Non-cellular distributed uplink receiving method and system based on diffusion model
The cellular-free distributed uplink reception method based on the diffusion model solves the problem of limited fronthaul link capacity in cellular-free networks, achieves high-performance distributed signal processing, removes accumulated noise interference, and improves the equalization effect.
Patent Information
- Application Number
- CN202510782442.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-12
- Publication Date
- 2025-09-12
AI Technical Summary
In non-cellular wireless access networks, the fronthaul link capacity between distributed devices is limited, resulting in suboptimal information transmission and cumulative noise interference affecting the equalization effect.
The cell-free distributed uplink reception method adopts a diffusion model, which removes accumulated noise and achieves optimal equalization through local equalization and signal combining at the RF access point layer, information aggregation at the fronthaul layer, and denoising processing at the central processing unit layer.
A high-performance, massively scalable cellular-free system is realized, effectively eliminating noise interference in the fronthaul link and improving the effect of distributed signal processing.
Smart Images
Figure CN120640352A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of communication technology, and in particular to a cellular-free distributed uplink receiving method and system based on a diffusion model. Background Art
[0002] In non-cellular wireless access networks, distributed uplink equalization can achieve scalable non-cellular uplink reception. However, in practical systems, to improve the performance of distributed signal processing and fully leverage the advantages of non-cellular networks, effective collaboration between distributed devices is often required. A typical approach is for distributed devices to first perform local equalization, then report limited information upward. A central processor aggregates this distributed information for further equalization optimization. However, due to cost constraints, fronthaul capacity may not provide lossless information transmission. Distributed devices accumulate noise when reporting combined information, ultimately causing significant noise interference to the aggregated information and affecting the final equalization effect. Summary of the Invention
[0003] Purpose of the invention: The present invention provides a method and system for cellular-free distributed uplink reception based on a diffusion model, which can solve the problems of capacity limitation of the fronthaul link and unsatisfactory information transmission in cellular-free distributed uplink.
[0004] Technical solution: The present invention provides a non-cellular distributed uplink reception method based on a diffusion model, comprising the following steps:
[0005] Step 1: The wireless access point (AP) in the radio access point layer performs user selection, local channel estimation, and preliminary equalization, and uploads uplink combining information for assisting in combining signals.
[0006] Step 2: The forwarding layer collects the AP's uplink combined information and transmits it to the next level until it is aggregated to the central processing unit layer;
[0007] Step 3: The central processing unit layer aggregates the uplink combined information, performs denoising using a diffusion model, and re-equalizes the uplink signal using the denoised information.
[0008] Furthermore, in step 1, the AP performs user selection, local channel estimation, and preliminary equalization, and uploads uplink combining information for assisting in combining signals, specifically including the following steps:
[0009] Step 11: The AP performs user selection and local channel estimation, obtains the channel estimation from the AP to the selected user, the related estimation error, and the receiving noise variance parameter;
[0010] Step 12: AP performs local balancing. Local balancing does not destroy the global optimal balance. In theory, the optimal balancing effect can be achieved by further processing by the central processor. The local balancing matrix used is expressed as
[0011]
[0012] in is the channel estimation, C lk is the estimation error matrix, p k is the user power, is the received noise variance;
[0013] Step 13: AP calculates uplink combined information, which is expressed as
[0014]
[0015] That is, equivalent channel.
[0016] Furthermore, in step 2, the forwarding layer aggregates the AP's uplink combined information and transmits it to the next level until it is aggregated to the central processing unit layer. Specifically, the steps include:
[0017] Step 21: The fronthaul layer aggregates the AP's uplink combined information. Each node performs bitwise addition on the received noisy information. AP1 and AP2 are connected in series. AP1 uploads the noisy uplink combined information T1+N1 to AP2. AP2 adds its own noisy uplink combined information T2+N2 to the combined information and then uploads the noisy uplink combined information T1+T2+N1+N2 to the upper level.
[0018] Step 22: The forwarding layer further passes the information upward to the CPU layer.
[0019] Furthermore, in step 3, the central processing unit layer aggregates the uplink combined information, performs denoising processing using a diffusion model, and uses the denoised information to re-equalize the uplink signal. Specifically, the steps include:
[0020] Step 31: The CPU layer performs a final accumulation to obtain the final noisy uplink combined information T+N. In large-scale non-cellular networks, according to the law of large numbers, N eventually tends to a complex Gaussian distribution. This conclusion applies to any distribution law of the noise introduced at each step in the uplink combining process.
[0021] Step 32: The CPU layer uses the diffusion model to perform denoising. The CPU layer regards the process of gradually accumulating noise in the upstream merge as a Brownian motion process. This is a special case called a Brownian bridge. Because the final state of the contaminated information is known, a diffusion model can be used to describe the denoising and denoising process of the upstream merge information T. Then the forward noise addition process of the diffusion model can be described as
[0022] q(x m |x m-1 ,x M )
[0023] Such a conditional distribution, and reverse denoising uses
[0024] p θ (x m-1 |x m )
[0025] To express it, where θ is the deep learning network parameter to be learned.
[0026] Step 33: The CPU layer uses the denoised information Re-equalize the uplink signal:
[0027]
[0028] in is the local signal estimate.
[0029] Accordingly, a cellular-free distributed uplink receiving system based on a diffusion model includes: a radio frequency access point layer, a fronthaul layer, and a central processing unit layer; the AP in the radio frequency access point layer completes beamforming, transmission and reception of wireless radio frequency signals in each frequency band, digital-to-analog and analog-to-digital conversion, and waveform signal processing of the physical channel; the fronthaul layer is composed of switches connected according to certain topologies and edge computing units with certain computing capabilities, which are responsible for forwarding information and simple calculation and merging; the central processing unit layer is composed of a CPU cluster, which performs final calculations and information merging, and the diffusion model is deployed in the central processing unit layer with a graphics processing unit.
[0030] Furthermore, the AP performs user selection, local channel estimation, and preliminary equalization, and uploads uplink combining information used to assist in combining signals. The AP performs user selection and local channel estimation, obtains the channel estimate from the AP to the selected user and the related estimation error, and receives the noise variance parameter. The AP performs local equalization, which does not destroy the global optimal equalization. In theory, further processing by the central processor can ultimately achieve the optimal equalization effect.
[0031] Furthermore, the fronthaul layer collects the uplink combined information of the AP and transmits it to the next level until it is aggregated to the central processing unit layer.
[0032] Furthermore, the fronthaul layer aggregates the uplink combined information of the APs, and each node adds the received noisy information bit by bit. AP1 and AP2 are connected in series. AP1 uploads the noisy uplink combined information to AP2, and AP2 adds its own noisy uplink combined information to it before uploading the noisy uplink combined information to the upper level.
[0033] Furthermore, the central processing unit layer aggregates the uplink combined information, performs denoising using a diffusion model, and uses the denoised information to re-equalize the uplink signal.
[0034] Furthermore, the CPU layer performs a final accumulation to obtain the final noisy uplink merge information T+N. In large-scale non-cellular networks, the law of large numbers states that noise ultimately tends to a complex Gaussian distribution. This conclusion applies to any distribution law of noise introduced at each step in the uplink merging process. The CPU layer uses a diffusion model for denoising. The CPU layer regards the process of gradually accumulating noise in the uplink merging as a Brownian motion process, which is a special case called a Brownian bridge. Because the final state of the contaminated information is known, a diffusion model can be used to describe the noise addition and denoising process of the uplink merge information.
[0035] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: distributed uplink reception takes into account the propagation noise and quantization noise of the fronthaul link. Regardless of the topology structure adopted by the cellular-free fronthaul link, the uplink process can be equivalent to the diffusion process described in the present invention. The central processor removes noise and merges uplink information through the deployed diffusion model, thereby realizing a high-performance, large-scale, scalable cellular-free system with limited fronthaul link capacity. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 Schematic diagram of the method of the present invention.
[0037] Figure 2 The figure is a flow chart of the present invention for performing uplink combined information denoising using a diffusion model.
[0038] Figure 3 This is a trend comparison chart of the NMSE of the uplink combined information versus the number of diffusion model training rounds in the present invention.
[0039] Figure 4 This is a trend comparison chart of the average uplink per-user spectral efficiency versus the number of diffusion model training rounds in the present invention. DETAILED DESCRIPTION
[0040] like Figure 1 and Figure 2 As shown, a non-cellular distributed uplink receiving method based on a diffusion model includes the following steps:
[0041] Step S1: The AP performs user selection, local channel estimation, and preliminary equalization, and uploads uplink combining information for assisting in combining signals.
[0042] The specific steps include:
[0043] Step S101: The AP performs user selection and local channel estimation to obtain parameters such as channel estimation from the AP to the selected user, related estimation error, and reception noise variance.
[0044] Step S102: AP performs local balancing. Local balancing does not destroy the global optimal balance. In theory, the optimal balancing effect can be achieved by further processing by the central processing unit layer. The local balancing matrix used can be expressed as
[0045]
[0046] in is the channel estimation, C lk is the estimation error matrix, p k is the user power, is the received noise variance.
[0047] Step S103: The AP calculates uplink merging information. Specifically, the uplink merging information can be expressed as
[0048]
[0049] The AP uploads this information to the fronthaul network. Due to the constraints of the fronthaul capacity, the uplink combined information cannot be perfectly uploaded to the central processing unit layer. It usually carries a noise, and the noise accumulates as the propagation path increases. Such noise can be expressed as N l , and T l The dimensions are exactly the same and N l The elements are independent of each other.
[0050] Step S2: The fronthaul layer collects the uplink combined information of the AP and transmits it to the upper level until it is aggregated to the central processing unit layer.
[0051] The specific steps include:
[0052] Step S201: The fronthaul layer aggregates the AP's uplink combined information, and each node performs bitwise addition on the received noisy information. Optionally, AP1 and AP2 are connected in series. AP1 uploads the noisy uplink combined information T1+N1 to AP2, and AP2 adds its own noisy uplink combined information T2+N2 to it before uploading the noisy uplink combined information T1+T2+N1+N2 to the upper level.
[0053] Step S202: The forwarding layer further passes the information upward to the CPU layer.
[0054] Step S3: The central processing unit layer aggregates the uplink combined information, performs denoising processing using a diffusion model, and re-equalizes the uplink signal using the denoised information.
[0055] The specific steps include:
[0056] In step S301, the CPU layer performs a final accumulation to obtain the final noisy uplink combined information T+N. In particular, in large-scale non-cellular networks, the law of large numbers states that N eventually tends to a complex Gaussian distribution. This conclusion applies to any distribution law of the noise introduced at each step in the uplink combining process.
[0057] Step S302: The CPU layer performs denoising using a diffusion model. Specifically, the CPU considers the process of gradually accumulating noise in the uplink merge as a Brownian motion process. This is a special case called a Brownian bridge because the final state of the contaminated information is known. A diffusion model can be used to describe the denoising and denoising process of the uplink merged information T. Then the forward noise addition process of the diffusion model can be described as
[0058] q(x m |x m-1 ,x M )
[0059] Such a conditional distribution, and reverse denoising can be used
[0060] p θ (x m-1 |x m )
[0061] To express it, where θ is the deep learning network parameter to be learned.
[0062] Step S303: The CPU layer uses the denoised information Re-equalize the uplink signal. In particular, the following formula can be used for processing:
[0063]
[0064] in is the local signal estimate.
[0065] Through the above three steps, the non-cellular distributed uplink reception based on the diffusion model is completed. The effect example can be seen Figure 3 and Figure 4 . Figure 3The NMSE of the uplink merged information is plotted as the number of diffusion model training rounds increases. A comparison is made with the classic denoising deep learning network DnCNN. It is clear that DnCNN does not capture the characteristics of the uplink merged information well, with little improvement as the number of training rounds increases. Figure 4 This is a trend chart of the average uplink per-user spectral efficiency versus the number of diffusion model training rounds, comparing the classic denoising deep learning network DnCNN and the distributed uplink reception method LTMMSE. LTMMSE can be viewed as the upper limit that can be achieved without exchanging instantaneous auxiliary information between distributed processors. Clearly, the method described in this invention is far superior to the distributed non-collaborative method.
[0066] Accordingly, a non-cellular distributed uplink receiving system based on a diffusion model includes: a radio frequency access point layer (Access Points Layer, APL), a fronthaul layer (Fronthaul Layer, FL) and a central processing unit layer (Central Processing Unit Layer, CPUL);
[0067] The AP layer of the radio access point mainly performs beamforming, transmission and reception of wireless radio frequency signals in various frequency bands, digital-to-analog and analog-to-digital conversion, and waveform signal processing of the physical channel.
[0068] The fronthaul layer consists of switches connected according to certain topologies and edge computing units (EDUs) with certain computing capabilities. It is mainly responsible for information forwarding and simple calculation merging.
[0069] The central processing unit layer consists of a CPU cluster that performs final calculations and information merging;
[0070] The diffusion model is deployed at the central processing unit layer with a graphics processing unit (GPU). The entire method is jointly implemented by various units in the cellular-free network. Specifically, the computing processing resources on the AP, EDU and CPU run a software program that implements the cellular-free distributed uplink reception method of the diffusion model.
[0071] The AP performs user selection, local channel estimation, and preliminary equalization, and uploads uplink combining information used to assist in combining signals. The AP performs user selection and local channel estimation, obtains the channel estimate from the AP to the selected user and the related estimation error, and receives the noise variance parameter. The AP performs local equalization, which does not destroy the global optimal equalization. In theory, further processing by the central processor can ultimately achieve the optimal equalization effect.
[0072] The fronthaul layer collects the uplink combined information of the AP and transmits it to the next level until it is aggregated to the central processing unit layer.
[0073] The fronthaul layer aggregates the uplink combined information of the APs. Each node adds the received noisy information bit by bit. AP1 and AP2 are connected in series. AP1 uploads the noisy uplink combined information to AP2. AP2 adds its own noisy uplink combined information to the information and then uploads the noisy uplink combined information to the upper level.
[0074] The central processing unit layer aggregates the uplink combined information, performs denoising using a diffusion model, and uses the denoised information to re-equalize the uplink signal.
[0075] The CPU layer performs a final accumulation to obtain the final noisy uplink merge information T+N. In large-scale non-cellular networks, the law of large numbers indicates that noise ultimately tends to a complex Gaussian distribution. This conclusion applies to any distribution law of noise introduced at each step in the uplink merging process. The CPU layer uses a diffusion model for denoising. The CPU layer regards the process of gradually accumulating noise in the uplink merging as a Brownian motion process, which is a special case called a Brownian bridge. Because the final state of the contaminated information is known, a diffusion model can be used to describe the noise addition and denoising process of the uplink merge information.
[0076] The distributed uplink reception of this invention takes into account the propagation noise and quantization noise of the fronthaul link. Regardless of the topology of the cellular-free fronthaul link, the uplink process is equivalent to the diffusion process described in this invention. The central processor then uses the deployed diffusion model to remove noise and merge uplink information, thereby achieving a high-performance, massively scalable cellular-free system with limited fronthaul link capacity.
Claims
1. A cell-free distributed uplink reception method based on a diffusion model, characterized in that: The steps include: Step 1: The wireless access point (AP) in the radio access point layer performs user selection, local channel estimation, and preliminary equalization, and uploads uplink combining information for assisting in combining signals. Step 2: The forwarding layer collects the AP's uplink combined information and transmits it to the next level until it is aggregated to the central processing unit layer; Step 3: The central processing unit layer aggregates the uplink combined information, performs denoising using a diffusion model, and re-equalizes the uplink signal using the denoised information.
2. The method for non-cellular distributed uplink reception based on a diffusion model according to claim 1, wherein: In step 1, the AP performs user selection, local channel estimation, and preliminary equalization, and uploads uplink combining information for assisting in combining signals. Specifically, the steps include: Step 11: The AP performs user selection and local channel estimation, obtains the channel estimation from the AP to the selected user, the related estimation error, and the receiving noise variance parameter; Step 12: AP performs local balancing. The local balancing matrix is expressed as in is the channel estimation, C lk is the estimation error matrix, p k is the user power, is the received noise variance; Step 13: AP calculates uplink combined information, which is expressed as That is, equivalent channel.
3. The method for non-cellular distributed uplink reception based on a diffusion model according to claim 1, wherein: In step 2, the forwarding layer aggregates the AP's uplink combined information and transmits it to the next level until it is aggregated to the central processing unit layer. The specific steps include: Step 21: The fronthaul layer aggregates the AP's uplink combined information. Each node performs bitwise addition on the received noisy information. AP1 and AP2 are connected in series. AP1 uploads the noisy uplink combined information T1+N1 to AP2. AP2 adds its own noisy uplink combined information T2+N2 to the combined information and then uploads the noisy uplink combined information T1+T2+N1+N2 to the upper level. Step 22: The forwarding layer further passes the information upward to the CPU layer.
4. The method for non-cellular distributed uplink reception based on a diffusion model according to claim 1, wherein: In step 3, the central processing unit layer aggregates the uplink combined information, performs denoising using a diffusion model, and uses the denoised information to rebalance the uplink signal. Specifically, the steps include: Step 31: The central processing unit layer performs a final accumulation to obtain the final noisy uplink combined information T+N; Step 32: The CPU layer uses a diffusion model to perform denoising. The CPU layer regards the process of gradually accumulating noise in uplink merging as a Brownian motion process and uses a diffusion model to describe the denoising and denoising process of the uplink merging information T. Then the forward noise addition process of the diffusion model is described as q(x m |x m-1 ,x M ) Such a conditional distribution, and reverse denoising uses p θ (x m-1 |x m ) To express it, where θ is the deep learning network parameter to be learned. Step 33: The CPU layer uses the denoised information Re-equalize the uplink signal: in is the local signal estimate.
5. A system for implementing the cell-free distributed uplink reception method based on a diffusion model as claimed in claim 1, characterized in that: include: Radio access point layer, fronthaul layer, and central processing unit layer; The AP in the radio access point layer performs beamforming, transmission and reception of radio frequency signals in various frequency bands, digital-to-analog and analog-to-digital conversion, and waveform signal processing of the physical channel. The fronthaul layer consists of switches connected according to certain topologies and edge computing units with certain computing capabilities, which are responsible for forwarding information and simple calculation merging; the central processing unit layer consists of a CPU cluster, which performs final calculations and information merging. The diffusion model is deployed in the central processing unit layer with a graphics processing unit.
6. The non-cellular distributed uplink receiving system based on the diffusion model according to claim 5, characterized in that: The AP performs user selection, local channel estimation, and preliminary equalization, and uploads uplink combining information to assist in combining signals. The AP performs user selection and local channel estimation, obtains the channel estimate from the AP to the selected user, the associated estimation error, and the receive noise variance parameter. The AP performs local balancing, and the central processor ultimately processes it further to achieve the optimal balancing effect.
7. The non-cellular distributed uplink receiving system based on the diffusion model according to claim 5, characterized in that: The fronthaul layer collects the uplink combined information of the AP and transmits it to the next level until it is aggregated to the central processing unit layer.
8. The non-cellular distributed uplink receiving system based on the diffusion model according to claim 5, characterized in that: The fronthaul layer aggregates the uplink combined information of the APs. Each node adds the received noisy information bit by bit. AP1 and AP2 are connected in series. AP1 uploads the noisy uplink combined information to AP2. AP2 adds its own noisy uplink combined information to the information and then uploads the noisy uplink combined information to the upper level.
9. The non-cellular distributed uplink receiving system based on the diffusion model according to claim 5, characterized in that: The central processing unit layer aggregates the uplink combined information, performs denoising using a diffusion model, and uses the denoised information to re-equalize the uplink signal.
10. The cell-free distributed uplink receiving system based on a diffusion model according to claim 5, wherein: The CPU layer performs a final accumulation to obtain the final noisy uplink merge information T+N. The CPU layer uses a diffusion model for denoising. The CPU layer regards the process of gradually accumulating noise in uplink merging as a Brownian motion process and uses a diffusion model to describe the denoising and denoising process of the uplink merge information.