component, region, layer, or portion from another element, component, region, layer, or portion. Therefore, without departing from the spirit and scope of this disclosure, the first element, component, region, layer, or portion discussed herein may be referred to as the second element, component, region, layer, or portion.

[0126] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. As used herein, the terms “substantially,” “about,” and similar terms are used as approximate terms rather than terms of degree and are intended to take into account the inherent biases of measurements or calculations that will be recognized by one of ordinary skill in the art.

[0127] As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. It will be further understood that, when used in this specification, the terms “comprising” and / or “comprising…” indicate the presence of the stated 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, components, and / or combinations thereof. As used herein, the term “and / or” includes any and all combinations of one or more of the associated listed items. Expressions such as “at least one of…” modify the entire column of elements when preceding a column of elements, rather than individual elements of the column. Furthermore, the use of “may” in describing embodiments of this disclosure means “one or more embodiments of this disclosure.” Additionally, the term “exemplary” is intended to indicate or illustrate. As used herein, the terms “use,” “using…,” and “being used” may be considered synonymous with the terms “utilize,” “using…,” and “being exploited,” respectively.

[0128] It should be understood that when a component or layer is referred to as being "on," "connected to," "coupled to," or "adjacent to" another component or layer, it may be directly on, connected to, coupled to, or adjacent to the other component or layer, or one or more intermediate components or layers may exist. Conversely, when a component or layer is referred to as being "directly on," "directly connected to," "directly coupled to," or "immediately adjacent to" another component or layer, no intermediate components or layers exist.

[0129] Any numerical range described herein is intended to include all subranges of the same numerical precision falling within the range. For example, the range “1.0 to 10.0” is intended to include all subranges between (and including) the minimum value 1.0 and the maximum value 10.0, i.e., a minimum value equal to or greater than 1.0 and a maximum value equal to or less than 10.0, such as, for example, 2.4 to 7.6. Any maximum numerical limit described herein is intended to include all lower numerical limits falling within it, and any minimum numerical limit described in this specification is intended to include all higher numerical limits falling within it.

[0130] While this disclosure has been described in conjunction with certain exemplary embodiments, it should be understood that this disclosure is not limited to the disclosed embodiments, but rather is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims and their equivalents.

Claims

1. A method for estimating the Doppler spread of a wireless channel, comprising: The processing circuitry of a radio receiver extracts one or more features from the received signal, wherein the one or more features include a channel correlation estimated based on a reference signal in the current time slot, the estimated channel correlation indicating the rate of change of the Doppler frequency shift of the radio channel over time; and The processing circuit calculates the Doppler spread of the wireless channel by providing one or more features to one or more Doppler shift predictors trained on training data spanning the training signal-to-noise ratio (SNR) range and the training Doppler shift range. Each Doppler frequency shift predictor is trained on a portion of training data corresponding to a different part of the training SNR range or a different part of the training Doppler frequency shift range. Wherein, the one or more Doppler frequency shift predictors are trained using machine learning, and The training data includes data collected from operating radio communication systems that correlate channel correlation with Doppler spread.

2. The method according to claim 1, wherein, The estimated channel correlation includes a single channel correlation filtered by an infinite impulse response.

3. The method according to claim 1, wherein, The one or more features include one or more estimated channel correlations based on one or more reference signals in one or more previous time slots.

4. The method according to claim 1, wherein, The reference signal is a tracking reference signal.

5. The method according to claim 1, wherein, Each of the one or more Doppler shift predictors is trained based on different sub-ranges of the training SNR range, each sub-range having a lower bound and an upper bound, and The method further includes: Determine the current SNR of the received signal; and Based on the current SNR, a Doppler shift predictor is selected from the one or more Doppler shift predictors, wherein the lower limit of the corresponding subrange of the selected Doppler shift predictor is higher than the current SNR.

6. The method according to claim 5, wherein, The lower limit of the corresponding subrange of the selected Doppler frequency shift predictor is closest to the current SNR among the lower limits of the subranges above the current SNR.

7. The method according to claim 1, wherein, Each of the one or more Doppler frequency shift predictors is trained based on different sub-ranges of the training Doppler frequency shift range, and The method further includes: The processing circuit calculates one or more classification probabilities by providing the one or more features to a Doppler shift classifier network, wherein each of the one or more classification probabilities corresponds to a different Doppler shift predictor among the one or more Doppler shift predictors. Wherein, the one or more features are provided to the one or more Doppler frequency shift predictors to calculate one or more predicted Doppler frequency shifts, and The step of calculating the Doppler spread includes: combining the one or more predicted Doppler shifts according to the one or more classification probabilities.

8. The method according to claim 7, wherein, The step of combining the one or more predicted Doppler shifts includes summing the one or more predicted Doppler shifts multiplied by one or more products of the corresponding classification probabilities among the one or more classification probabilities.

9. The method according to claim 7, wherein, The step of combining the one or more predicted Doppler shifts includes: outputting the predicted Doppler shift with the highest probability among the one or more predicted Doppler shifts that corresponds to the highest classification probability among the one or more classification probabilities.

10. The method according to claim 1, wherein, The training SNR range of the training data is greater than the operating SNR range of the radio receiver.

11. The method according to claim 1, wherein, Each of the one or more Doppler shift predictors is trained to calculate the predicted Doppler shift based on a regression model.

12. The method according to claim 1, wherein, Each of the one or more Doppler shift predictors is trained to classify the one or more features by calculating one or more probabilities corresponding to each of the one or more ranges of the Doppler shift.

13. The method according to claim 1, wherein, Each of the one or more Doppler shift predictors is a multilayer perceptron.

14. A radio receiver including a channel estimator, wherein, The channel estimator includes: A feature extractor is configured to extract one or more features from a received signal, wherein the one or more features include a channel correlation estimated based on a reference signal in the current time slot, the estimated channel correlation indicating the rate of change of the Doppler frequency shift of the wireless channel over time; and A Doppler spread estimator is configured to estimate the Doppler spread of the wireless channel by feeding the one or more features into one or more Doppler shift predictors trained on training data spanning the training signal-to-noise ratio (SNR) range and the training Doppler shift range. Each Doppler frequency shift predictor is trained on a portion of training data corresponding to a different part of the training SNR range or a different part of the training Doppler frequency shift range. Wherein, the one or more Doppler frequency shift predictors are trained using machine learning, and The training data includes data collected from operating radio communication systems that correlate channel correlation with Doppler spread.

15. The radio receiver according to claim 14, wherein, The estimated channel correlation includes a single channel correlation filtered by an infinite impulse response.

16. The radio receiver according to claim 14, wherein, The one or more features include one or more estimated channel correlations based on one or more reference signals in one or more previous time slots.

17. The radio receiver according to claim 16, wherein, The reference signal is a tracking reference signal.

18. The radio receiver according to claim 14, wherein, Each of the one or more Doppler shift predictors is trained based on different sub-ranges of the training SNR range, each sub-range having a lower bound and an upper bound, and The channel estimator further includes: An SNR extractor is configured to extract the current SNR of the received signal; and A predictor selector is configured to select a Doppler shift predictor from one or more Doppler shift predictors based on the current SNR, wherein the lower limit of the corresponding subrange of the selected Doppler shift predictor is higher than the current SNR.

19. The radio receiver according to claim 18, wherein, The lower limit of the corresponding subrange of the selected Doppler frequency shift predictor is closest to the current SNR among the lower limits of the subranges above the current SNR.

20. The radio receiver according to claim 14, wherein, Each of the one or more Doppler frequency shift predictors is trained based on different sub-ranges of the training Doppler frequency shift range, and The Doppler expansion estimator includes a Doppler frequency shift classifier network, wherein the Doppler frequency shift classifier network is configured to calculate one or more classification probabilities that the one or more features belong to the categories corresponding to the one or more Doppler frequency shift predictors. The Doppler extension estimator is configured to provide the one or more features to the one or more Doppler frequency shift predictors to calculate one or more predicted Doppler frequency shifts, and The Doppler spread estimator is configured to compute the Doppler spread by combining one or more predicted Doppler shifts based on one or more classification probabilities.

21. The radio receiver according to claim 20, wherein, The Doppler extension estimator is configured to combine the one or more predicted Doppler shifts by summing the products of the one or more predicted Doppler shifts multiplied by the corresponding classification probabilities of the one or more classification probabilities.

22. The radio receiver according to claim 20, wherein, The Doppler expansion estimator is configured to combine the one or more predicted Doppler shifts by outputting the predicted Doppler shift with the highest probability among the one or more classification probabilities.

23. The radio receiver according to claim 14, wherein, The training SNR range of the training data is greater than the operating SNR range of the radio receiver.

24. The radio receiver according to claim 14, wherein, Each of the one or more Doppler shift predictors is trained to calculate the predicted Doppler shift based on a regression model.

25. The radio receiver according to claim 14, wherein, Each of the one or more Doppler shift predictors is trained to classify the one or more features by calculating one or more probabilities corresponding to each of the one or more ranges of the Doppler shift.

26. The radio receiver according to claim 14, wherein, Each of the one or more Doppler shift predictors is a multilayer perceptron.

Images