Method and apparatus for machine learning based symbol timing recovery
By using the time offset generated by machine learning in the communication system and combining the frequency domain channel correlation characteristics, the time offset problem caused by timing drift is solved, and the timing recovery performance and stability of the communication system are improved.
Patent Information
- Application Number
- CN202210379389.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2021-06-25
- Filing Date
- 2022-04-12
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-04-12
AI Technical Summary
In the process of signal processing, existing communication systems are difficult to quickly and accurately adjust the transformation window, especially between infrequently transmitted reference signals, resulting in errors and system performance degradation.
Using a machine learning-based method, an estimated time offset is generated through a neural network, using the frequency domain channel correlation characteristics extracted from the channel estimation, the transformation window in the communication system is adjusted, and the relatively frequent and infrequent reference signals are combined to improve the Accuracy and stability of timing recovery.
It realizes more stable and accurate time offset estimation, reduces the system complexity, energy consumption and processing time, improves the performance of symbol timing recovery, and is suitable for various communication systems.
Smart Images

Figure CN115208729B_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates generally to communication systems, and more particularly to systems, methods, and apparatus for machine learning based symbol timing recovery. Background Art
[0002] The communication system may use a transform such as a fast Fourier transform (FFT) to determine the frequency content of a signal. An FFT may be performed on samples of a signal in a timing window that may drift.
[0003] The above information disclosed in this Background section is only for enhancement of understanding of the background of the invention and therefore it may contain information that does not constitute the prior art. Summary of the invention
[0004] A method may include generating an estimated time offset based on a reference signal in a communication system, and adjusting a transform window in the communication system based on the estimated time offset, wherein the estimated time offset is generated based on machine learning. Generating the estimated time offset may include applying machine learning to one or more channel estimates. Generating the estimated time offset may include extracting one or more features from the one or more channel estimates, and generating the estimated time offset based on the one or more features. Extracting the one or more features may include determining a correlation between a first channel and a second channel. The correlation may include a frequency domain correlation between the first channel and the second channel. Extracting the one or more features may include extracting a subset of a set of features of the one or more channel estimates. The set of features may include a set of frequency domain channel correlations. The set of frequency domain channel correlations may be calculated based on a frequency domain channel separated by Δ subcarriers. The set of frequency domain channel correlations may be based on channel correlations corresponding to a range of Δ, and the subset may include frequency domain channel correlations based on one or more highest Δ and one or more lowest Δ. A neural network may generate an estimated time offset based on the one or more features. Generating the estimated time offset based on the one or more features may be performed at least in part by a neural network. The neural network may generate one or more classification outputs based on the one or more features. Generating the estimated time offset may include combining two or more classification outputs from the neural network. The reference signal may include a first reference signal, the estimated time offset may include a first estimated time offset, and the method may also include generating a second estimated time offset based on a second reference signal, which second reference signal may be generated less frequently than the first reference signal, and adjusting the transform window based on a combination of the first estimated time offset and the second estimated time offset.
[0005] A method may include estimating a time offset based on a first reference signal, estimating a first arrival path based on a second reference signal in a communication system, the second reference signal may be sent less frequently than the first reference signal, and adjusting a transform window based on the time offset and the first arrival path. Estimating the time offset based on the first reference signal may include generating one or more channel estimates based on the first reference signal, extracting one or more features from the one or more channel estimates, and generating the time offset based on the one or more features through a machine learning process. The one or more channel estimates may provide a set of frequency domain channel correlations between channels separated by Δ subcarriers, and extracting the one or more features may include extracting a subset of the set of frequency domain channel correlations. The number of candidate frequency domain channel correlations may be based on a reference signal configuration and a resource block bundling configuration. The machine learning process may be based on a neural network.
[0006] An apparatus may include: a feature extractor configured to extract one or more features from one or more channel estimates based on a reference signal; trained logic configured to generate an estimated time offset based on the one or more features; and a window processor configured to adjust a transform window based on the estimated time offset. The feature extractor may include a channel correlator, and the trained logic may include a neural network. BRIEF DESCRIPTION OF THE DRAWINGS
[0007] The drawings are not necessarily drawn to scale, and for the purpose of illustration, elements of similar structure or function are generally represented by similar reference numerals or portions thereof in all drawings. The drawings are intended only to facilitate the description of the various embodiments described herein. The drawings do not describe every aspect of the teachings disclosed herein and do not limit the scope of the claims. In order to prevent the drawings from becoming obscure, not all components, connections, etc. may be shown, and not all components may have reference numerals. However, the mode of component configuration can be easily seen from the drawings. The drawings, together with the specification, illustrate example embodiments of the present disclosure and, together with the description, are used to explain the principles of the present disclosure.
[0008] Figure 1A is a timing diagram illustrating a timing window of an FFT with a positive time offset in a communication system using orthogonal frequency-division multiplexing (OFDM) according to an example embodiment of the present disclosure.
[0009] Figure 1B is a timing diagram illustrating a timing window of an FFT with a negative time offset in a communication system using OFDM according to an example embodiment of the present disclosure.
[0010] Figure 2An embodiment of a system for estimating time offset according to the present disclosure is shown.
[0011] Figure 3 An example embodiment of a system for estimating time offset using channel correlation according to the present disclosure is shown.
[0012] FIG. 4A to FIG. 4E An example of a subset of channel correlations that may be used as input to a neural network according to an example embodiment of the present disclosure is shown.
[0013] Figure 5 An example embodiment of a multilayer perceptron according to the present disclosure is shown.
[0014] Figure 6 Another example embodiment of a multilayer perceptron according to the present disclosure is shown.
[0015] Figure 7 An example embodiment of a system for adjusting the timing of a transform window according to the present disclosure is shown.
[0016] Figure 8 is a timing diagram illustrating the operation of an embodiment of a system for combining timing estimates from multiple sources according to the present disclosure.
[0017] Fig. 9 An example embodiment of a user device according to the present disclosure is shown.
[0018] Fig.10 An example embodiment of a base station according to the present disclosure is shown.
[0019] Fig.11 An embodiment of a method for adjusting the timing of a transform window according to the present disclosure is shown. DETAILED DESCRIPTION
[0020] Overview
[0021] In some communication systems according to example embodiments of the present disclosure, a transform such as a Fast Fourier Transform (FFT) may be performed on a signal to determine the frequency content of symbols that may be transmitted in the signal. The transform may be performed in a timing window that may be aligned with the symbol. Timing drift may cause the window to shift relative to the symbol, resulting in errors in the transform. Therefore, one or more symbol timing recovery (STR) techniques may be used to adjust the timing of the transform window to keep it synchronized with the symbol.
[0022] In some embodiments, the STR technique can be based on a periodic broadband reference signal (RS), such as a synchronization signal block (SSB) in a 5G wireless network. The SSB can be used to estimate the first arrival path (FAP), which can then be used to adjust the window timing. However, such reference signals can be sent relatively infrequently. Therefore, it may be difficult to use one of these reference signals to adjust a timing window that may drift rapidly during the time between infrequently sent reference signals.
[0023] In some embodiments, STR techniques can be based on more frequently transmitted reference signals, such as the demodulation reference signal (DMRS) for the physical downlink shared channel (PDSCH) in 5G wireless networks. For example, some embodiments can adjust the timing window based on the difference between two center of mass (COM) estimates of the frequency domain correlation between channels. However, some COM techniques may have accuracy, stability and / or residual offset issues and require storage of previous COM estimates to calculate the difference.
[0024] In some STR schemes according to example embodiments of the present disclosure, machine learning techniques may be used to estimate time offsets to adjust transform windows. For example, in some embodiments, a neural network may be trained to generate an estimated time offset based on one or more features extracted from a channel estimate. The channel estimate may include timing information based on relatively frequent reference signals such as a PDSCH DMRS. Although not limited to any particular feature, in some embodiments, one or more frequency domain (FD) channel correlations may be used as input features to the neural network.
[0025] Depending on implementation details, STR schemes according to example embodiments of the present disclosure may provide a direct estimate of the time offset, and therefore, no prior time estimate may be used. Additionally or alternatively, some embodiments include a residual time offset within the estimated time offset. Additionally or alternatively, because some embodiments may estimate the time offset based on multiple FD channel correlations, they may provide an estimate that may be more stable, more accurate, and / or more robust.
[0026] In some embodiments, time estimates generated using machine learning based on relatively frequent reference signals can be combined with time estimates based on relatively infrequent reference signals. Depending on implementation details, this can leverage existing devices and / or processing, thereby reducing cost, complexity, energy consumption, processing time, etc.
[0027] In some embodiments, features used as input to a neural network may be selected as a subset of a larger number of candidate features. For example, in embodiments using FD channel correlations as input features, there may be a total number of candidate FD channel correlations based on, for example, reference signal configuration and / or resource block bundling size. By using only a subset of candidate FD channel correlations, some embodiments may reduce the cost, complexity, processing time, storage space, integrated circuit (IC) area, power consumption, etc. of a neural network and / or associated apparatus.
[0028] Depending on implementation details, some embodiments may operate on any type of channel, regardless of system bandwidth, channel selectivity, etc.
[0029] The present disclosure includes many inventive principles related to symbol timing recovery. These principles may have independent utility and may be implemented separately, and not every embodiment may utilize every principle. In addition, these principles may also be embodied in various combinations, some of which may enhance the benefits of the individual principles in a synergistic manner.
[0030] Example Embodiments
[0031] Some example embodiments of systems, devices, apparatuses, processes, methods, etc. are described herein, showing some possible implementation details according to the present disclosure. These examples are provided to illustrate the principles of the present disclosure, but these principles are not limited to or defined by these embodiments, implementation details, etc. For example, some embodiments may be described in the context of 5G and / or New Radio (NR) wireless communication systems, but these principles may also be applied to any other type of wired and / or wireless system, including 3G, 4G and / or future generation wireless networks, and / or any other communication system that can implement symbol timing recovery, frequency conversion window adjustment, etc.
[0032] Timing offset
[0033] Figure 1A and Figure 1BThe timing windows of FFTs with positive time offset (TO) and negative TO, respectively, are shown in accordance with an example embodiment of the present disclosure in a communication system using orthogonal frequency division multiplexing (OFDM). Each OFDM symbol may include a data portion and a cyclic prefix (CP) portion. In some embodiments, the CP may be implemented as a copy of the final portion of the data.
[0034] Item 1 idx can represent the starting point of the correct FFT timing window of the OFDM symbol, and the term l TO It can represent the absolute value of TO which can be measured, for example, in terms of FFT samples. Figure 1A In the positive TO case shown, the starting point of the FFT window can be determined by l idx -l TO Given. Figure 1B In the case of negative TO shown in the figure, the starting point of the FFT window can be determined by l idx +l TO Given.
[0035] In the case of negative TO, inter-symbol interference (ISI) may occur because the FFT window may occupy a part of the next OFDM symbol (in Figure 1B In the example shown, the CP part of the next OFDM symbol).
[0036] In the case of positive TO, the presence of CP can prevent ISI depending on the magnitude of TO. CP Can represent the CP length, and l DS It can represent the delay spread (DS) of the channel. idx -l TO >l idx -l CP +l DS (ie, l CP >l TO +l DS ), there may be essentially no ISI.
[0037] In some embodiments, the sampling position can be set to be several samples ahead of the estimated sampling position to prevent negative TO conditions. However, if TO is too large, or the delay spread of the channel is too large, so that l CP <l TO +l DS , the FFT window may occupy part of the previous OFDM symbol and inter-symbol interference may occur. Regardless of the source, ISI may degrade the performance of the system.
[0038] Effects of Timing Offset
[0039] To illustrate the effect of time offset, the received signal at time sample n can be expressed as y n , n = 0, ..., N-1. After FFT, the received signal at subcarrier k can be expressed as Y k , k = 0, ..., N-1. The transmitted reference signal symbol and frequency domain channel at subcarrier k can be represented as X k and H k Therefore, the received signal can be given by
[0040] Y k =H k X k +W k (1)
[0041] Where W k can represent the noise at subcarrier k. The time offset can be expressed as l TO Sample, l cp can represent the cyclic prefix (CP) length, and l DS It can represent the channel delay spread length. If there is essentially no ISI from the previous or next OFDM symbol (i.e., 0 < l TO <l cp -l DS ), then there is a time offset l TO The received signal at the kth subcarrier can be given by
[0042]
[0043] And therefore, the channel estimate at the kth subcarrier can be given by
[0044]
[0045] It can be a phase-rotated version of the original channel plus some noise. TO The information can be embedded in the channel estimation middle.
[0046] Symbol Timing Recovery Based on Power Delay Profile
[0047] Because channel estimates can contain information about time offsets, some STR techniques according to example embodiments of the present disclosure can extract timing information from one or more channel estimates (which can include reference signal information) to adjust the timing of the FFT window. For example, in some embodiments, the STR technique can be based on a periodic wideband reference signal, such as a synchronization signal block (SSB) or a tracking reference signal (TRS) in a 5G wireless network. A power delay profile (PDP) can be calculated from one or more channel estimates based on the wideband reference signal. The PDP information can be used to estimate the first arrival path (FAP), which can then be used to adjust the timing of the FFT window.
[0048] In some embodiments, a wideband reference signal such as SSB or TRS may be sent relatively infrequently. Therefore, if the time offset of the FFT window drifts too much between reference signal transmissions, e.g., due to a large sampling clock error, the PDP-based STR algorithm may not be able to adjust the timing of the FFT window with acceptable accuracy.
[0049] Centroid-Based Symbol Timing Recovery
[0050] Some STR techniques according to example embodiments of the present disclosure may track timing drift based on more frequently transmitted reference signals, such as PDSCH DMRS signals in 5G wireless systems. For example, some embodiments may estimate the center of mass (COM) based on frequency domain correlations between channels. However, in some embodiments, a single COM estimate may not be used to directly adjust the timing window. Instead, the previous COM estimate may be saved, and the difference between two consecutive COM estimates may be used to determine the time offset. Furthermore, in some embodiments, the COM estimation technique may not be able to adapt to residual time offsets. Additionally, because the COM estimate may be based on a single FD channel correlation, the COM estimate may be unstable. Furthermore, because the COM-based time offset estimate may only indicate changes in the time offset, it may require additional STR techniques (such as PDP-based STR techniques) to provide complete reference timing.
[0051] Learning-based symbol timing recovery
[0052] Some STR techniques according to example embodiments of the present disclosure may use machine learning to estimate time offsets for adjusting transform windows. Machine learning-based logic may act as a mapping function that maps one or more input features with reference signal timing information to one or more time offsets. For example, the logic may be trained based on data collected over a wide range of channel types, signal-to-noise ratios (SNRs), etc.
[0053] In some embodiments, learning-based STR may be characterized as a data-driven technique, while the above-described PDP and / or COM-based STR techniques may be characterized as model-driven.
[0054] Figure 2 An embodiment of a system for estimating time offset according to the present disclosure is shown. Figure 2 The illustrated embodiment may include a machine learning based feature extractor 202 and logic 204. In some embodiments, the machine learning based logic 204 may be referred to as trained logic. The feature extractor 202 may receive one or more channel estimates that may be estimated based on a reference signal. (eg, frequency domain channel estimation.) Examples of reference signals may include SSB, TRS, PDSCH DMRS, etc. In some embodiments, using reference signals that may be sent relatively frequently, such as PDSCH DMRS, may enable the system to track fast timing changes in the transition window.
[0055] The feature extractor 202 can estimate the One or more features g are extracted from the feature extractor 202. Examples of features may include frequency domain (FD) channel correlation, delay spread estimate (DSE), etc. In some embodiments, depending on the subcarrier distance, the feature extractor 202 may calculate multiple FD channel correlations as candidate features.
[0056] The logic 204 may implement any type of machine learning to estimate one or more time offsets l based on one or more input features. TO For example, in some embodiments, supervised learning may be used to enable the logic 204 to learn a mapping function from one or more frequency-domain channel correlations (input features) to one or more time offsets (output labels). In some embodiments, other types of machine learning such as unsupervised learning and / or reinforcement learning (e.g., learning from experience) may be used for the trained logic 204.
[0057] In some embodiments, the logic 204 can be trained offline (e.g., before operation), online (e.g., during real-time operation), or in a hybrid combination with initial offline training, where the initial offline training can be refined or enhanced based on the online training. In some embodiments, the logic 204 can be implemented with one or more neural networks of any type, such as a convolutional neural network (CNN), a deep neural network (DNN), a perceptron, a multi-layer perceptron (MLP) (e.g., a feed-forward MLP), etc.
[0058] Frequency Domain Channel Correlation
[0059] In some embodiments, frequency domain (FD) channel correlation can be used as an input feature for learning-based STR, since time offset may result in linear phase rotation at different subcarriers. Therefore, the FD channel correlation between channel estimates from the reference signal can be used to determine the time offset.
[0060] Figure 3 An example embodiment of a system for estimating time offset using channel correlation according to the present disclosure is shown. Figure 3 The system shown in the embodiment may illustrate some example implementation details that may be used, for example, to implement Figure 2 The system shown in .
[0061] Figure 3 The system shown may include a channel estimation block 300, a channel correlator 302, and a neural network, which in this example may be implemented as an MLP 304. The channel estimation block 300 may provide any number of channel estimates. Where k may indicate, for example, the number of subcarriers, such as k=0, ..., N-1. The channel correlator 302, which may operate as a feature extractor, may be based on the channel estimation Generate any number of channel correlations Channel Correlation may be applied as input features to the MLP 304, where the MLP 304 may generate an estimated time offset l based on, for example, a mapping learned during training. TO .
[0062] In some embodiments, one or more different precoding matrices may be applied to PDSCH DMRS in different bundles. Therefore, in some embodiments, the frequency domain channel correlation may be calculated based on each bundle.
[0063] In some embodiments, It may refer to the frequency domain (FD) channel correlation between channels separated by Δ subcarriers, which may be estimated, for example, based on the PDSCH DMRS channels in the mth bundle. It can be calculated as
[0064]
[0065] in may refer to the frequency domain channel estimate at the kth subcarrier in the mth bundle, and K Δ may refer to a set of subcarrier indices that may satisfy two conditions: (1) the subcarrier index corresponds to a PDSCH DMRS position, and (2) both the subcarrier index k and the subcarrier index k+Δ belong to the same bundle.
[0066] In an example embodiment, resources may be allocated for DMRS configuration type 1, the DMRS may start from subcarrier 0, and the bundle size may be configured as 2. In this example, the set of possible subcarrier indices corresponding to PDSCH DMRS positions in the bundle may be {0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22}. In some embodiments, it may be convenient to discard the final entry. Therefore, for purposes of illustration, in this example embodiment, the set of possible subcarrier indices corresponding to PDSCH DMRS positions in the bundle may be {0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20}. Continuing with this example, for K Δ , Δ = 2, because subcarrier index k and subcarrier index k + Δ both belong to the same bundle, K 2 ={0, 2, 4, 6, 8, 10, 12, 14, 16, 18}. Similarly, for K Δ ,Δ=20,K 20 ={0}. However, in other embodiments, different values may be used for the DMRS configuration type, the value of Δ, and / or any other parameters.
[0067] Calculate for the mth bundle After that, the final frequency domain channel correlation It can be given by the following formula
[0068]
[0069] Where M may represent the total number of configured bundles. Then, one or more final frequency domain channel correlations can be used as input features to a neural network, where the neural network can be implemented as an MLP 304 in this example.
[0070] Feature Selection - Subset of Channel Correlations
[0071] In some embodiments, the candidate channel correlation The total number of may depend on the DMRS type and / or the bundling size. In an example where the DMRS is configured as type 1 and the bundling size is configured as 2, the candidate set of Δ may be {2, 4, 6, 8, 10, 12, 14, 16, 18, 20}, so that the total number of possible Δ values may be 10. Because Can be plural, so Average real part and the imaginary part can be used as the input of the neural network respectively. As a result, there are a total of 20 candidate features that make Δ=2, 4, 6, 8, 10, 12, 14, 16, 18, 20.
[0072] However, in some embodiments, some of the candidate features may be omitted while still providing acceptable performance. By using only a subset of the set of possible (e.g., available) input features, the number of inputs to the neural network may be reduced. Depending on the implementation details, this may reduce the size, cost, complexity, computational load, power consumption, storage space, etc. of the neural network.
[0073] FIG. 4A to FIG. 4E Channel correlations that can be used as input to a neural network according to an example embodiment of the present disclosure are shown. Some examples of subsets of . FIG. 4A to FIG. 4E In the illustrated embodiment, the shaded blocks indicate the channel correlations for which and will The values of Δ as input to the neural network, while the non-shaded blocks indicate for which channel correlation can be omitted The value of Δ.
[0074] Figure 4A shows that all available channel correlations can be calculated And use it as an example of the input of the neural network. Figure 4B shows the correlation of only the outermost channels Some of (e.g., its two end portions Δ=2, 4, 6, 16, 18, 20) can be calculated and applied as input to an embodiment of a neural network. Figure 4C shows that only the left channel correlation Some of (e.g., the left portion Δ=2, 4, 6, 8, 10, 12) may be computed and applied as input to an embodiment of a neural network. Figure 4D shows that only the right channel correlation Some of (e.g., the right side portion Δ=10, 12, 14, 16, 18, 20) may be computed and applied as input to an embodiment of a neural network. Figure 4E shows that only the innermost channel correlation Some of (e.g., the middle portion Δ=6, 8, 10, 12, 14, 16) may be computed and applied as input to an embodiment of a neural network.
[0075] In some embodiments, utilizing features of only the two end portions of a set of FD channel correlations can provide performance close to that provided by a set of all correlations while reducing the number of input features to the neural network. For example, the number of input features can be reduced from 20 features (e.g., Δ=2,4,6,8,10,12,14,16,18,20) is reduced to 12 features (e.g. = 2, 4, 6, 16, 18, 20), while still providing adequate performance. Thus, some of the embodiments discussed below may be described in the context of an implementation where only the outermost FD channel correlations (e.g., the end portion) may be used, but in other embodiments, the full set of available channel correlations or any subset thereof may be used. Likewise, any other type of feature may be used, such as DSE.
[0076] Feature Selection - Input Features for Different DMRS Types
[0077] In some example embodiments for DMRS Type 1 configuration, the following 12 features may be applied as input to the neural network regardless of whether the bundling size is configured as 2 or 4: Δ=2, 4, 6, 16, 18, 20.
[0078] In some example embodiments for DMRS type 2 configuration, the DMRS may start from subcarrier 0 and the bundle size may be configured as 2. In this configuration, the set of all possible subcarrier indices that may correspond to PDSCH DMRS positions in the bundle may be {0, 1, 6, 7, 12, 13, 18, 19}. In some embodiments, it may be convenient to drop some of the subcarrier indices and use {0, 6, 12, 18} as the set of all possible subcarrier indices. Thus, the candidate delta for DMRS type 2 may be {6, 12, 18}. The real and imaginary parts of can be processed separately. However, in this example, since there may be only three candidate Δ in total, it may be disadvantageous or infeasible to use only a subset of the possible input feature set, and all candidate Δ may be included as input features for DMRS Type 2. Therefore, in some example embodiments, the input features of the neural network with DMRS Type 2 may be Δ = 6, 12, 18. This set of six features can be applied as the input of the neural network regardless of whether the bundle size is configured as 2 or 4.
[0079] Table 1 shows some example input features of neural networks of different DMRS types based on the above example embodiments. These details are provided only to illustrate the principles of the present disclosure, and any other values may be used.
[0080] Table 1
[0081]
[0082] Multilayer Perceptron
[0083] Figure 5 An example embodiment of an MLP according to the present disclosure is shown. Figure 5 The MLP shown in can be used, for example, to implement Figure 3 The MLP 304 shown in FIG. 1 is configured as a DMRS of type 1 as shown in Table 1. The MLP may include an input layer including input nodes 5A-1, 5A-2, ..., 5A-12, each of which may receive an input FD channel correlation The corresponding one in .
[0084] The MLP may also include an output layer including output nodes 5C-1, 5C-2, ..., 5C-5, each of which may provide an output label corresponding to a probability that the time offset is a certain value. Figure 5 In the example shown, each output label can correspond to a time offset l TO The probability P of x is such that P(l TO =x), where x∈{-4,-2,0,2,4}.
[0085] The MLP may also include a hidden layer including hidden nodes 5B-1, 5B-2, ..., 5B-N, each of which may receive one or more inputs from one or more of the input nodes 5A-1, 5A2, ..., 5A-12, and provide outputs to one or more of the output nodes 5C-1, 5C-2, ..., 5C-5. In various embodiments, the hidden layer may include any number of nodes, such as 64 nodes.
[0086] Figure 6 Another example embodiment of an MLP according to the present disclosure is shown. For example, Figure 6 The illustrated embodiment can be used to implement Figure 3 The MLP 304 shown is configured as a DMRS of type 2 as shown in Table 1. The MLP may include an input layer, which includes input nodes 6A-1, 6A-2, ..., 6A-6, each of which may receive an input FD channel correlation The corresponding one in .
[0087] The MLP may also include an output layer including output nodes 6C-1, 6C-2, ..., 6C-5, each of which may provide an output label corresponding to a probability that the time offset is a certain value. Figure 6 In the example shown, each output label can correspond to a time offset l TO The probability P of x is such that P(l TO =x), where x∈{-4,-2,0,2,4}.
[0088] The MLP may also include a hidden layer including hidden nodes 6B-1, 6B-2, ..., 6B-N, each of which may receive one or more inputs from one or more of the input nodes 6A-1, 6A2, ..., 6A-6 and provide outputs to one or more of the output nodes 6C-1, 6C-2, ..., 6C-5.
[0089] To illustrate the principles of the present disclosure, some embodiments have been described with example implementation details, but these principles may be implemented with any other implementation details. For example, as described above, any type of neural network or other logic based on machine learning (such as CNN, DNN, etc.) may be used. In embodiments implemented with MLP or other neural networks, any number of layers including hidden layers may be used, and any number of nodes may be used in any layer. The output of the machine learning-based logic may be implemented as a continuous output (e.g., for regression), a discrete output, or a hybrid combination thereof. In some embodiments where timing may be implemented in units of samples, an implementation with discrete outputs may be beneficial because one or more values may already be in discrete form.
[0090] As another example, any set and / or subset of input features may be applied to the neural network. For example, some embodiments may have been described as using 12 FD channel correlations corresponding to Δ=2, 4, 6, 16, 18, 20 (for DMRS type 1), however, other subsets such as Δ=2, 4, 6, 8, 14, 16, 18, 20 or Δ=2, 4, 18, 20 may also be used, as well as other subsets of all, middle, end, right, left, etc.
[0091] Time offset estimation
[0092] In some embodiments, the learning-based STR algorithm according to the present disclosure can be formulated as a classification problem. For example, in Figure 5 and Figure 6 In the illustrated embodiment, a neural network (e.g., MLP) may have an output layer with 5 nodes corresponding to time offsets of -4, -2, 0, 2, or 4, thereby providing a total of 5 categories (in some embodiments, these time offsets may be implemented in terms of sampling clock cycles). In an embodiment implemented as a classification problem, a cross entropy cost function may be used, and a soft-max function may be applied after the nodes in the output layer. As a result, the output of the MLP may correspond to the probability that the input feature belongs to each category, such as the probability that the time offset is equal to -4, -2, 0, 2, or 4.
[0093] In some embodiments, an average combination technique may be used to obtain a final estimated time offset, in which the final time offset may be estimated by taking a weighted average of one or more of the output time offsets as described below, where the weights may be equal to the output of the MLP network
[0094]
[0095] Where P can represent the time offset l TO is the probability of some value x.
[0096] In some embodiments, the final estimated time offset can be obtained using a maximum combination technique as shown below, in which the final time offset can be estimated as the time offset label corresponding to the maximum output of the MLP network
[0097]
[0098] Although some example embodiments may have been shown with five output labels corresponding to time offsets of samples -4, -2, 0, 2, 4, any number of labels with any value of time offset covering a smaller or larger range may be used. For example, some embodiments may implement three output labels corresponding to time offsets of -2, 0, 2, five output labels corresponding to time offsets of -8, -4, 0, 4, 8, etc.
[0099] Estimation of the combination
[0100] In some embodiments, the estimated time offset generated by the machine learning based logic can be used to directly adjust the timing window of the FFT or other transform. However, in some embodiments, it may be beneficial to combine timing estimates from multiple sources. For example, depending on the range of time offsets that can be used for a particular implementation, the neural network can use a relatively large number of labels to cover that range with sufficient resolution (e.g., 17 output labels corresponding to time offsets of -16, -14, -12, -10, -8, -6, -4, -2, 0, 2, 4, 6, 8, 10, 12, 14, 16). In turn, this may involve using a relatively large amount of resources for the neural network, such as memory, processing resources, training resources, etc.
[0101] In some embodiments, the estimated time offset generated by the trained logic can be combined with another timing estimate, such as an estimated FAP based on a PDP from one or more channel estimates. Depending on the implementation details, this type of combination can enable the STR system to cover a wide range of time offsets while reducing the amount of resources used by the trained logic.
[0102] Figure 7 An example embodiment of a system for adjusting the timing of an FFT or other transform window according to the present disclosure is shown. Figure 7 The illustrated embodiment may include a first processing path having a feature extractor 702 and trained logic 704 , a second processing path including a PDP processor 706 and a STR block 708 , and a window processor 710 .
[0103] In some embodiments, the feature extractor 702 and the trained logic 704 may be implemented in a manner similar to Figure 2 The feature extractor may receive one or more FD channel estimates based on a reference signal (such as PDSCH DMRS) that may be transmitted relatively frequently. The feature extractor 702 may generate one or more input features g, which may be applied to the trained logic 704, which may generate an estimated time offset l. TO .
[0104] The PDP processor 706 may estimate the FD channel from one or more FD channel estimates based on a reference signal that may be sent relatively infrequently (e.g., a periodic wideband reference signal such as SSB or TRS). Generate PDP Information The STR block 708 may use the PDP information to estimate the FAP.
[0105] The window processor 710 may combine the estimated FAP and the estimated time offset l TO , to adjust the timing window of the FFT or other transforms.
[0106] exist Figure 7 In the illustrated embodiment, the trained logic 704 may be implemented using a relatively small neural network (eg, an MLP) that may be trained to provide a relatively small number of output labels (eg, labels that may represent time offsets l TO =-4, -2, 0, 2, 4), so the maximum time offset that can be compensated by the learned STR can be + / -4. However, these relatively small number of output labels can be combined with the underlying PDP-based STR algorithm in a synergistic manner to provide complete reference timing over a wider range of time offsets. In addition, because the feature extractor 702 and the trained logic 704 can operate based on reference signals that can be sent relatively frequently (such as PDSCH DMRS), they can track relatively rapid timing changes that may occur between transmissions of wideband reference signals such as SSB or TRS.
[0107] Figure 8 is a timing diagram illustrating the operation of an embodiment of a system that can combine timing estimates from multiple sources according to the present disclosure. Figure 8 The timing diagram shown in can show, for example, Figure 7 Some aspects of the operation of the system shown in .
[0108] Figure 8 708. For example, the PDP-based STR implemented by the PDP processor 706 and the STR block 708 may be implemented in the time period T 2 , and the learning-based STR implemented by the feature extractor 702 and the trained logic 704 can be used in the time period T 1 On operation.
[0109] The time period T of STR based on PDP 2 It may depend on the configuration of the period of the base reference signal. For example, for an SSB reference signal with a period of 10 ms, the PDP-based STR may have a period of T 2=10 ms, which may correspond to 80 time slots for a subcarrier spacing set to 120 KHz. However, if the time offset drifts during the period between transmissions of the wideband reference signal, e.g., due to a large sampling clock error, the PDP-based STR algorithm may not be able to fully track the fast time offset drift, since the next reference signal has not yet been sent.
[0110] To track fast timing drifts, learning-based STR can exploit the information from possibly more frequently transmitted (e.g., with a shorter period T 1 ) of the PDSCH DMRS reference signal. In some embodiments, the PDSCH DMRS may be sent in each time slot in which PDSCH data is sent. In some embodiments, the period T of the learned STR may be used to estimate the channel of the PDSCH DMRS reference signal. 1 can be implemented as a configurable parameter that can be set based on the expected rate of timing drift that the system can be configured to track. For example, if the fastest timing drift that the system can be configured to track is 1 sample / ms, then based on the period T of the learned STR 1 Can be set to T 1 =1 millisecond, where if the SCS is set to 120KHz, this can correspond to 8 time slots.
[0111] Iterative Pruning
[0112] In some embodiments, iterative pruning can be used to reduce the number of parameters in the trained logic, such as the number of weights in a neural network. Depending on the implementation details, this can reduce the cost, complexity, processing time, storage space, IC area, power consumption, etc. of the neural network and / or related devices.
[0113] In an example embodiment of an iterative pruning process according to the present disclosure, at each iteration, training can be performed using masking values applied to the weights and / or gradients of the neural network, and if some elements of the weights are less than a predetermined threshold and the product of the standard deviation of some or all weights, these elements can be forced to zero. The set of masking values can be increased after one or more iterations, and the output weights in the last iteration can be used as the final output weights of the trained neural network.
[0114] User Equipment
[0115] Fig. 9 An example embodiment of a user equipment (UE) according to the present disclosure is shown. Fig. 9The illustrated embodiment 900 may include a radio transceiver 902 and a controller 904, which may control the operation of the transceiver 902 and / or any other components in the UE 900. For example, the UE 900 may be used to implement any of the functions described in the present disclosure, including learning-based STR, PDP-based STR, feature extraction, FD channel correlation, neural network training and / or inference, estimation of time offset, combinations of time estimates and / or offsets, etc.
[0116] The transceiver 902 may transmit / receive one or more signals to / from a base station and may include an interface unit for such transmission / reception. For example, the transceiver 902 may receive a PDSCH signal and / or a reference signal such as SSB, TRS, PDSCH DMRS, etc. from a base station.
[0117] The controller 904 may include, for example, one or more processors 906 and a memory 908, which may store instructions for the one or more processors 906 to execute code to implement any of the functions described in the present disclosure. For example, the controller 904 may be used to implement learning-based STR, PDP-based STR, feature extraction, FD channel correlation, neural network training and / or inference, estimation of time offset, a combination of time estimation and / or offset, etc.
[0118] Base Station
[0119] Fig.10 An example embodiment of a base station according to the present disclosure is shown. Fig.10 The illustrated embodiment 1000 may include a radio transceiver 1002 and a controller 1004, which may control the operation of the transceiver 1002 and / or any other components in the base station 1000. For example, the base station 1000 may be used to implement any of the functions described in the present disclosure, including learning-based STR, PDP-based STR, feature extraction, FD channel correlation, neural network training and / or inference, estimation of time offset, combinations of time estimates and / or offsets, etc.
[0120] The transceiver 1002 may transmit / receive one or more signals to / from the user equipment and may include an interface unit for such transmission / reception. For example, the transceiver 1002 may transmit a PDSCH signal and / or a reference signal such as SSB, TRS, PDSCH DMRS, etc. to the user equipment.
[0121] The controller 1004 may include, for example, one or more processors 1006 and a memory 1008, which may store instructions for the one or more processors 1006 to execute code to implement any base station functions described in the present disclosure. For example, the controller 1004 may be used to implement learning-based STR, PDP-based STR, feature extraction, FD channel correlation, neural network training and / or inference, estimation of time offset, a combination of time estimation and / or offset, etc.
[0122] exist Fig. 9 and 10 In the illustrated embodiment, transceivers 902 and 1002 may be implemented with various components (such as amplifiers, filters, modulators and / or demodulators, A / D (analog-to-digital) and / or D / A (digital-to-analog) converters, antennas, switches, phase shifters, detectors, couplers, conductors, transmission lines, etc.) to receive and / or transmit RF signals. Controllers 904 and 1004 may be implemented with hardware, software, and / or any combination thereof. For example, all or part of a hardware implementation may include combinatorial logic, sequential logic, timers, counters, registers, gate arrays, amplifiers, synthesizers, multiplexers, modulators, demodulators, filters, vector processors, complex programmable logic devices (CPLDs), field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), systems on chips (SOCs), state machines, data converters such as ADCs and DACs, etc. All or part of a software implementation may include one or more processor cores, memories, programs, and / or data storage, etc., which may be located locally and / or remotely and may be programmed to execute instructions to perform one or more functions of the controller. Some embodiments may include one or more CPUs (such as complex instruction set computer (CISC) processors (such as x86 processors) and / or reduced instruction set computer (RISC) processors (such as ARM processors)), graphics processing units (GPUs), neural processing units (NPUs), tensor processing units (TPUs), etc.) that execute instructions stored in any type of memory.
[0123] Additional Embodiments
[0124] Fig.11An embodiment of a method for adjusting the timing of a transform window according to the present disclosure is shown. The method may start at operation 1102. At operation 1104, the method may generate an estimated time offset based on a reference signal in a communication system. At operation 1106, the method may adjust the transform window based on the estimated time offset in the communication system, wherein the estimated time offset may be generated based on machine learning. The method may end at operation 1108.
[0125] exist Fig.11 In the illustrated embodiment, the components and / or operations shown are exemplary only. Some embodiments may relate to various additional components and / or operations not shown, and some embodiments may omit some components and / or operations. In addition, in some embodiments, the arrangement of components and / or the temporal order of operation may vary. Although some components may be shown as separate components, in some embodiments, some components shown separately may be integrated into a single component, and / or some components shown as a single component may be implemented with multiple components.
[0126] The embodiments disclosed herein may be described in the context of various implementation details, but the principles of the present disclosure are not limited to these or any other specific details. Some functions have been described as being implemented by certain components, but in other embodiments, the functions may be distributed between different systems and components in different locations. Reference to a component or element may refer only to a portion of a component or element. Terms such as "first" and "second" used in the present disclosure and claims may be used solely to distinguish the things they modify, and may not indicate any spatial or temporal order unless it becomes obvious from the context. Mentioning the first thing does not mean the existence of the second thing. In addition, according to the inventive principles of this patent disclosure, the various details and embodiments described above may be combined to generate additional embodiments. For convenience, various organizational aids, such as section titles, etc., may be provided, but the subject matter arranged according to these aids and the principles of the present disclosure is not defined or limited by these organizational aids.
[0127] Since the inventive concepts of this patent disclosure may be modified in arrangement and detail without departing from the inventive concept, such changes and modifications are considered to fall within the scope of the appended claims.
Claims
1. A method for symbol timing recovery, include: generating, in a communication system, a first estimated time offset based on a first reference signal that periodically recurs with a first period, wherein the first estimated time offset is generated based on machine learning; generating, in the communication system, a second estimated time offset based on a second reference signal that periodically recurs with a second period that is longer than the first period; and A transform window is adjusted in the communication system based on a combination of a first estimated time offset and a second estimated time offset.
2. The method according to claim 1, in, Generating the first estimated time offset includes applying the machine learning to one or more channel estimates derived from the first reference signal.
3. The method according to claim 1, in, Generating a first estimated time offset includes: extracting one or more features from one or more channel estimates derived from the first reference signal; and A first estimated time offset is generated based on the one or more features.
4. The method according to claim 3, in, Extracting the one or more features includes determining a correlation between the first channel and the second channel.
5. The method according to claim 4, in, The correlation includes a frequency domain correlation between the first channel and the second channel.
6. The method according to claim 4, in, Extracting the one or more features includes extracting a subset of a set of features of the one or more channel estimates.
7. The method according to claim 6, in, The set of features includes a set of frequency domain channel correlations.
8. The method according to claim 7, in: The set of frequency-domain channel correlations is calculated on a per-bundle basis based on frequency-domain channels separated by bundle-specific subcarriers, wherein each bundle includes a predetermined number of bundle-specific subcarriers; The set of frequency-domain channel correlations is based on a series of bundled channel correlations corresponding to different numbers of bundle-specific subcarriers; and The subset includes frequency domain channel correlations based on one or more highest numbers of the bundle-specific subcarriers and one or more lowest numbers of the bundle-specific subcarriers.
9. The method according to claim 3, in, The neural network generates a first estimated time offset based on the one or more features.
10. The method according to claim 3, in, Generating a first estimated time offset based on the one or more features is performed at least in part by a neural network.
11. The method according to claim 10, in, The neural network generates one or more classification outputs based on the one or more features.
12. The method according to claim 11, in, Generating a first estimated time offset includes combining two or more classification outputs from the neural network.
13. A method for symbol timing recovery, include: estimating a time offset in a communication system based on a first reference signal that periodically recurs with a first period, wherein the time offset is estimated using a machine learning process; estimating a first arrival path based on a second reference signal periodically transmitted with a second cycle longer than the first cycle in the communication system; and A transform window is adjusted in the communication system based on a combination of the time offset and the first arrival path.
14. The method according to claim 13, in, Estimating the time offset based on the first reference signal comprises: generating one or more channel estimates based on the first reference signal; extracting one or more features from the one or more channel estimates; and The time offset is generated by the machine learning process based on the one or more features.
15. The method according to claim 14, in: The one or more channel estimates provide a set of frequency domain channel correlations between channels separated by a predetermined number of subcarriers; and Extracting the one or more features includes extracting a subset of the set of frequency-domain channel correlations.
16. The method according to claim 15, in, The number of candidate frequency-domain channel correlations is based on a signal configuration of the first reference signal and a resource block bundling configuration.
17. The method according to claim 14, in, The machine learning process is based on neural networks.
18. An apparatus for symbol timing recovery, include: a feature extractor configured to extract one or more features from one or more channel estimates based on a first reference signal that periodically recurs with a first period; a trained logic unit configured to generate a first estimated time offset based on the one or more features; a timing recovery block configured to generate a second estimated time offset based on a second reference signal that periodically recurs with a second period that is longer than the first period; and A window processor is configured to adjust a transform window based on a combination of the first estimated time offset and the second estimated time offset.
19. The device according to claim 18, in: The feature extractor comprises a channel correlator; and The trained logic unit includes a neural network.
Citation Information
Patent Citations
Digital communication signal synchronous demodulation method and device
CN111884685A
Regular sign restoring device and method for orthogonal frequency-division multiplexing receiver
CN1433166A
OFDM signal demodulation method and device thereof
US20120114031A1
A method of determining a carrier frequency offset between a transmitter and a receiver
WO2008082367A1
