A method and accelerator for channel state information feedback for edge devices
By employing a lightweight neural network and parallel quantization processing with cosine function companding mapping on an edge device, combined with a TPU and dedicated hardware circuitry, the problem of real-time, stable, and low-power CSI feedback in existing technologies is solved, achieving low-latency and high-reliability CSI feedback.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-04-15
- Publication Date
- 2026-07-10
AI Technical Summary
Existing deep learning CSI feedback schemes are difficult to meet the requirements of real-time performance, low power consumption, and high reliability on edge devices. In particular, the complex network structure, floating-point computation, and large-scale matrix operation overhead, as well as the lack of dedicated hardware design and hardware-software co-engineering architecture, make it difficult to implement the algorithm efficiently.
A lightweight neural network structure and cosine function companding mapping are adopted, combined with parallel quantization processing and hardware-software co-engineering architecture. Through an end-to-end architecture of encoder-quantizer-decoder, the computational complexity is reduced and stable feature representation under low precision is achieved. Parallel quantization processing is performed using TPU and dedicated hardware circuit modules to generate feedback bitstream.
Low-latency, high-stability, and low-power CSI feedback was implemented on edge devices, reducing computational load and parameter scale, and meeting the low-latency and high-reliability requirements of actual systems.
Smart Images
Figure CN122372167A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wireless communication and artificial intelligence accelerators, and more particularly to a method and accelerator for channel state information feedback for edge devices. Background Technology
[0002] In frequency division duplex massive MIMO systems, channel state information (CSI) feedback is crucial for downlink precoding and link adaptation. Current mainstream solutions fall into three categories: codebook-based, compressed sensing, and deep learning. Codebook-based feedback relies on predefined codebooks and index searches, offering good standard compatibility, but its complexity and overhead increase dramatically in massive MIMO scenarios. Compressed sensing methods utilize channel sparsity to achieve low-dimensional feedback, but reconstruction accuracy decreases in complex channel environments, and the iterative computation at the base station is substantial. Deep learning end-to-end encoding / decoding schemes offer superior accuracy and do not rely on sparsity assumptions, making them the current mainstream research direction. However, existing solutions largely remain at the algorithm and software simulation level. When deployed on edge user devices, they generally face constraints in computing power, power consumption, and real-time performance, making it difficult to meet the low-latency, high-reliability CSI feedback requirements of real-world systems.
[0003] Current deep learning CSI feedback still suffers from several technical shortcomings: the network structure based on Transformer and attention mechanisms is complex, with high overhead from floating-point computation and large-scale matrix operations, and low-precision inference is prone to performance degradation; traditional codeword quantization operations are non-differentiable, and the coarse gradient approximation leads to instability in end-to-end training, with the accumulation of multi-level quantization errors further reducing reconstruction accuracy; codeword quantization involves a large number of comparison, judgment, and memory access operations, resulting in low execution efficiency on general-purpose processors and easily creating latency bottlenecks. Meanwhile, existing hardware acceleration focuses primarily on neural network inference, lacking dedicated hardware designs and hardware-software co-architectures for codeword quantization, making it difficult to efficiently deploy algorithms and achieve real-time, stable, and low-power CSI feedback on edge devices. Summary of the Invention
[0004] The present invention aims to at least partially solve one of the technical problems in the related art.
[0005] Therefore, the first objective of this invention is to provide a method for channel state information feedback for edge devices, comprising the following steps:
[0006] S1, acquire channel state information, and encode the channel state information through a neural network encoder to obtain continuous codeword data; S2, The continuous codeword data is compressed and expanded using the cosine function to map the data distribution to the compressed domain, thus obtaining compressed domain data; S3, the compressed domain data is subjected to parallel quantization processing by the quantization processing unit, and continuous values are mapped to discrete quantization indices based on a preset quantization level; S4. Generate a feedback bit stream based on the discrete quantization index and send it to the receiving end via the uplink.
[0007] In one embodiment of the present invention, S1 includes: S11, deploy an encoder network on the user equipment side and a decoder network on the base station side to build an end-to-end architecture of encoder-quantizer-decoder; S12, the spatial correlation and global characteristics of the channel state information are modeled by lightweight attention and feedforward structure to obtain a continuous codeword representation suitable for subsequent codeword quantization processing; S13 avoids using computational forms that are highly sensitive to numerical precision in its structural selection and operator design, and reduces its reliance on high-cost operators such as exponentiation and division.
[0008] In one embodiment of the present invention, S13 includes: S131 adopts the concept of quantization perception to fully consider the subsequent low-precision weight quantization deployment requirements in the network design stage, so that the network can still maintain stable feature expression ability under low-precision conditions. S132 employs a unified quantization-aware training strategy that integrates network weight quantization and codeword quantization, preventing errors from accumulating continuously during multi-level quantization.
[0009] In one embodiment of the present invention, S2 includes: S21, First, the continuous codewords output by the encoder are subjected to companding mapping based on the cosine function, which maps the original codeword distribution to the compression domain, so that the codewords are relatively uniformly distributed in the value domain; S22, then a uniformly distributed quantization level is constructed within the compressed domain, and a hard quantization operation is performed based on the nearest neighbor criterion to map continuous values to discrete quantization levels or corresponding indices.
[0010] In one embodiment of the present invention, S4 includes: S41, the discrete quantization index is converted into a finite bit representation according to the preset quantization bit width to generate a feedback bit stream for transmission on the uplink; S42 supports different quantization bit widths, different network sizes, and different hardware implementations through a hardware and software co-working CSI feedback system architecture.
[0011] To achieve the above objectives, a second aspect of the present invention provides an accelerator for channel state information feedback in edge devices, comprising: The data acquisition module acquires channel state information and encodes the channel state information using a neural network encoder to obtain continuous codeword data. The compressed domain data generation module uses a cosine function to perform compressive-expanded mapping on the continuous codeword data, mapping the data distribution to the compressed domain to obtain compressed domain data. The parallel quantization processing module performs parallel quantization processing on the compressed domain data through the quantization processing unit, and maps continuous values to discrete quantization indices based on a preset quantization level. The feedback bitstream uplink transmission module generates a feedback bitstream based on the discrete quantization index and transmits it to the receiving end via the uplink.
[0012] In one embodiment of the present invention, the data acquisition module is used for: Deploy an encoder network on the user equipment side and a decoder network on the base station side to build an end-to-end encoder-quantizer-decoder architecture; By modeling the spatial correlation and global characteristics of the channel state information using lightweight attention and feedforward structures, a continuous codeword representation suitable for subsequent codeword quantization processing is obtained. In terms of structure selection and operator design, avoid using operation forms that are highly sensitive to numerical precision, and reduce reliance on high-cost operators such as exponentiation and division.
[0013] In one embodiment of the present invention, the compressed domain data generation module is used for: First, the continuous codewords output by the encoder are subjected to companding mapping based on the cosine function, which maps the original codeword distribution to the compression domain, so that the codewords are relatively uniformly distributed in the value domain. Subsequently, a uniformly distributed quantization level is constructed within the compressed domain, and a hard quantization operation is performed based on the nearest neighbor criterion to map continuous values to discrete quantization levels or corresponding indices.
[0014] In one embodiment of the present invention, the feedback bitstream uplink transmission module is used for: The discrete quantization index is converted into a finite bit representation according to the preset quantization bit width to generate a feedback bit stream for transmission in the uplink. The CSI feedback system architecture, which combines software and hardware, supports different quantization bit widths, different network sizes, and different hardware implementations.
[0015] To achieve the above objectives, a third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.
[0016] The method, system, and storage medium of this invention optimize the CSI encoding network design stage to address the computing power and power consumption constraints of edge devices. It employs a lightweight network structure and avoids the use of exponential and complex normalization operators that demand high computational resources, enabling the encoder to operate stably even under low-precision conditions. Compared to existing solutions that rely on complex attention or high-dimensional convolutional structures, this technical solution effectively reduces the computational load and parameter scale on the UE side while ensuring CSI reconstruction performance, thereby reducing dependence on hardware performance and making it easier to deploy in actual edge devices.
[0017] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0018] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for channel state information feedback for an edge device according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the CSI feedback process based on deep learning; Figure 3 This is a schematic diagram of a channel state information feedback architecture for edge devices according to an embodiment of the present invention; Figure 4 This is a flowchart of the compassion mapping process for cosine compassion codeword quantization according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the proxy gradient for boundary alignment according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the hardware architecture of the dedicated hardware circuit module for code word quantization; Figure 7 This is a structural diagram of an accelerator for channel state information feedback of an edge device according to an embodiment of the present invention. Detailed Implementation
[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0021] The following description, with reference to the accompanying drawings, describes a channel state information feedback method and accelerator for edge devices according to an embodiment of the present invention.
[0022] Example 1 Figure 1 This is a flowchart of a channel state information feedback method according to an embodiment of the present invention.
[0023] like Figure 1 As shown, the channel state information feedback method includes the following steps: S1. Obtain channel state information and encode the channel state information using a neural network encoder to obtain continuous codeword data.
[0024] This step aims to utilize the nonlinear mapping capability of neural networks to compress high-dimensional channel state information into a low-dimensional continuous representation. Obtaining the channel state information includes obtaining a channel matrix representation in the frequency or spatial domain through a channel estimation process. The neural network encoder is deployed at the transmitter, and its network structure is configured to extract spatial correlations and global features from the channel state information, mapping the extracted features to the continuous codeword data. The continuous codeword data is an undiscretized floating-point or fixed-point tensor, serving as the input basis for subsequent quantization processing; its dimensionality is significantly lower than the original channel state information.
[0025] Specifically, such as Figure 2 The diagram shows a CSI feedback process in the prior art. As shown in the figure, the process typically includes: the encoder network on the user equipment side, the codeword quantization module, and the decoder network on the base station side.
[0026] On the user equipment side, the CSI is first input to the encoder neural network for feature compression, outputting a continuous codeword representation. This continuous codeword is mapped to a finite-bit representation via a quantization module and transmitted to the base station via the uplink feedback link. On the base station side, the received quantized codeword is dequantized and then input to the decoder neural network to reconstruct the CSI result, which is used for subsequent downlink precoding or resource scheduling. In this process, the encoder, quantizer, and decoder are typically implemented in software, with the quantization module often existing as an independent processing step outside the network.
[0027] However, most existing CSI feedback methods based on deep learning introduce complex convolutional structures or attention mechanisms, especially Transformer-based methods, which typically involve large-scale matrix multiplication, softmax exponentiation, and global normalization. These operators have high requirements for computational resources, storage bandwidth, and numerical accuracy. When the computing power, power consumption, and storage resources on the UE side are limited, it is difficult to meet real-time and energy consumption constraints, thus limiting their direct application in practical systems.
[0028] To address this technical challenge, this invention designs a CSI feedback neural network structure for edge device deployment. This structure compresses and encodes CSI on the user equipment side and decodes and reconstructs the feedback codewords on the base station side. The network employs an end-to-end architecture of encoder-quantizer-decoder, with the encoder deployed on the user equipment and the decoder on the base station side. Unlike existing schemes that only design network structures under floating-point precision conditions, this invention introduces quantization awareness from the network design stage, fully considering the impact of subsequent low-precision inference and codeword quantization on performance. Specifically, the encoder network avoids using computational forms highly sensitive to numerical precision in its structure selection and operator design, reducing reliance on high-cost operators such as exponentiation and division, enabling the network to maintain stable feature representation capabilities even under low-precision conditions such as INT8. Simultaneously, the network structure models the spatial correlation and global characteristics of CSI, reducing overall computational complexity and parameter scale while ensuring reconstruction performance through lightweight attention and feedforward structures, thereby reducing the demand on edge device computing power and storage resources. Through this design, the encoder network can output continuous codeword representations suitable for subsequent codeword quantization processing while ensuring deployability.
[0029] like Figure 3 The diagram illustrates the overall hardware architecture and data flow of the CSI feedback proposed in this invention. The system is divided into two parts: the user equipment side and the base station side, and is implemented using a hardware-software co-processing approach. On the user equipment side, the CSI data, after preprocessing, is input to a TPU-based encoder neural network acceleration unit. This acceleration unit performs feature extraction and compression of the CSI data, generating continuous codewords. The continuous codewords are then fed into a dedicated hardware circuit module for codeword quantization, where they undergo low-bit quantization and output the corresponding bitstream or quantization index. The quantized bit information is then transmitted to the base station side via the uplink. On the base station side, the received quantized bits first enter a dequantization module to recover the continuous codeword representation; subsequently, the CSI reconstruction is completed by a decoder neural network based on a TPU or high-performance computing unit. The diagram also illustrates the data interface and control relationship between the TPU acceleration unit and the codeword quantization circuit module, demonstrating the functional separation and collaborative operation of neural network computation and discrete quantization processing at the hardware level.
[0030] The aforementioned techniques effectively reduce the transmission dimensionality of channel state information while ensuring the recoverability of coded features, laying the foundation for subsequent quantization and feedback. Furthermore, the quantization-aware network design improves stability under low-precision inference, reduces the computational complexity and deployment threshold of edge devices, and achieves coordinated optimization of algorithm structure and hardware constraints.
[0031] S2, the continuous codeword data is compressed and expanded using a cosine function to map the data distribution to the compressed domain, thus obtaining compressed domain data.
[0032] Specifically, addressing the problem of mapping continuous codewords to finite bit representations during CSI feedback, this invention proposes a boundary-aligned cosine companded codeword quantization method to achieve low-bit, high-stability codeword quantization and end-to-end training.
[0033] like Figure 4 The diagram illustrates the process of mapping continuous codewords to the compressed domain via a cosine companding function. It demonstrates the monotonically nonlinear mapping between the original numerical space and the compressed domain, as well as the distribution of quantization levels within the compressed domain after compression. Assume the input codeword... It is monotonically mapped to a compression domain z:
[0034] In this method, the continuous codewords output by the encoder are first companded using a cosine function, mapping the original codeword distribution to the compressed domain. This results in a relatively uniform distribution of codewords within the value range, enhancing the resolution of the original low-amplitude region. Subsequently, uniformly distributed quantization levels are constructed within the compressed domain, and hard quantization is performed based on the nearest neighbor criterion, mapping continuous values to discrete quantization levels or corresponding indices.
[0035] To address the issue of hard quantization being non-differentiable and difficult to participate in end-to-end training, this invention further constructs a surrogate gradient function at the quantization boundary that is strictly aligned with the quantization step. For example... Figure 5 The diagram illustrates a boundary-aligned surrogate gradient. This surrogate gradient function has a non-zero value near the quantization step transition position, used to approximate the gradient information of the quantization function during backpropagation; far from the quantization boundary, the surrogate gradient value is zero to reflect the piecewise constant characteristic of hard quantization. For the quantization bit width... Construct within the compression domain z One uniform quantization level:
[0036] Obviously, It always contains 0s and 1s. Then, nearest neighbor quantization is performed in the z-domain to obtain the quantization result. The quantization calculation process is defined as follows:
[0037] The surrogate gradient function is also constructed based on the cosine function, possessing non-zero values only near the quantization boundary, and its shape remains consistent with the quantization step change. This ensures that the gradient information obtained during backpropagation accurately reflects the impact of quantization error on the model parameters. In the compressed domain z, the transition between adjacent quantization levels occurs at the midpoint:
[0038] In the formula, Indicates the first k The boundary of the step transition. Therefore, the first step is defined. k Local normalized offset of each step boundary:
[0039] In the formula, The half-width of the window at the k-th boundary. This parameter controls the relative width of the smooth transition region; as λ approaches 0, the surrogate approaches a hard step. To involve the codeword quantization process in the overall optimization, we further construct a differentiable gradient function that approximates the boundary of the quantization interval, keeping the original function as close as possible to the hard-quantized step structure. At each transition boundary... Construct a tightly supported cosine window function nearby:
[0040] Therefore, the final surrogate gradient is defined as the sum of all boundary windows:
[0041] This design maintains a high degree of consistency between forward vectorization and backward gradient propagation, effectively avoiding the training instability issues caused by traditional pass-through estimators or coarse approximations of gradients. This quantization method is not only applicable to software implementation but also provides a clear structural basis for the design of subsequent dedicated hardware quantization modules.
[0042] By employing companding based on cosine functions, the distribution characteristics of continuous codeword data can be effectively improved, making the data more suitable for uniform quantization in the compressed domain. This not only improves the quantization accuracy of signals in low-amplitude regions, avoiding the problem of insufficient resolution in low-amplitude regions caused by traditional uniform quantization, but also enhances the consistency between the quantization process and backpropagation, which is beneficial to improving the stability of end-to-end training and the final reconstruction performance, while providing clear computational logic for hardware implementation.
[0043] S3, the compressed domain data is subjected to parallel quantization processing by the quantization processing unit, and continuous values are mapped to discrete quantization indices based on a preset quantization level.
[0044] Based on a preset set of quantization levels, this unit determines the quantization interval of each continuous value using comparison logic based on the nearest neighbor criterion, thereby generating the corresponding discrete quantization index. During this process, the quantization processing unit can contain multiple parallel processing channels, each independently executing numerical comparison and decision logic to achieve high-speed quantization of vector-level data.
[0045] Specifically, at the hardware implementation level, to efficiently execute matrix multiplication and linear transformation operations in the CSI feedback neural network, this invention employs a Tensor Processing Unit (TPU) as the hardware unit for CSI feedback network computation. Through massively parallel multiply-accumulate arrays and on-chip cache, the TPU can efficiently execute matrix multiplication, feedforward neural network computation, and attention-related linear transformation operations in the encoder and decoder networks. Compared to general-purpose CPUs or GPUs, TPUs are more suitable for executing tensor computation tasks with custom data streams, providing stable and predictable inference performance under low power consumption. In the system of this invention, the TPU is responsible for handling the main computational load of the CSI encoding and decoding networks, without undertaking the discrete control logic related to codeword quantization, thereby fully leveraging its advantages in continuous tensor computation.
[0046] To address the latency bottleneck issue in practical deployments of codeword quantization, this invention further designs a method such as... Figure 6 The dedicated hardware circuit module for codeword quantization described above is used to replace the software quantization implementation on general-purpose processors. The codeword quantization circuit module includes a quantization level configuration unit, multiple parallel quantization processing units, and an output packing unit. In the quantization level configuration unit, reference values corresponding to each quantization level are pre-stored or loaded. Continuous codeword vectors are fed into multiple parallel quantization processing units through an input interface. Each processing unit performs a comparison operation between the corresponding codeword element and each quantization level, and determines the quantization index according to the minimum distance criterion. Each quantization processing unit can operate in parallel using a pipelined structure to improve overall throughput. The quantization index is integrated by the output packing unit to form a bit stream or index sequence for feedback transmission. The figure also illustrates the relationship between control signals and data paths to explain the execution flow and parallel mechanism of the quantization module at the hardware level.
[0047] This dedicated hardware module achieves high-speed nearest-neighbor quantization of input codeword vectors through parallel comparison and pipelined architecture. Internally, the module includes a quantization level configuration unit, a multi-channel parallel quantization processing unit, and a bitstream output packing unit. Each quantization processing unit can complete the quantization decision for a codeword element within a single clock cycle or a few clock cycles, with multiple processing units operating in parallel to support vector-level quantization. Because the module consists entirely of comparison, subtraction, and simple control logic, without relying on complex arithmetic operations, it is easily implemented on an Application Specific Integrated Circuit (ASIC) and can be parameterized according to quantization bit width, parallelism, and throughput requirements. By employing this dedicated module, the codeword quantization process achieves deterministic, ultra-low latency execution characteristics, thus avoiding the unstable latency caused by branching, memory access, and scheduling on CPUs or GPUs.
[0048] By employing a parallel quantization processing mechanism, this step significantly reduces the execution latency of quantization operations, avoiding the cumulative latency caused by serial processing. Simultaneously, the mapping method based on preset quantization levels ensures the determinism of the quantization process, which is beneficial for meeting the low latency and stability requirements of real-time communication scenarios.
[0049] S4, generate a feedback bit stream based on the discrete quantization index, and send it to the receiving end via the uplink. The process of generating the feedback bitstream in this invention includes mapping a discrete quantization index into a binary bit sequence according to a preset encoding protocol. This mapping process ensures that the index information is encapsulated into a data format that meets the transmission requirements of the communication link, enabling efficient transmission with limited bandwidth resources. During this process, the frame structure and timing configuration of the data transmission are adapted to optimize spectral efficiency and transmission reliability. Subsequently, the feedback bitstream is sent to the receiving end via the uplink, enabling the receiving end to recover channel state information based on the received bitstream.
[0050] At the system level, this invention organically combines the aforementioned algorithm with hardware modules to construct a hardware-software co-designed CSI feedback system architecture. In this architecture, neural network computation and codeword quantization are explicitly divided into two functional subsystems: continuous tensor computation is accelerated by a TPU, while discrete quantization and bitstream mapping are performed by a dedicated quantization circuit module. The two types of modules interact via an AXI (Advanced eXtensible Interface) bus interface, enabling the system to be highly integrated on a single-chip platform or distributed on heterogeneous platforms. This architecture supports different quantization bit widths, different network scales, and different hardware implementations, exhibiting good scalability and portability. Through hardware-software co-design, this invention achieves low-complexity, low-latency, and low-power edge CSI feedback while ensuring CSI reconstruction performance, providing a feasible technical solution for practical deployment in next-generation wireless communication systems.
[0051] This technical step achieves efficient feedback of channel state information by converting discrete quantization indexes into a feedback bitstream and transmitting it via the uplink. This process ensures that quantized information can be transmitted in a compact format within limited bandwidth, effectively reducing uplink load. Simultaneously, combined with pre-quantization processing, it guarantees the accuracy and timeliness of the receiver's reconstruction of channel state information, meeting the low-latency communication requirements of edge devices.
[0052] Example 2 like Figure 7 As shown, the present invention proposes an accelerator 10 for channel state information feedback of edge devices, comprising: The data acquisition module 100 acquires channel state information and encodes the channel state information through a neural network encoder to obtain continuous codeword data. The compressed domain data generation module 200 uses a cosine function to perform compressive-expanded mapping on the continuous codeword data, mapping the data distribution to the compressed domain to obtain compressed domain data. The parallel quantization processing module 300 performs parallel quantization processing on the compressed domain data through the quantization processing unit, and maps continuous values to discrete quantization indices based on a preset quantization level. The feedback bit stream uplink transmission module 400 generates a feedback bit stream based on the discrete quantization index and transmits it to the receiving end via the uplink.
[0053] Furthermore, the data acquisition module is used for 100: Deploy an encoder network on the user equipment side and a decoder network on the base station side to build an end-to-end encoder-quantizer-decoder architecture; By modeling the spatial correlation and global characteristics of the channel state information using lightweight attention and feedforward structures, a continuous codeword representation suitable for subsequent codeword quantization processing is obtained. In terms of structure selection and operator design, avoid using operation forms that are highly sensitive to numerical precision, and reduce reliance on high-cost operators such as exponentiation and division.
[0054] Furthermore, the compressed domain data generation module 200 is used for: First, the continuous codewords output by the encoder are subjected to companding mapping based on the cosine function, which maps the original codeword distribution to the compression domain, so that the codewords are relatively uniformly distributed in the value domain. Subsequently, a uniformly distributed quantization level is constructed within the compressed domain, and a hard quantization operation is performed based on the nearest neighbor criterion to map continuous values to discrete quantization levels or corresponding indices.
[0055] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for channel state information feedback for edge devices.
[0056] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0057] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.
Claims
1. A method for channel state information feedback for edge devices, characterized in that, Includes the following steps: S1, acquire channel state information, and encode the channel state information through a neural network encoder to obtain continuous codeword data; S2, The continuous codeword data is compressed and expanded using the cosine function to map the data distribution to the compressed domain, thus obtaining compressed domain data; S3, the compressed domain data is subjected to parallel quantization processing by the quantization processing unit, and continuous values are mapped to discrete quantization indices based on a preset quantization level; S4. Generate a feedback bit stream based on the discrete quantization index and send it to the receiving end via the uplink.
2. The method according to claim 1, characterized in that, S1 includes: S11, deploy an encoder network on the user equipment side and a decoder network on the base station side to build an end-to-end architecture of encoder-quantizer-decoder; S12, the spatial correlation and global characteristics of the channel state information are modeled by lightweight attention and feedforward structure to obtain a continuous codeword representation suitable for subsequent codeword quantization processing; S13 avoids using computational forms that are highly sensitive to numerical precision in its structural selection and operator design, and reduces its reliance on high-cost operators such as exponentiation and division.
3. The method according to claim 2, characterized in that, S13 includes: S131 adopts the concept of quantization perception to fully consider the subsequent low-precision weight quantization deployment requirements in the network design stage, so that the network can still maintain stable feature expression ability under low-precision conditions. S132 employs a unified quantization-aware training strategy that integrates network weight quantization and codeword quantization, preventing errors from accumulating continuously during multi-level quantization.
4. The method according to claim 1, characterized in that, S2 includes: S21, First, the continuous codewords output by the encoder are subjected to companding mapping based on the cosine function, which maps the original codeword distribution to the compression domain, so that the codewords are relatively uniformly distributed in the value domain; S22, then a uniformly distributed quantization level is constructed within the compressed domain, and a hard quantization operation is performed based on the nearest neighbor criterion to map continuous values to discrete quantization levels or corresponding indices.
5. The method according to claim 1, characterized in that, S4 includes: S41, the discrete quantization index is converted into a finite bit representation according to the preset quantization bit width to generate a feedback bit stream for transmission on the uplink; S42 supports different quantization bit widths, different network sizes, and different hardware implementations through a hardware and software co-working CSI feedback system architecture.
6. An accelerator for channel state information feedback in edge devices, characterized in that, include: The data acquisition module acquires channel state information and encodes the channel state information using a neural network encoder to obtain continuous codeword data. The compressed domain data generation module uses a cosine function to perform compressive-expanded mapping on the continuous codeword data, mapping the data distribution to the compressed domain to obtain compressed domain data. The parallel quantization processing module performs parallel quantization processing on the compressed domain data through the quantization processing unit, and maps continuous values to discrete quantization indices based on a preset quantization level. The feedback bitstream uplink transmission module generates a feedback bitstream based on the discrete quantization index and transmits it to the receiving end via the uplink.
7. The accelerator according to claim 6, characterized in that, The data acquisition module is used for: Deploy an encoder network on the user equipment side and a decoder network on the base station side to build an end-to-end encoder-quantizer-decoder architecture; By modeling the spatial correlation and global characteristics of the channel state information using lightweight attention and feedforward structures, a continuous codeword representation suitable for subsequent codeword quantization processing is obtained. In terms of structure selection and operator design, avoid using operation forms that are highly sensitive to numerical precision, and reduce reliance on high-cost operators such as exponentiation and division.
8. The accelerator according to claim 6, characterized in that, The compressed domain data generation module is used for: First, the continuous codewords output by the encoder are subjected to companding mapping based on the cosine function, which maps the original codeword distribution to the compression domain, so that the codewords are relatively uniformly distributed in the value domain. Subsequently, a uniformly distributed quantization level is constructed within the compressed domain, and a hard quantization operation is performed based on the nearest neighbor criterion to map continuous values to discrete quantization levels or corresponding indices.
9. The accelerator according to claim 6, characterized in that, The feedback bitstream uplink transmission module is used for: The discrete quantization index is converted into a finite bit representation according to the preset quantization bit width to generate a feedback bit stream for transmission in the uplink. The CSI feedback system architecture, which combines software and hardware, supports different quantization bit widths, different network sizes, and different hardware implementations.
10. A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method as claimed in any one of claims 1-5.