Channel estimation methods, chips, electronic devices, and storage media

By adopting an adaptive channel estimation method, the calculation mode and number of iterations are dynamically switched according to the real-time signal-to-noise ratio and channel dynamic characteristics, which solves the problem of balancing accuracy and efficiency in wireless channel estimation, adapts to different signal-to-noise ratio scenarios, optimizes the use of hardware resources, and improves the stability and efficiency of channel estimation.

CN123074879APending Publication Date: 2026-11-10GREE ELECTRIC APPLIANCE INC OF ZHUHAI +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611585349.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing wireless channel estimation techniques struggle to balance accuracy and efficiency, especially in high and low signal-to-noise ratio scenarios, where traditional methods fall short in terms of resource utilization and estimation performance.

Method used

By receiving signals from the wireless communication system, the real-time signal-to-noise ratio and channel dynamic characteristics are determined. The calculation mode and number of iterations are dynamically switched, and an adaptive channel estimation method is adopted, including Full-Patch and Quad-Patch modes. The number of iterations is adjusted according to conditions such as signal-to-noise ratio and pilot residuals. The hardware resource utilization is optimized by combining a relaxation self-attention module and a content-aware compression module.

Benefits of technology

It achieves automatic adjustment of calculation mode and iteration number under different signal-to-noise ratio scenarios, reduces resource utilization while maintaining stable estimation performance, realizes dynamic balance between accuracy and efficiency, and adapts to non-stationary and highly dynamic wireless channel environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123074879A_ABST
    Figure CN123074879A_ABST
Patent Text Reader

Abstract

This application provides a channel estimation method, chip, electronic device, and storage medium. The method includes: receiving a received signal from a wireless communication system, and determining a real-time signal-to-noise ratio (SNR) and channel dynamic characteristics based on the received signal; determining a target computation mode and a target number of iterations based on the real-time SNR and channel dynamic characteristics; and determining a channel estimation result for the wireless communication system based on the target computation mode and the target number of iterations. Embodiments of this application can dynamically switch between different computation modes and adjust the number of iterations of the model based on the real-time SNR and channel dynamic characteristics. In high SNR scenarios, the number of iterations is automatically reduced to save power, while in low SNR scenarios, the number of iterations is increased to improve accuracy. Compared to a fixed configuration scheme, this mechanism can maintain stable estimation performance while reducing resource utilization, achieving a dynamic balance between accuracy and efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the technical field of wireless channel estimation, specifically relating to a channel estimation method, chip, electronic device, and storage medium. Background Technology

[0002] Wireless channel estimation is one of the core signal processing tasks in the physical layer (PHY) of wireless communication, and its performance directly determines the transmission quality, throughput, and reliability of the communication system. Finding a balance between estimation accuracy and efficiency has become one of the most pressing problems to be solved. Summary of the Invention

[0003] In view of the above problems, a channel estimation method, chip, electronic device, and storage medium are proposed to overcome or at least partially solve the above problems, including: A channel estimation method, the method comprising: Receive the received signal from the wireless communication system, and determine the real-time signal-to-noise ratio and channel dynamic characteristics based on the received signal; Based on the real-time signal-to-noise ratio and the channel dynamic characteristics, the target calculation mode and the target number of iterations are determined; Based on the target calculation mode and the target iteration number, the channel estimation result for the wireless communication system is determined.

[0004] In some embodiments, determining the target calculation mode and the target number of iterations based on the real-time signal-to-noise ratio and the channel dynamic characteristics includes: Based on the real-time signal-to-noise ratio, determine the average signal-to-noise ratio for the current period and the change in signal-to-noise ratio for adjacent sampling periods; The target calculation mode and the target number of iterations are determined based on the current period's average signal-to-noise ratio, the change in signal-to-noise ratio in adjacent sampling periods, and the channel dynamic characteristics.

[0005] In some embodiments, the method further includes: Based on the channel estimation results, determine the error feedback results; Based on the error feedback results, the target calculation mode and / or the target number of iterations are adjusted.

[0006] In some embodiments, the method further includes: Write the target calculation mode and target iteration number determined for the current cycle into the first register group, and write the calculation mode and iteration number determined for the next cycle into the second register group; When the next cycle arrives, the computation mode and iteration number in the second register set are used to perform channel estimation for the wireless communication system.

[0007] In some embodiments, the target computation mode includes a first computation mode or a second computation mode; determining the channel estimation result for the wireless communication system based on the target computation mode and the target number of iterations includes: When the target calculation mode is the first calculation mode, the current time-frequency resource block with the starting address, row and column size and span configured as a complete block is determined according to the received signal, and the channel estimation result is determined according to the current time-frequency resource block and the target iteration number. When the target calculation mode is the second calculation mode, the current time-frequency resource block with a starting address, row and column size and span configured as a complete block is determined according to the received signal, and the current time-frequency resource block is divided into multiple sub-blocks; the channel estimation result is determined according to the multiple sub-blocks and the target iteration number.

[0008] This application embodiment also provides a channel estimation chip, the chip comprising: The data preprocessing module is used to receive the received signal from the wireless communication system and determine the real-time signal-to-noise ratio and channel dynamic characteristics based on the received signal. An adaptive control module is used to determine the target calculation mode and the target iteration number based on the real-time signal-to-noise ratio and the channel dynamic characteristics. A hardware acceleration module is used to determine the channel estimation result for the wireless communication system based on the target computation mode and the target iteration number.

[0009] In some embodiments, the hardware acceleration module includes: A relaxed self-attention module is used to determine attention enhancement features based on the received signal; The model module is used to determine local shallow time-frequency features and local deep time-frequency features based on the attention enhancement features; and to determine the channel estimation result based on the local shallow time-frequency features and the local deep time-frequency features.

[0010] In some embodiments, the hardware acceleration module further includes: A content-aware compression module is used to load model weights; the model weights are stored in the chip after being compressed in a content-aware manner.

[0011] In some embodiments, the chip further includes: an interface module, the interface module comprising: A data input interface is used to receive signals received by the wireless communication system. A data output interface is used to output the channel estimation results to the baseband processing unit; The control interface is pre-written with a pilot pattern or pilot position mask consistent with the communication protocol.

[0012] In some embodiments, the chip further includes: a storage module, the storage module comprising: A parameter storage area is used to store the parameters that generate the channel estimation results; A data buffer is used to cache data generated during the process of producing the channel estimation results; The result storage area is used to store the channel estimation results.

[0013] This application also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the computer program is executed by the processor, it implements the channel estimation method described above.

[0014] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the channel estimation method described above.

[0015] The embodiments of this application have the following advantages: In this embodiment, the received signal from the wireless communication system is received, and the real-time signal-to-noise ratio (SNR) and channel dynamic characteristics are determined based on the received signal. Based on the real-time SNR and channel dynamic characteristics, the target calculation mode and target iteration count are determined. Based on the target calculation mode and target iteration count, the channel estimation result for the wireless communication system is determined. This embodiment can dynamically switch between different calculation modes and adjust the number of iterations of the model based on the real-time SNR and channel dynamic characteristics. In high SNR scenarios, the number of iterations is automatically reduced to save power, while in low SNR scenarios, the number of iterations is increased to improve accuracy. Compared to a fixed configuration scheme, this mechanism can maintain stable estimation performance while reducing resource utilization, achieving a dynamic balance between accuracy and efficiency. Attached Figure Description

[0016] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart illustrating the steps of a channel estimation method according to an embodiment of this application; Figure 2 This is a flowchart illustrating the steps of another channel estimation method according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a channel estimation chip according to an embodiment of this application; Figure 4 This is a schematic diagram of the structure of another channel estimation chip according to an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a hardware acceleration module according to an embodiment of this application; Figure 6 This is a flowchart illustrating an embodiment of an adaptive deep learning-based method for accelerating wireless channel estimation. Figure 7 This is a schematic diagram of the structure of an adaptive deep learning model according to an embodiment of this application. Detailed Implementation

[0017] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0018] Wireless channel estimation (which is essentially the process of transmitting signals through a spatial wireless link, resulting in amplitude attenuation, phase rotation, and multipath delay) is one of the core signal processing tasks of the physical layer of wireless communication. Its performance directly determines the transmission quality, throughput, and reliability of the communication system.

[0019] In scenarios such as B5G / 6G ultra-reliable and low-latency communications (HRLLC), massive MIMO (Multiple Input Multiple Output), and millimeter-wave communications, the channel exhibits characteristics such as non-stationarity, high dynamics, and strong sparsity.

[0020] The reason why traditional methods cannot achieve both high accuracy and low complexity is that the LS (Least Squares) method only needs to perform division or least squares at the pilot position, which has a low computational cost. However, it does not utilize the time-frequency correlation of the channel and prior statistical information, and will directly amplify noise in low SNR (Signal to Noise Ratio), sparse pilots and fast time-varying scenarios. If the accuracy is improved by increasing the number of pilots, expanding the interpolation neighborhood or increasing the number of iterations, the spectral overhead and processing delay will increase.

[0021] The LMMSE (Linear Minimum Mean Square Error) method can suppress noise using second-order channel statistics, but it requires obtaining and continuously updating the channel covariance matrix. Traditional direct implementations also require matrix inversion or decomposition related to the estimation dimension, resulting in a computational complexity typically of O(N). 3 The storage complexity is O(N). 2 In non-stationary channels, statistical mismatch of covariance can also lead to a decrease in estimation accuracy.

[0022] Therefore, LS is biased towards low complexity but has limited accuracy, while LMMSE is biased towards higher accuracy but has large computational, storage and update overhead. Neither can meet the requirements of future communication systems for extremely high accuracy and extremely low latency at the same time.

[0023] In recent years, deep learning (DL) technology has been widely used in wireless channel estimation. By mining deep features of channel data, it can significantly improve the estimation accuracy in complex channel scenarios, outperforming traditional statistical methods. However, existing deep learning-based channel estimation schemes have significant bottlenecks: In related technologies, deep collaborative optimization between algorithms and hardware has not been achieved, and computing resources cannot be dynamically adjusted according to channel quality, signal-to-noise ratio (SNR), etc., resulting in wasted computing power in high channel quality scenarios and insufficient accuracy in low channel quality scenarios.

[0024] To address this, embodiments of this application provide an adaptive channel estimation method that can balance accuracy and efficiency.

[0025] Reference Figure 1 The diagram illustrates a flowchart of a channel estimation method according to an embodiment of this application, which may include the following steps: Step 101: Receive the received signal from the wireless communication system, and determine the real-time signal-to-noise ratio and channel dynamic characteristics based on the received signal.

[0026] In this embodiment, the received signal from the wireless communication system can be received first. For example, the received signal can be an OFDM (Orthogonal Frequency Division Multiplexing) signal. A wireless communication system refers to a communication system capable of transmitting and receiving OFDM and other radio waveforms over the air, and it can consist of a transmitter, a wireless channel, and a receiver. The received signal can be a time-domain baseband / RF waveform generated by multi-carrier modulation.

[0027] After receiving the received signal, it can be preprocessed and analyzed to obtain the real-time signal-to-noise ratio and channel dynamic characteristics. The channel dynamic characteristics may include frequency band identifiers, stationary state bits, the average signal-to-noise ratio of the current period, and the change in signal-to-noise ratio between adjacent sampling periods. This application embodiment does not limit these characteristics.

[0028] Step 102: Determine the target calculation mode and the number of target iterations based on the real-time signal-to-noise ratio and channel dynamic characteristics.

[0029] After determining the real-time signal-to-noise ratio and channel dynamic characteristics, the target computation mode and target iteration number for the current computation cycle can be determined based on the real-time signal-to-noise ratio and channel dynamic characteristics.

[0030] For example, in high signal-to-noise ratio scenarios, a Full-Patch computing mode can be used, and the number of iterations can be reduced to save power consumption; in low signal-to-noise ratio scenarios, a Quad-Patch computing mode can be used, and the number of iterations can be increased to improve accuracy.

[0031] Step 103: Determine the channel estimation result for the wireless communication system based on the target calculation mode and the target iteration number.

[0032] After determining the target calculation mode and the target number of iterations, the channel estimation result of the wireless communication system can be calculated based on the received signal according to the target calculation mode. During the calculation process, the target number of iterations can be used as the threshold for iteration. When the target number of iterations is reached, the calculated result is output, thus outputting the channel estimation result of the wireless communication system. The channel estimation result is the channel estimate obtained after estimating the channel of the wireless communication system, that is, the channel estimate value of the subcarrier at each pilot position.

[0033] In this embodiment, the received signal from the wireless communication system is received, and the real-time signal-to-noise ratio (SNR) and channel dynamic characteristics are determined based on the received signal. Based on the real-time SNR and channel dynamic characteristics, the target calculation mode and target iteration count are determined. Based on the target calculation mode and target iteration count, the channel estimation result for the wireless communication system is determined. This embodiment can dynamically switch between different calculation modes and adjust the number of iterations of the model based on the real-time SNR and channel dynamic characteristics. In high SNR scenarios, the number of iterations is automatically reduced to save power, while in low SNR scenarios, the number of iterations is increased to improve accuracy. Compared to a fixed configuration scheme, this mechanism can maintain stable estimation performance while reducing resource utilization, achieving a dynamic balance between accuracy and efficiency.

[0034] Reference Figure 2 The diagram illustrates a flowchart of another channel estimation method according to an embodiment of this application, which may include the following steps: Step 201: Receive the received signal from the wireless communication system, and determine the real-time signal-to-noise ratio and channel dynamic characteristics based on the received signal.

[0035] In this embodiment, the received signal from the wireless communication system can be received first. After receiving the received signal, it can be preprocessed and analyzed to obtain the real-time signal-to-noise ratio and channel dynamic characteristics.

[0036] Step 202: Based on the real-time signal-to-noise ratio, determine the average signal-to-noise ratio of the current period and the change in signal-to-noise ratio between adjacent sampling periods.

[0037] After obtaining the real-time signal-to-noise ratio (SNR) for the current period, the mean SNR for the current period, i.e., the mean SNR for the current period, and the change in SNR between adjacent sampling periods can be calculated based on the real-time SNR for the current period.

[0038] For example, if the real-time signal-to-noise ratio is γ(k), then the change in signal-to-noise ratio between adjacent sampling periods is Δγ(k) = |γ(k) - γ(k-1)|; where γ(k-1) is the real-time signal-to-noise ratio of the previous period.

[0039] Based on the real-time signal-to-noise ratio, the mean signal-to-noise ratio γavg(k) for the current period can also be determined using the following formula: γavg(k)=α·γavg(k-1)+(1-α)·γ(k) Wherein, γavg(k-1) is the average signal-to-noise ratio determined in the previous period, and α is a preset weighting coefficient, which can be configured to 0.5-0.85. This application embodiment does not limit this.

[0040] Step 203: Determine the target calculation mode and the target iteration number based on the current period's average signal-to-noise ratio, the change in signal-to-noise ratio in adjacent sampling periods, and the channel dynamic characteristics.

[0041] After determining the average SNR of the current period, the SNR change in adjacent sampling periods, and the channel dynamic characteristics, this application can determine the target calculation mode and the target number of iterations based on the average SNR of the current period, the SNR change in adjacent sampling periods, the channel dynamic characteristics, and the thresholds set for each parameter.

[0042] For example, the channel dynamic characteristics may include the normalized pilot residual e_p(k), which can be calculated by the following formula: e_p(k) = [Σ|Y_p-X_p·H_P| 2 ] / [Σ|Y_p| 2 +ε] Where Y_p is the received pilot, X_p is the locally known pilot, H_P is the channel response at the currently estimated pilot position, and ε is a positive number to prevent the denominator from being zero.

[0043] If γavg(k) ≤ Tlow, or Δγ(k) > Tγ, or e_p(k) > Eth, then the target computation mode is determined to be Full-Patch mode and the number of iterations is increased; if γavg(k) ≥ Thigh, and Δγ(k) ≤ Tγ and e_p(k) ≤ Eth are valid for M consecutive sampling windows, then the target computation mode is determined to be Quad-Patch mode and the number of iterations is reduced.

[0044] Among them, Tlow, Tγ, Eth, and Thigh are preset thresholds.

[0045] A hysteresis range can be set between Tlow and Thigh, for example, Tlow=18dB and Thigh=22dB, to avoid frequent switching.

[0046] Step 204: The target calculation mode includes a first calculation mode or a second calculation mode; when the target calculation mode is the first calculation mode, the current time-frequency resource block with the starting address, row and column size and span configured as a complete block is determined according to the received signal, and the channel estimation result is determined according to the current time-frequency resource block and the target iteration number.

[0047] In some embodiments, the target computing mode may include a first computing mode or a second computing mode; wherein the first computing mode may be a Full-Patch mode; and the second computing mode may be a Quad-Patch mode.

[0048] After determining the target computation mode and the target number of iterations, the channel estimation result can be determined based on the target computation mode and the target number of iterations.

[0049] In some embodiments, if the target computation mode is the first computation mode, the current time-frequency resource block, configured as a complete block with a starting address, row and column size, and span, can be determined based on the received signal. Then, the channel estimation result can be estimated based on this complete current time-frequency resource block and the target iteration number. The first computation mode retains a larger contextual receptive field and is suitable for scenarios with low SNR, large pilot residuals, or rapid channel changes due to high-speed movement.

[0050] Step 205: When the target calculation mode is the second calculation mode, determine the current time-frequency resource block with the starting address, row and column size and span configured as a complete block according to the received signal, and divide the current time-frequency resource block into multiple sub-blocks; determine the channel estimation result according to the multiple sub-blocks and the target iteration number.

[0051] In some embodiments, if the target computation mode is the second computation mode, the current time-frequency resource block, whose starting address, row and column size, and span are configured as a complete block, can be determined first based on the received signal; then, the current time-frequency resource block can be divided into multiple sub-blocks.

[0052] In the second computation mode, the channel estimation result can be determined based on multiple sub-blocks and the number of target iterations. This division does not change the front-end input signal or the final output range. Its function is to reduce the context and cache working set of each attention calculation, thereby reducing the on-chip cache read / write volume and power consumption. It is suitable for scenarios with high SNR and stable channels.

[0053] In this embodiment, a received signal from a wireless communication system is received, and based on the received signal, the real-time signal-to-noise ratio (SNR) and channel dynamic characteristics are determined. Based on the real-time SNR, the average SNR of the current period and the SNR change in adjacent sampling periods are determined. Based on the average SNR of the current period, the SNR change in adjacent sampling periods, and the channel dynamic characteristics, a target calculation mode and a target number of iterations are determined. The target calculation mode includes a first calculation mode or a second calculation mode. When the target calculation mode is the first calculation mode, based on the received signal, a current time-frequency resource block with a starting address, row and column size, and span configured as a complete block is determined, and the channel estimation result is determined based on the current time-frequency resource block and the target number of iterations. When the target calculation mode is the second calculation mode, based on the received signal, a current time-frequency resource block with a starting address, row and column size, and span configured as a complete block is determined, and the current time-frequency resource block is divided into multiple sub-blocks. The channel estimation result is determined based on the multiple sub-blocks and the target number of iterations. This embodiment can dynamically switch between different calculation modes and adjust the number of model iterations based on the real-time SNR and channel dynamic characteristics. In high signal-to-noise ratio (SNR) scenarios, the number of iterations is automatically reduced to save power, while in low SNR scenarios, the number of iterations is increased to improve accuracy. Compared to a fixed configuration scheme, this mechanism can maintain stable estimation performance while reducing resource utilization, achieving a dynamic balance between accuracy and efficiency.

[0054] In some embodiments of this application, any of the above embodiments may further include the following steps: Based on the channel estimation results, determine the error feedback results; based on the error feedback results, adjust the target calculation mode and / or the number of target iterations.

[0055] In some embodiments, the preliminary results can be corrected under the MSE (Mean Square Error) criterion to further suppress the influence of noise. This design not only adapts to the Sub-6GHz band but also effectively addresses the high dynamic characteristics of the millimeter-wave band, enhancing the versatility and stability of the channel estimation method across multiple scenarios.

[0056] Specifically, after obtaining the preliminary channel estimation results, the error feedback results can be calculated; based on the error feedback results, the target calculation mode and / or the number of target iterations can be adjusted.

[0057] In some embodiments of this application, any of the above embodiments may further include the following steps: The target computation mode and target iteration number determined for the current cycle are written into the first register group, and the computation mode and iteration number determined for the next cycle are written into the second register group; when the next cycle arrives, the computation mode and iteration number in the second register group are used to perform channel estimation for the wireless communication system.

[0058] In some embodiments, a first register group and a second register group can be designed; after determining the target computation mode and target iteration number for the current cycle, this information can be written into the first register group; when performing channel estimation based on the information in the first register group, if the received signal for the next cycle is received, the computation mode and iteration number for the next cycle can be determined based on the received signal for the next cycle, and the computation mode and iteration number for the next cycle can be written into the second register group.

[0059] When the channel estimation for the current period is completed, the calculation mode and iteration number in the second register group can be used to perform channel estimation for the wireless communication system, thereby achieving parallel data processing.

[0060] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of this application are not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of this application.

[0061] Reference Figure 3 The diagram illustrates the structure of a channel estimation chip according to an embodiment of this application, which may include: Data preprocessing module 1 is used to receive the received signal from the wireless communication system and determine the real-time signal-to-noise ratio and channel dynamic characteristics based on the received signal. Adaptive control module 2 is used to determine the target calculation mode and the number of target iterations based on the real-time signal-to-noise ratio and channel dynamic characteristics; Hardware acceleration module 3 is used to determine the channel estimation result for the wireless communication system based on the target calculation mode and the target iteration number.

[0062] In this embodiment of the application, the channel estimation chip 300 may include a data preprocessing module 1, an adaptive control module 2, and a hardware acceleration module 3; wherein, the data preprocessing module 1 can be used to receive the received signal of the wireless communication system and determine the real-time signal-to-noise ratio and channel dynamic characteristics based on the received signal.

[0063] After obtaining the real-time signal-to-noise ratio and channel dynamic characteristics, the data preprocessing module 1 can send this information to the adaptive control module 2.

[0064] After obtaining the real-time signal-to-noise ratio and channel dynamic characteristics, the adaptive control module 2 can determine the appropriate target calculation mode and target iteration number for the current period based on the real-time signal-to-noise ratio and channel dynamic characteristics.

[0065] After obtaining the target computation mode and the target iteration number, the hardware acceleration module 3 can process the received signal based on the target computation mode and the target iteration number to obtain the channel estimation result of the wireless communication system.

[0066] In some embodiments of this application, the hardware acceleration module 3 includes: A relaxed self-attention module is used to determine attention enhancement features based on the received signal; The model module is used to determine local shallow time-frequency features and local deep time-frequency features based on attention enhancement features; and to determine the channel estimation result based on the local shallow time-frequency features and local deep time-frequency features.

[0067] In some embodiments, the hardware acceleration module 3 may include a relaxed self-attention module and a model module; it is designed with a unit that integrates relaxed self-attention (RSA) approximation operation and adopts a hybrid processing flow of bilinear interpolation combined with MSE error correction unit 33, which significantly reduces the complexity of hardware implementation while ensuring accuracy.

[0068] For example, the relaxation self-attention module can be used to determine attention enhancement features based on the received signal; for example, the data received by the relaxation self-attention module is compressed feature data obtained by attention processing based on the received signal. After obtaining the compressed feature data, the relaxation self-attention module can perform attention approximation calculation to obtain the attention enhancement features.

[0069] The model module can extract shallow local time-frequency features from the attention-enhanced features, which are characterized by the pilot resource unit (the smallest resource unit on the time-frequency grid specifically used to carry the pilot) and its neighboring resource units (resource units around the pilot).

[0070] In addition, the model module can also extract higher-level local deep time-frequency features from the attention-enhanced features.

[0071] After obtaining the local shallow time-frequency features and the local deep time-frequency features, they can be fused, and then the channel estimation result can be determined based on the fused features.

[0072] In some embodiments of this application, the hardware acceleration module 3 further includes: The content-aware compression module is used to load model weights; the model weights are stored in the chip after being compressed in a content-aware manner.

[0073] In some embodiments, channel adaptive compression technology can be used to compress model weights, and combined with dedicated on-chip SRAM (Static Random Access Memory) partition management, the model weights can be loaded on the entire chip, eliminating the latency and power consumption caused by EMA (External Memory Access).

[0074] In some embodiments, the model weights, after being compressed with content awareness, can be stored in a chip, such as in storage module 4.

[0075] In some embodiments of this application, the chip further includes: an interface module 5, which includes: Data input interface 51 is used to receive signals from a wireless communication system; Data output interface 52 is used to output channel estimation results to the baseband processing unit; The control interface 53 is pre-written with a pilot pattern or pilot position mask consistent with the communication protocol.

[0076] In some embodiments, in addition to the modules described above, the chip may also include an interface module 5, which can be used for data output, input, etc.

[0077] For example, interface module 5 may include data input interface 51, data output interface 52, and control interface 53.

[0078] The data input interface 51 can be used to receive the received signal from the wireless communication system and send the received signal to the data preprocessing module 1 for preprocessing.

[0079] The data output interface 52 can be used to output channel estimation results to the baseband processing unit; specifically, after the hardware acceleration module 3 generates the channel estimation results, it can output them to the baseband processing unit that needs the channel estimation results through the data output interface 52. The baseband processing unit is responsible for performing signal processing in the digital baseband domain.

[0080] The control interface 53 can pre-write a pilot pattern or pilot position mask that is consistent with the communication protocol. The coordinates of the pilot identifiers in the mask constitute a set of pilot positions.

[0081] In some embodiments of this application, the chip further includes: a storage module 4, which includes: Parameter storage area 41 is used to store the parameters used to generate the channel estimation results; Data buffer 42 is used to buffer data during the process of generating channel estimation results; Result storage area 43 is used to store channel estimation results.

[0082] In some embodiments, the chip may further include a storage module 4, which may be used to store threshold parameters and cache data generated during the calculation process.

[0083] For example, parameter storage area 41 can be used to store parameters that generate channel estimation results, such as various thresholds, model weights, etc.

[0084] Data buffer 42 can be used to cache data during the process of generating channel estimation results, such as intermediate results.

[0085] The result storage area 43 can be used to temporarily store the obtained channel estimation results and output them through the data output interface 52.

[0086] The following, combined with Figure 4 - Figure 7 The document also includes specific examples to further illustrate the channel estimation methods and chips described above. Reference Figure 4 This diagram illustrates the structure of another channel estimation chip according to an embodiment of this application. Figure 5 The diagram illustrates the structure of a hardware acceleration module 3 according to an embodiment of this application. Figure 4-5 As shown, the channel estimation chip may include a data preprocessing module 1, an adaptive control module 2, a hardware acceleration module 3, a storage module 4, an interface module 5, and an on-chip bus 6.

[0087] The data preprocessing module 1 includes an SNR detection unit 11; the adaptive control module 2 includes a mode switching agent 21 (MSA); the hardware acceleration module 3 includes an adaptive deep learning operation unit 31, an interpolation operation unit 32, and an error correction unit 33; the adaptive deep learning operation unit 31 includes a CAC (Content-Aware Compression) compression module (i.e., content-aware compression module), an RSA module 312 (i.e., relaxation self-attention module), and an IResNet core unit 313 (i.e., model module); the storage module 4 includes a parameter storage area 41, a data cache area 42, and a result storage area 43; and the interface module 5 includes a data input interface 51, a data output interface 52, and a control interface 53.

[0088] The chip in this example supports Sub-6GHz (FR1) and millimeter wave (FR2 / 3) multi-band adaptation, and its specific structure is as follows: The data preprocessing module 1 adopts a pipelined architecture with a working clock frequency of 1GHz. It supports demodulation processing of OFDM signals, and the pilot extraction accuracy is ≤0.1dB. The SNR detection unit 11 has a detection range of 0-40dB and a detection error of ≤0.5dB. It adopts a 12-bit fixed-point design, which reduces data storage overhead by 60%.

[0089] The adaptive control module 2 uses a microcontroller with a RISC-V (Reduced Instruction Set Computer-Five) architecture, with a working clock frequency of 500MHz and a built-in mode switching agent 21 (MSA).

[0090] MSA includes a channel quality analysis subunit, a mode decision state machine, an iteration number generator, a configuration register group, and a configuration distribution interface. It can switch between Full-Patch and Quad-Patch modes. It supports dynamically adjusting the number of iterations of the deep learning model based on SNR information (configurable from 1 to 10 times), and the iteration step size can be flexibly set to achieve a balance between accuracy and latency.

[0091] Specifically, the SNR detection unit 11 outputs the instantaneous SNR value γ(k) (i.e., real-time signal-to-noise ratio) with a sampling period of one OFDM symbol or one pilot resource block, and calculates the change in signal-to-noise ratio Δγ(k) = |γ(k) - γ(k-1)| between adjacent sampling periods.

[0092] Error correction unit 33 or data preprocessing module 1 calculates the normalized pilot residual e_p(k) = [∑|Y_p-X_p·H_P|] at the known pilot position. 2 ] / [∑|Y_p| 2+ε], where Y_p is the received pilot, X_p is the locally known pilot, H_P is the channel response at the currently estimated pilot position, and ε is a positive number to prevent the denominator from being zero.

[0093] γ(k), Δγ(k), e_p(k), and the frequency band identifier are written together into the configured channel quality status register.

[0094] The channel quality analysis subunit performs a moving average on γ(k) to obtain γavg(k) = α·γavg(k-1) + (1-α)·γ(k), where α can be configured to be 0.5-0.875. When Δγ(k) ≤ Tγ and e_p(k) ≤ Eth for M consecutive sampling windows, the channel is determined to be in a stationary state; when any index exceeds the corresponding threshold, the channel is determined to be in a dynamic state.

[0095] In one example, Tγ = 1 dB / window, Eth = 0.01, and M = 3. This forms the quality vector Q(k) = {γavg(k), Δγ(k), e_p(k), band identifier, stationary state bit}, which serves as the input to the MSA mode decision state machine.

[0096] The mode decision state machine generates control results using a combination of threshold tables and hysteresis decisions. The threshold table is stored in parameter storage area 41 and includes at least the low SNR threshold Tlow, the high SNR threshold Thigh, the SNR change threshold Tγ, the pilot residual threshold Eth, and the number of continuous windows M. If γavg(k) ≤ Tlow, or Δγ(k) > Tγ, or e_p(k) > Eth, then the MSA outputs Full-Patch mode and increases the number of iterations; if γavg(k) ≥ Thigh, and Δγ(k) ≤ Tγ and e_p(k) ≤ Eth are valid for M consecutive sampling windows, then the MSA outputs Quad-Patch mode and decreases the number of iterations.

[0097] Set a hysteresis interval between Tlow and Thigh, for example, Tlow=18dB and Thigh=22dB, to avoid frequent switching when the SNR fluctuates around 20dB.

[0098] The iteration number generator calculates the iteration number Niter of the model based on the quality vector Q(k) and writes it to the iteration number register ITER_REG.

[0099] In one example, the configuration is as follows: when γavg(k) ≥ 22dB and the stationarity criteria of Δγ(k) ≤ Tγ and e_p(k) ≤ Eth are satisfied for M consecutive sampling windows, Niter is taken 1-3 times; when 18dB < γavg(k) < 22dB, Niter is taken 4-6 times. When γavg(k) ≤ 18dB, or Δγ(k) > Tγ, e_p(k) > Eth and the channel is determined to be changing rapidly, Niter is taken 7-10 times. If the MSE fed back by the error correction unit 33 is greater than the error threshold, Niter is increased 1-2 times in the next processing window; if the MSE is lower than the error threshold for M consecutive windows, Niter is decreased by 1 time until the lower limit allowed by this mode is reached.

[0100] The configuration register group is located inside the MSA of the adaptive control module 2. It includes the MODE_REG (mode selection value), ITER_REG (iteration number), PE_MASK_REG (processing unit enable mask), PATCH_CFG_REG (patch start address, size and span), ERR_TH_REG (error threshold) and CFG_VALID (configuration validity flag) bits, and sets two physical register groups, the shadow register group and the effective register group, for each configuration field.

[0101] MODE_REG is used to select the Full-Patch or Quad-Patch data path, ITER_REG is used to limit the number of IResNet loops, PE_MASK_REG is used to determine the number of PEs and groups involved in the calculation, PATCH_CFG_REG is used to control the read and write-back range of the address generator, ERR_TH_REG is used to determine whether the error stop or feedback adjustment conditions are met, and CFG_VALID is used to indicate whether the above fields can be put into execution as a complete configuration.

[0102] The effective register group continuously outputs the current control parameters to the hardware acceleration module 3. The current control parameters are the MODE_REG mode selection value, ITER_REG iteration count, PE_MASK_REG processing unit enable mask, PATCH_CFG_REG patch start address, size and span, ERR_TH_REG error threshold, and CFG_VALID configuration valid flag, which have been latched by hardware atom and are being executed.

[0103] The shadow register set allows MSA to write complete parameters for the next processing task in the background, thus not changing the configuration being executed during the write process.

[0104] Patch computation refers to the process by which the hardware acceleration module 3 selects a rectangular time-frequency feature block formed by the data preprocessing module 1 for the current time-frequency resource block from the data buffer 42 according to the PATCH_CFG_REG in the current effective configuration, and sequentially completes the process of address reading, IResNet / RSA feature extraction, interpolation or intermediate feature writing back. The time and frequency dimensions of the rectangular time-frequency feature block (referring to the current patch) are determined by the patch's starting address, row and column dimensions, and span, respectively. The normalized features of the rectangular time-frequency feature block serve as input to the PE cluster, and the calculation results of the rectangular time-frequency feature block are used to form preliminary channel estimation of pilot positions or complete time-frequency grid results. The normalized features are multi-channel tensor data for a specific patch (time-frequency block) that have undergone normalization processing and incorporated received signal (real / imaginary parts), preliminary LS estimation results, relative position coding, pilot mask, and SNR information.

[0105] After the last intermediate feature is written back, the PATCH_DONE signal is generated. After the MSA completes the decision, it first writes to the shadow configuration register and sets CFG_VALID when PATCH_DONE or pipeline is cleared. Hardware acceleration module 3 latches all fields of the shadow register group to the effective register group at the same clock boundary, and then returns CFG_ACK. The next patch runs with the new configuration.

[0106] This double-buffering configuration mechanism can prevent the time-sharing updates of fields such as MODE_REG, ITER_REG, and address boundaries from corrupting the pilot data and intermediate feature maps being processed.

[0107] Hardware acceleration module 3 adopts a reconfigurable PE (Processing Element) cluster architecture, which can contain 32 PE processing units, an address generator, a data selector, and an intermediate feature cache, with an operating clock frequency of 1.2GHz. The adaptive deep learning computation unit 31 integrates a lightweight IResNet model (8-layer structure, replacing the traditional 23-layer CNN). Through CAC compression technology, the model weights are compressed from 1024KB to 51.2KB, achieving full on-chip SRAM loading and eliminating external memory access. The model weights refer to the same set of IResNet convolutional weights and biases read from the parameter storage area 41 and loaded into the on-chip SRAM by the CAC compression module 311. These weights are reused in each loop, and the set of parameters is not reloaded or changed between loops.

[0108] The RSA module 312 uses a first-order Taylor expansion to approximate the Softmax operation and replaces GeLU with ReLU, reducing the latency of the MHA (Multi-Head Attention) part by 83%. The interpolation unit 32 adopts a bilinear interpolation algorithm with an interpolation delay of ≤0.1μs, and is adapted to different pilot patterns (pilot density is configurable). The error correction accuracy of the error correction unit 33 is ≤0.01, and the feedback delay is ≤5ns.

[0109] The address generator generates the read and write addresses of the data buffer 42 according to PATCH_CFG_REG. The data selector selects the data entry point of the complete patch or the four sub-patches according to MODE_REG. The intermediate feature buffer is used to store the output of each layer and each iteration of IResNet.

[0110] The current time-frequency resource block is a set of rectangular resource units consisting of N_t OFDM symbols × N_f subcarriers formed by the data preprocessing module 1 within a processing window.

[0111] Among them, the pilot resource element is a resource element in the pilot position set P that carries the known reference symbol X_p of both the transmitter and receiver. Its received value Y_p is used for LS coarse estimation, pilot residual calculation and as the anchor point for network channel estimation. The pilot resource element is a resource element taken from the pilot position set P that carries the known reference symbol X_p.

[0112] A data resource unit is a resource unit in the data location set D that carries service data. Its received value Y_d is used for subsequent equalization and demodulation and can provide the received signal context around the pilot.

[0113] The data preprocessing module 1 normalizes or encodes the above-mentioned Y_p, Y_d, known pilot values ​​(referring to the known reference symbol X_p agreed upon by the transmitting and receiving parties in advance), LS coarse estimation, pilot position mask, relative position encoding and SNR respectively, and combines them by channel to form a normalized feature tensor corresponding to the current time-frequency resource block position; the normalized feature is written into the data buffer 42 before processing, and then read by the address generator and used as the input of the RSA module 312 and the IResNet core unit 313. After processing, a complete channel estimation grid corresponding one-to-one with the resource unit position is formed.

[0114] A patch is a rectangular feature block that the hardware acceleration module 3 uses to schedule and compute the resource block.

[0115] In Full-Patch mode, PATCH_CFG_REG configures the starting address, row and column size, and span of the current time-frequency resource block as a complete patch.

[0116] The address generator is an address generation circuit within hardware acceleration module 3, which performs continuous burst readings of the entire two-dimensional time-frequency grid according to the parameters described above. Pilot characteristics are derived from pilot resource units and include at least the real and imaginary parts of the received pilot, known pilot values, and a coarse LS estimate, used to constrain and output a preliminary channel estimate of the pilot position. The neighborhood features are derived from the pilot resource units and data resource units surrounding the pilot, and include at least the received signal features of adjacent resource units, pilot position masks, relative position codes and SNR features. These features are used to provide local time-frequency context to the RSA module 312 and the IResNet core unit 313 to extract the aforementioned local time-frequency features.

[0117] Continuous burst read refers to the address generator continuously retrieving the aforementioned pilot features and neighborhood features from the data buffer 42 in adjacent address order and forming a regular data stream, which is then directly sent to the PE cluster, thereby expanding the context range available for a single calculation and improving the utilization rate of on-chip storage bandwidth.

[0118] All 32 PE processing units are enabled. The IResNet core unit 313 performs feature extraction on the entire patch, and the RSA module 312 performs attention approximation calculation across the entire patch. This mode preserves a larger contextual receptive field and is suitable for scenarios with low SNR, large pilot residuals, or rapid channel changes due to high-speed movement.

[0119] In Quad-Patch mode, the input is still the current time-frequency resource block defined in the previous stage, and it has the same start time, subcarrier range, normalized feature source and output grid range as Full-Patch mode; the two modes process the same batch of resource unit data, the only difference is the hardware scheduling granularity.

[0120] PATCH_CFG_REG divides the current time-frequency resource block into four sub-patches according to the midline of the time dimension and the midline of the frequency dimension, and records the starting address, size, and span of each sub-patches. Configurable overlap boundaries are set between sub-patches, such as overlapping by two pilot intervals or two resource blocks, to maintain boundary continuity. PE_MASK_REG enables 16 or 32 PE processing units and divides them into four PE groups. The four PE groups process the four sub-patches respectively, using the same set of compressed model weights. The output is then stitched together by interpolation unit 32 according to boundary weighted averaging or priority coverage rules to form a complete time-frequency grid result corresponding to the original time-frequency resource block. This division does not change the front-end input signal or the final output range. Its function is to reduce the context and cache working set of each attention calculation, thereby reducing on-chip cache read / write operations and power consumption, which is suitable for high SNR and channel stability scenarios.

[0121] The switching between Full-Patch and Quad-Patch is controlled by MODE_REG, which controls the data selector, address generator step size, PE group enable mask, and intermediate buffer write-back address, respectively. The data selector determines whether to send a complete set of Patch data streams or four sets of sub-Patch data streams to the PE cluster. The address generator step size determines whether to address continuously across the entire grid or address separately for each of the four sub-regions; The PE group enable mask determines whether to enable all PEs or enable PEs in groups; The write-back address of the intermediate cache determines whether the feature is written to the full-patch contiguous cache segment or the four sub-patch cache segments.

[0122] The decision is made based on the real-time SNR and channel dynamic characteristics, while MODE_REG is merely a hardware control register that carries and executes the decision result. MSA first generates the mode control word based on the quality vector Q(k), and then encodes and writes it into MODE_REG.

[0123] MODE_REG=01 indicates the selection of the Full-Patch data path, and MODE_REG=10 indicates the selection of the Quad-Patch data path. These two values, along with the aforementioned adaptive decision-making logic, relate to the sequence of decision logic, control coding, and hardware execution. ITER_REG directly serves as the upper limit of the loop count for the IResNet core unit 313, and PE_MASK_REG controls the entry into clock-gated state when no PE is used. Therefore, The adaptive control module 2 not only provides the mode selection result, but also changes the actual computing scale and energy consumption status of the hardware acceleration module 3 through executable register configuration.

[0124] Storage module 4 uses on-chip SRAM with a total capacity of 34KB. The parameter storage area 41 has a capacity of 31KB (storing model weights, fitting parameters, etc.), the data cache area 42 has a capacity of 3KB, and the result storage area 43 has a capacity of 1KB. It supports high-speed read and write with a read / write latency of ≤1ns and no external memory access (Zero EMA).

[0125] Interface module 5 supports the LVDS interface standard. The input rate of data input interface 51 is 14.4Gb / s, the output rate of data output interface 52 is 14.4Gb / s, and the control interface 53 adopts the SPI interface, which supports online configuration and status monitoring of chip parameters.

[0126] Reference Figure 6 The diagram illustrates a flowchart of an accelerated wireless channel estimation method based on adaptive deep learning, according to an embodiment of this application; see reference. Figure 7 The diagram shows a structural schematic of an adaptive deep learning model according to an embodiment of this application: Step 1: Data Acquisition and Preprocessing. The OFDM received signal (rate of 14.4 Gb / s) of the wireless communication system is received through the data input interface 51 of the interface module 5. The data preprocessing module 1 sequentially completes timing / frequency offset synchronization, cyclic prefix removal, serial-to-parallel conversion and N-point FFT, transforming the time-domain sample into the frequency-domain receive resource grid Y(l, n), where l is the OFDM symbol index and n is the subcarrier index.

[0127] The control interface 53 is pre-written with a pilot pattern or pilot position mask consistent with the communication protocol. The coordinates of the pilot identifier in the mask constitute the pilot position set P. The data preprocessing module 1 reads the received value Y_p=Y(l,n) on P from Y(l,n) according to the mask, where (l,n) belongs to P, forming a pilot signal, and combines the real part, imaginary part and corresponding position encoding of Y_p into the aforementioned pilot feature.

[0128] The transmitting symbol X_p corresponding to the pilot signal is a known reference symbol agreed upon in advance by both the transmitting and receiving parties, which is used to calculate the LS coarse estimate and the pilot residual; The aforementioned pilot features, X_p, and LS coarse estimates are written together into the data buffer 42, and then used as inputs to the RSA module 312 and the IResNet core unit 313, and used to constrain the pilot position for preliminary channel estimation.

[0129] The effective service data coordinates, excluding the positions of the protection subcarrier, DC subcarrier, and pilot, constitute the data location set D. The data preprocessing module 1 reads the received value Y_d=Y(l,n) on D from Y(l,n), where (l,n) belongs to D, forming a data signal.

[0130] Y_d is retained for subsequent equalization and demodulation. Furthermore, the real and imaginary parts of Y_d within a predetermined time-frequency window around the pilot position are encoded and combined with their corresponding relative positions to form the pilot neighborhood reception feature. This feature, along with the pilot position mask and SNR feature, is written into the data buffer 42 for the RSA module 312 to calculate attention and the IResNet core unit 313 to extract local time-frequency features. The SNR detection unit 11 collects the current channel's signal-to-noise ratio information in real time. Assuming the current SNR is 15dB, the standardized channel input data (i.e., the received signal) and SNR information (15dB) are sent to the adaptive control module 2 via the on-chip bus 6 and written into the data buffer 42 for the hardware acceleration module 3 to read.

[0131] Step 2: Adaptive Mode and Parameter Configuration. Channel mobility characteristics are not externally preset quantities, but are obtained by the channel quality analysis subunit based on a coarse estimate of the pilot channels in adjacent processing windows. Let H_P(k) and H_P(k-1) be the pilot channel vectors of the current window and the previous window, respectively. Calculate the normalized time correlation coefficient: ρ(k) = |∑H_P (k)·H_P(k-1)|] / [||H_P(k)||·||H_P(k-1)||+ε] in The complex conjugate is represented by ρ(k). Since ρ(k) represents the normalized correlation of the pilot channel vectors of adjacent windows, the closer its value is to 1, the more stable the channel is. Therefore, we directly take the complementary quantity μ(k) = 1 - ρ(k) from the calculation result of the previous formula, and obtain a mobility index that the larger the value, the lower the time correlation and the stronger the mobility. Therefore, μ(k) is a quantity derived from the aforementioned ρ(k) formula.

[0132] The channel quality analysis subunit compares μ(k) with the threshold Tμ, and combines the comparison result with Δγ(k), pilot residual and Doppler frequency shift estimated by phase change to form the channel mobility characteristics in the quality vector Q(k), which is used by MSA to select the patch mode and the number of iterations.

[0133] In one example, the SNR is 15dB, and the subsequent readings show Δγ = 2.8dB / window, pilot residual = 0.018, and μ(k) > Tμ, thus indicating strong channel dynamics. Based on this, the adaptive control module 2 selects the Full-Patch mode via the mode switching agent 21 (MSA) and configures the computational parameters of the hardware acceleration module 3: the deep learning model iteration count is 8, the PE cluster uses 32 PE processing units, the interpolation algorithm step size is 2, and the error correction threshold is 0.01.

[0134] In this step, the specific processing procedure for MSA is as follows: First, the channel quality parsing subunit reads the channel quality status register and obtains γavg=15dB, Δγ=2.8dB / window, and pilot residual=0.018. Second, the mode decision state machine compares γavg with the threshold table. Since γavg≤Tlow and the pilot residual is greater than Eth, a Full-Patch control word is generated. Third, the iteration number generator generates Niter=8 based on the low SNR configuration range and combines it with the error threshold to generate ERR_TH_REG=0.01. Finally, the MSA writes MODE_REG=01, ITER_REG=8, PE_MASK_REG=0xFFFF_FFFF, and PATCH_CFG_REG=the complete patch address range into the shadow configuration register.

[0135] Once the patch calculation for the current OFDM symbol is complete, MSA sets CFG_VALID, and hardware acceleration module 3 returns CFG_ACK and loads the above configuration. After loading, the address generator reads the standardized channel input data from data buffer 42 according to the Full-Patch address range, all 32 PE processing units are activated, and no external storage access occurs.

[0136] Step 3: Deep learning channel estimation acceleration computation. Hardware acceleration module 3 performs the computation according to the configuration parameters: (1) The CAC compression module 311 of the adaptive deep learning operation unit 31 reads the compressed IResNet weights from the parameter storage area 41 and loads them into the on-chip SRAM; RSA module 312 uses a first-order Taylor expansion to approximate the Softmax operation, reducing computational complexity; The IResNet core unit 313 reads ITER_REG=8 and iteratively extracts features from the standardized channel input data with an upper limit of 8 iterations. In each iteration, the PE cluster performs convolution, residual accumulation, and attention approximation calculations in parallel according to PE_MASK_REG, and outputs the preliminary channel estimation results of the pilot positions with a computation delay of 0.5μs.

[0137] In the Full-Patch mode of this embodiment, the complete patch covers the pilot neighborhood corresponding to the 1024×256 time-frequency grid, and the address generator generates read and write addresses according to the starting address, row span, and column span given by PATCH_CFG_REG. The intermediate feature map is written to the Full-Patch cache segment of data buffer 42. If MODE_REG is switched to Quad-Patch, the address generator generates four sets of sub-Patch addresses. The PE cluster processes these addresses in parallel as four PE groups, and retains overlapping features at the boundaries of the sub-Patches for subsequent stitching.

[0138] (2) The interpolation unit 32 uses a bilinear interpolation algorithm to interpolate the preliminary estimation result of the pilot position to the entire time-frequency grid (assuming the time-frequency grid is 1024×256) to obtain the preliminary channel estimation result in the entire time-frequency domain. The interpolation delay is 0.1μs.

[0139] (3) The error correction unit 33 performs error correction on the full-time frequency domain preliminary channel estimation results output by the interpolation unit 32 based on the MSE criterion, and integrates the coarse estimation given by the LS estimator at the pilot position to obtain the optimized complete channel estimation grid. The optimized complete channel estimation grid, referred to as the optimization result in this paper, is then stored in the result storage area 43 and output to the baseband processing unit, with an error correction delay of 0.05μs.

[0140] The error information is not externally provided, but is calculated online by the error correction unit 33 using known pilots: The corresponding estimated value is extracted from the above optimization results according to the pilot position set P and denoted as H_P. H_P is multiplied by the locally known pilot X_p to obtain the reconstructed pilot X_p·H_P, which is then compared with the received pilot Y_p extracted by the data preprocessing module 1 to calculate the normalized mean square error (i.e., the error feedback result) MSE = [∑|Y_p-X_p·H_P|]. 2 ] / [∑|Y_p| 2 +ε]; This value can also be weighted and fused with the deviation of the optimization result from the coarse LS estimate at the pilot position. The error correction unit 33 writes the MSE=0.008 obtained in this embodiment into the error feedback register ERR_FB_REG and feeds it back to the adaptive control module 2 through the on-chip bus 6 to determine whether the current error threshold is met and to adjust the iteration number and patch mode of the next processing window.

[0141] Step 4: Result Output and Feedback Optimization. Hardware acceleration module 3 stores the final channel estimation result in the result storage area 43 of storage module 4, and outputs it to the baseband processing unit through data output interface 52; adaptive control module 2 reads the error information written by error correction unit 33 in ERR_FB_REG. Since MSE=0.008<threshold 0.01, the current Full-Patch parameter configuration is maintained to ensure the stability of the next round of channel estimation.

[0142] If an SNR increase is detected and maintained at 25dB for three consecutive processing windows, and the pilot residual is less than 0.01, then the MSA meets the Quad-Patch switching condition.

[0143] At this point, MSA updates MODE_REG to 10, ITER_REG to 2, PE_MASK_REG to a mask that enables only 16 PE processing units, and PATCH_CFG_REG to the starting address, size, and overlapping boundary of the four sub-patches. After CFG_ACK returns, hardware acceleration module 3 runs according to the Quad-Patch data path, and the processing results of the four sub-patches are output after boundary weighted concatenation.

[0144] If the SNR drops below 18dB again during Quad-Patch mode operation, or if the error correction unit 33 reports an MSE exceeding 0.01 for two consecutive windows, then the MSA will revert to Full-Patch mode and the Niter will be increased to 7-10. Through the closed-loop process of acquisition, decision-making, register issuance, hardware execution, and error feedback described above, this embodiment clearly implements the technical means of dynamically configuring computing resources according to changes in the channel environment.

[0145] The IResNet core layer (5:2) consists of a first residual block, a second residual block, and an output layer connected in sequence, as well as a residual branch spanning the two residual blocks and a residual fusion branch that performs element-wise addition at the output. The lightweight residual network (8 layers) represents the eight computational layers inside the IResNet core unit 313 after being expanded according to operations such as convolution, activation, and output mapping.

[0146] The RSA module 312 receives the compressed feature data output by the CAC compression module 311 and performs attention approximation calculation. The output of the RSA module 312 is the attention-enhanced feature. This attention-enhanced feature is used as the input of the IResNet core unit 313 for subsequent residual feature extraction.

[0147] The first residual block extracts shallow local time-frequency features characterized by the pilot resource unit and its neighboring resource units from the attention-enhanced features output by the RSA module 312. The second residual block extracts higher-level local time-frequency features based on this; the local time-frequency features reflect the channel correlation between adjacent OFDM symbols and subcarriers, and are used by the output layer to generate preliminary channel estimates for pilot positions.

[0148] The main branch feature is a deep feature obtained by passing the attention enhancement feature through the first residual block and the second residual block in sequence, while the short-circuit feature is a shallow pilot feature retained by the residual branch bypassing the above residual blocks. The residual fusion branch adds the two (main branch features and short-circuit features) element-wise to obtain the fused features. The output layer linearly combines the channels of the fused features through convolutional weights and outputs two channels, which serve as the real and imaginary parts of the complex channel response at the pilot position, respectively, for the interpolation unit 32 to extend to the complete time-frequency grid. ITER_REG controls the above core layer to execute 1-10 times in a loop; The fused features generated in each loop are written to the data buffer 42 as intermediate features of the loop, and read back by the IResNet core unit 313 as input in the next loop. After reaching the number of iterations or the error threshold configured by ITER_REG, the output layer result of the last loop is sent to the interpolation unit 32. This structure uses residual shorting to alleviate the degradation of deep network training, and reduces the computational load and on-chip storage overhead by reducing the number of network layers and reusing weights in the loop.

[0149] In one embodiment, the hardware acceleration module 3 employs a dedicated ASIC (Application-Specific Integrated Circuit) design. Alternatively, the hardware acceleration module 3 can be designed as a reconfigurable logic circuit based on an FPGA (Field-Programmable Gate Array). In this approach, the IResNet model, interpolation algorithm, and error correction logic are all stored in the FPGA's configuration memory in Verilog / VHDL code format.

[0150] The adaptive control module 2 alters hardware behavior by modifying the FPGA's configuration registers or reloading the bitstream. While this approach is slightly less efficient in terms of static area than ASICs, it offers exceptional flexibility, enabling rapid algorithm migration between different generations of communication standards (such as the evolution from 5G NR to 6G), making it suitable for prototype verification or small-batch customized production.

[0151] To further reduce power consumption and area, more aggressive model compression strategies can be adopted. In addition to CAC compression, Quantization-Aware Training (QAT) can be introduced. The weights and activation values ​​of the IResNet model are forced to be quantized from floating-point numbers (FP16 / FP32) to INT8 or even INT4 precision. At the hardware level, the multipliers in the PE unit are replaced with low-bit-width integer multipliers or adder trees.

[0152] At this point, the capacity of the parameter storage area 41 in storage module 4 can be further reduced. Although this will slightly sacrifice estimation accuracy (MSE may increase slightly), in high signal-to-noise ratio scenarios, this loss of accuracy is usually within an acceptable range, while it can bring significant energy efficiency improvements, making it suitable for battery-powered edge computing nodes or IoT terminal devices.

[0153] The above examples address the problems of high computational complexity, large memory consumption, and long execution latency in existing deep learning channel estimation algorithms, as well as the lack of flexibility in existing hardware architectures, their inability to adaptively adjust computing resources according to dynamic channel changes, and the high power consumption and high latency caused by frequent external memory access.

[0154] Specifically, on the one hand, this example achieves deep co-optimization between the algorithm and hardware, eliminating the external memory access bottleneck: CAC compression technology can compress model weights, and combined with an on-chip SRAM storage architecture, on-chip loading of model weights is achieved, eliminating external memory access. This design reduces the chip's channel estimation latency.

[0155] On the other hand, this example can also achieve adaptive resource scheduling based on channel state, balancing accuracy and energy efficiency: through the built-in adaptive control module 2 and SNR detection unit 11, the chip can dynamically switch between Full-Patch and Quad-Patch modes and adjust the number of model iterations (1-10 times) according to the real-time signal-to-noise ratio and channel dynamic characteristics. In high signal-to-noise ratio scenarios, the number of iterations is automatically reduced to save power, while in low signal-to-noise ratio scenarios, the number of iterations is increased to improve accuracy. Compared with a fixed configuration scheme, this mechanism reduces resource utilization while maintaining stable estimation performance, achieving a dynamic balance between accuracy and efficiency.

[0156] On the other hand, this example reduces computational complexity and improves hardware efficiency: it replaces the traditional deep CNN with a lightweight IResNet model, simplifies the Softmax nonlinear operation by combining RSA approximation, and uses bilinear interpolation instead of complex transposed convolution for channel interpolation. These optimizations significantly reduce the computational load on the PE cluster.

[0157] Furthermore, this example enhances the system's robustness and generalization ability: by fusing the output of a traditional LS estimator through error correction unit 33 and correcting the initial results under the MSE (Mean Square Error) criterion, the impact of noise is further suppressed. This structure not only adapts to the Sub-6GHz band but also effectively addresses the high dynamic characteristics of the millimeter-wave band, enhancing the chip's versatility and stability across multiple scenarios.

[0158] This application also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the computer program is executed by the processor, it implements the channel estimation method described above.

[0159] This application also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, it implements the channel estimation method described above.

[0160] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0161] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0162] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of this application can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0163] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0164] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0165] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0166] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0167] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0168] The above provides a detailed description of the channel estimation method, chip, electronic device, and storage medium. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for estimating a channel, characterized in that, The method includes: Receive the received signal from the wireless communication system, and determine the real-time signal-to-noise ratio and channel dynamic characteristics based on the received signal; Based on the real-time signal-to-noise ratio and the channel dynamic characteristics, the target calculation mode and the target number of iterations are determined; Based on the target calculation mode and the target iteration number, the channel estimation result for the wireless communication system is determined.

2. The method according to claim 1, characterized in that, The step of determining the target calculation mode and the target iteration number based on the real-time signal-to-noise ratio and the channel dynamic characteristics includes: Based on the real-time signal-to-noise ratio, determine the average signal-to-noise ratio for the current period and the change in signal-to-noise ratio for adjacent sampling periods; The target calculation mode and the target number of iterations are determined based on the current period's average signal-to-noise ratio, the change in signal-to-noise ratio in adjacent sampling periods, and the channel dynamic characteristics.

3. The method according to claim 1, characterized in that, The method further includes: Based on the channel estimation results, determine the error feedback results; Based on the error feedback results, the target calculation mode and / or the target number of iterations are adjusted.

4. The method according to claim 1, characterized in that, The method further includes: Write the target calculation mode and target iteration number determined for the current cycle into the first register group, and write the calculation mode and iteration number determined for the next cycle into the second register group; When the next cycle arrives, the computation mode and iteration number in the second register set are used to perform channel estimation for the wireless communication system.

5. The method according to claim 1, characterized in that, The target calculation mode includes a first calculation mode or a second calculation mode; determining the channel estimation result for the wireless communication system based on the target calculation mode and the target iteration number includes: When the target calculation mode is the first calculation mode, the current time-frequency resource block with the starting address, row and column size and span configured as a complete block is determined according to the received signal, and the channel estimation result is determined according to the current time-frequency resource block and the target iteration number. When the target calculation mode is the second calculation mode, the current time-frequency resource block with a starting address, row and column size and span configured as a complete block is determined according to the received signal, and the current time-frequency resource block is divided into multiple sub-blocks; the channel estimation result is determined according to the multiple sub-blocks and the target iteration number.

6. A channel estimation chip, characterized in that, The chip includes: The data preprocessing module is used to receive the received signal from the wireless communication system and determine the real-time signal-to-noise ratio and channel dynamic characteristics based on the received signal. An adaptive control module is used to determine the target calculation mode and the target iteration number based on the real-time signal-to-noise ratio and the channel dynamic characteristics. A hardware acceleration module is used to determine the channel estimation result for the wireless communication system based on the target computation mode and the target iteration number.

7. The chip according to claim 6, characterized in that, The hardware acceleration module includes: A relaxed self-attention module is used to determine attention enhancement features based on the received signal; The model module is used to determine local shallow time-frequency features and local deep time-frequency features based on the attention enhancement features; and to determine the channel estimation result based on the local shallow time-frequency features and the local deep time-frequency features.

8. The chip according to claim 7, characterized in that, The hardware acceleration module also includes: A content-aware compression module is used to load model weights; the model weights are stored in the chip after being compressed in a content-aware manner.

9. The chip according to claim 6, characterized in that, The chip further includes: an interface module, the interface module comprising: A data input interface is used to receive signals received by the wireless communication system. A data output interface is used to output the channel estimation results to the baseband processing unit; The control interface is pre-written with a pilot pattern or pilot position mask consistent with the communication protocol.

10. The chip according to claim 6, characterized in that, The chip further includes a storage module, the storage module comprising: A parameter storage area is used to store the parameters that generate the channel estimation results; A data buffer is used to cache data generated during the process of producing the channel estimation results; The result storage area is used to store the channel estimation results.

11. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the channel estimation method as described in any one of claims 1 to 5.

12. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the channel estimation method as described in any one of claims 1 to 5.