Superimposed channel state information feedback method with differential modulation assisted parallel branch fusion

Abstract

The application discloses a superimposed channel state information feedback method of differential modulation auxiliary parallel branch fusion, comprising: a user end: the user end carries out real value, quantization and superposition on a downlink CSI vector, maps and differentially modulates the UL-US to obtain a differential modulated signal and feeds back to a base station. A base station end: the received signal is processed in parallel according to a traditional branch and a learning branch, and then fused; finally, the recovery information of the downlink CSI and the initial characteristics of the CSI are sent into a fusion network Fushion_Net, so that the recovery precision of the downlink CSI fused with the signal characteristics of two paths is enhanced. The application avoids the user end from sending a guide, thereby saving bandwidth resources, reducing the energy consumption of the user end, having generalization, and improving the reconstruction precision of the downlink CSI.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning superimposed channel state information feedback, and particularly relates to a differential modulation assisted parallel branch fusion superimposed channel state information (CSI) feedback method. BACKGROUND

[0002] To reduce the feedback overhead and thus reduce the uplink bandwidth resource occupation and user equipment (UE) energy consumption, the existing methods mainly focus on the compressed feedback of downlink CSI. The CSI feedback method based on compressed sensing (CS) reduces the feedback overhead by developing the sparse structure of the signal. However, the CSI feedback based on CS is based on the assumption of sparse downlink CSI under a certain sparse basis, which sometimes does not match the actual situation, causing a significant decrease in the performance of the sparse reconstruction of the downlink CSI. The application of deep learning (DL) to CSI feedback not only effectively reduces the CSI feedback overhead but also improves the feedback accuracy of the CSI. The superimposed CSI feedback method based on DL superimposes the downlink CSI on the uplink user data sequence (UL-US) and feeds it back to the base station (BS) end, avoiding the occupation of additional uplink bandwidth resources due to the feedback of the downlink CSI and improving the spectral efficiency. Furthermore, in order for the BS end to demodulate the UL-US, the UE usually needs to send a pilot for uplink channel estimation to the BS, inevitably occupying the uplink bandwidth resources and consuming the energy of the UE. Fortunately, the differential modulation assisted parallel branch fusion superimposed channel state information feedback method avoids the feedback overhead and the pilot overhead for uplink channel estimation through differential methods and superimposition technology, thereby reducing the occupation of uplink bandwidth resources and saving the energy consumption of the UE end. However, differential detection causes a decrease in detection performance, and superimposed signal processing causes superimposed interference. To solve this problem, the present application adopts a parallel architecture of a traditional processing branch and a learning processing branch (based on a DL mode), reduces the superimposed interference, and makes up for the performance loss caused by differential detection. Unlike the DL processing mode, the present application utilizes the traditional processing branch to alleviate the generalization difficulty problem of the DL processing with the help of the parallel architecture; at the same time, the learning processing branch compensates for the deficiencies of the traditional processing branch in suppressing superimposed interference and differential detection. Therefore, the method of the present application fuses the traditional processing branch and the learning processing branch (based on the DL mode), improves the reconstruction accuracy of the downlink CSI, improves the detection performance of the UL-US, and has the DL generalization at the same time. SUMMARY

[0003] The application aims to provide a differential modulation auxiliary parallel branch fusion superimposed channel state information feedback method, which can improve the reconstruction accuracy of downlink CSI while saving uplink bandwidth resources and reducing user end energy consumption compared with the existing uplink pilot channel estimation CSI feedback.

[0004] The differential modulation auxiliary parallel branch fusion superimposed channel state information feedback method comprises the following steps:

[0005] User end processing:

[0006] S1. The modulated UL-US d is obtained by modulating the downlink CSI vector with real value and quantized spread bit stream form CSI w∈{0,1} P×1 After superimposition and mapping processing and differential modulation in turn, the differential modulated signal is obtained and fed back to the base station.

[0007] The real value is to convert the downlink CSI vector to real value CSI information

[0008] The quantization is a-bit quantization of to obtain the bit stream information m∈{0,1} 2aN×1 of real and imaginary parts.

[0009] The spread refers to spreading the spreading matrix with m through the formula w=Qm to obtain the spreading signal w with the same length as d.

[0010] The differential modulation is differential modulation of s=[s1,s2,...,s p ] T According to the rule differential modulation is performed, and x0=1 is discarded to obtain the differential modulation signal

[0011] The mapping processing further comprises the following sub-steps:

[0012] S11. According to the modulation order γ1 adopted by the element digital modulation in UL-US d, the modulation order γ2 of the element mapping symbol in the superimposed signal (d+w) is obtained: γ2=γ1+1.

[0013] S12. Each element in the superimposed signal (d+w) is mapped to a γ2-order modulation symbol, forming a digital modulation symbol with a modulation order 1 higher than that of d

[0014] Base station end processing:

[0015] The base station end processes the received signal in parallel with the traditional processing branch and the learning processing branch, and then fuses them;

[0016] S2. Traditional processing branch:

[0017] S21, the base station end processes the received signal to recover the bit stream form CSI by character by character differential demodulation and inverse mapping and the modulated UL-US

[0018] S22, the received signal is processed by despreading and dequantization to obtain the downlink CSI recovery information

[0019] The received signal r is obtained by r=x⊙h UL +n, where represents the equivalent uplink channel, represents additive noise;

[0020] The character by character differential demodulation is obtained by solving to obtain the differential demodulation signal where, represents the i-th decision value; r i , r i-1 represent the i-th and i-1-th elements in r respectively; s t represents the attempt constellation point value in the (γ+1)-order digital modulation constellation set , and the superscript * represents the conjugate operation, represents the s t selected from the set to maximize {·}, and Re[·] represents the real part;

[0021] S3. Learning processing branch: input the received signal into the learning branch network Infeature_Net to obtain the initial CSI feature

[0022] Step S3 includes the following sub-steps:

[0023] S31. Constructing a data set {r, h label} to train the learning branch network Infeature_Net to obtain the network parameters of the learning branch network Infeature_Net;

[0024] S32. The training label is obtained by real-valuing the downlink CSI vector h, that is:

[0025] h label= [Re (h1), Im (h1),..., Re (hN), Im (hN)] N N T

[0026] S33. In online operation, input r into the learning branch network Infeature_Net to obtain the initial features of CSI The specific process can be represented as:

[0027]

[0028] wherein f Infe (·) represents the network learning operation, and Θ Infe represents the network parameters of Enh-CsiNet.

[0029] S4. Send and to the fusion network Fushion_Net together to obtain the recovery accuracy enhanced downlink CSI fused with the features of the two signals

[0030] The superimposed channel state information feedback method of differential modulation assisted parallel branch fusion includes the following sub-steps in step S4.

[0031] S41. The despreading processing refers to using the same spreading matrix as the user end to obtain the despreading CSI in the form of a bit stream through the formula

[0032] S42. The dequantization processing refers to performing a-bit dequantization on to obtain the initial features of CSI

[0033] S43. The fusion network Fushion_Net is obtained by training the network parameters of Fushion_Net through constructing the data set {h T ,h label}, wherein h T is a group of training data in the offline training of Fushion_Net, and in online operation, input h T into the fusion network Fushion_Net to obtain the final downlink CSI The specific process can be represented as:

[0034] ​​​​

[0035] wherein, f Fus (·) represents the network learning operation, Θ Fus represents the network parameters of Fushion_Net.

[0036] The application utilizes differential modulation, superposition technology for CSI feedback, and then improves the reconstruction accuracy of downlink CSI through the lightweight feature extraction parallel network Infeature_Net and the fusion network Fushion_Net, and combines the advantages of quantization coding, digital modulation, mapping phase assignment and DL technology, feeds back the differential modulation signal after differential modulation to the base station through OFDM modulation technology, uses differential demodulation, inverse mapping and digital demodulation and dequantization processing at the base station end to obtain downlink CSI recovery information, at the same time, directly extracts downlink CSI from the received signal by using the feature extraction parallel network Infeature_Net, and finally uses the fusion network Fushion_Net to fuse the final reconstructed CSI according to the downlink CSI recovery information and the downlink CSI obtained from the feature extraction parallel network Infeature_Net, which greatly saves the uplink bandwidth resources and UE end energy while avoiding the UE end to send guide for BS end channel estimation and completely avoiding feedback overhead, and ensures the reconstruction accuracy of CSI.

[0037] Compared with the existing uplink guide channel estimation CSI feedback, the application saves the uplink bandwidth resources and reduces the user end energy consumption by using differential encoding technology, and improves the reconstruction accuracy of downlink CSI by using lightweight neural network. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is the overall flowchart of the application;

[0039] Figure 2 is the user end flowchart design of the application;

[0040] Figure 3 is the base station end flowchart design of the application. DETAILED DESCRIPTION

[0041] The application will be described in detail below in combination with embodiments and drawings, but it should be understood that the embodiments and drawings are only used to exemplarily describe the application, and cannot constitute any limitation on the protection scope of the application. All reasonable transformations and combinations within the scope of the inventive concept of the application fall within the protection scope of the application.

[0042] The differential modulation auxiliary parallel branch fusion superposition channel state information feedback method is shown in the flowchart as Figure 1 , which includes:

[0043] , which includes: Figure 2The user terminal processing is shown as follows:

[0044] S1. The modulated UL-US is superimposed with the quantized spread spectrum bit stream form CSI w∈{0, 1} P×1 and the mapping processing and differential modulation are sequentially performed to obtain the differential modulated signal and fed back to the base station.

[0045] The base station processing is shown as follows: Figure 3

[0046] The base station parallel processes the received signal according to the traditional processing branch and the learning processing branch and then fuses them.

[0047] S2. The traditional processing branch: the base station processes the received signal to recover the modulated UL-US and the downlink CSI recovery information through character-by-character differential demodulation, inverse mapping, despreading and dequantization processing.

[0048] S3. The learning processing branch: the received signal is input into the learning branch network Infeature_Net to obtain the CSI initial feature.

[0049] S4. The and the are sent into the fusion network Fushion_Net to obtain the downlink CSI with enhanced recovery accuracy fused with the features of the two signals.

[0050] The specific process of the embodiment is as follows:

[0051] a1. A specific implementation of mapping the superimposed signal (d+w) into a digital modulation symbol with modulation order one higher than d is as follows: taking the QPSK (modulation order is 2) modulation of the UL-US as an example, assuming that the modulated UL-US has a length of 8 and d is:

[0052]

[0053] The quantized spread spectrum bit stream form CSI w=[1, 1, 1, 1, 0, 0, 0, 0], and (d+w) is:

[0054]

[0055] (d+w) is mapped into an 8PSK (modulation order is 3) digital modulation symbol through mapping processing:

[0056] ​​

[0057] a2. A specific example of differential modulation of digital modulation symbol s to obtain differential modulation signal x is as follows: Taking QPSK modulation (modulation order γ = 2) and 4-bit quantization (a = 4) as an example, assuming the length of the mapped digital modulation symbol s is P = 8, s is:

[0058]

[0059] According to differential modulation rules The differential modulation signal x can be calculated as follows:

[0060]

[0061] a3. The base station uses a character-by-character differential demodulation method to obtain the differential demodulation signal from the received signal r. A specific example is as follows:

[0062] Assuming the received signal According to the character-by-character differential tone rule Differential demodulation is performed on the received signal r to obtain the differential demodulated signal. by The first element For example, the specific solution process is as follows:

[0063] Constellation Collection The CCP includes Four constellation points, first place the first constellation point Substitute Re[s] t ·r i * r i-1 The result obtained is 1.

[0064] Then, substituting the second to fourth constellation points in turn, the results are 0, 0, and -1 respectively;

[0065] The result is maximized when the first constellation point is substituted, then we get...

[0066] Solve for the remaining elements in r in the same way, and then the difference demodulation signal will be obtained. Ultimately, it can be expressed as:

[0067]

[0068] It should be noted that those skilled in the art will appreciate that the embodiments described herein are presented for the purpose of aiding the reader in understanding the method of implementing the present application, and should be understood as not limiting the scope of protection of the present application to such specific recitations and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical spirit of the present application disclosed herein without departing from the essence of the present application, and such modifications and combinations are still within the scope of protection of the present application.

Claims

1. A method of superimposed channel state information feedback with differential modulation aided parallel branch fusion, characterized in that, The method comprises the following steps: User terminal processing: S1. The modulated UL-US CSI w∈{0,1} P×1 The superimposed and sequentially mapped and differentially modulated signal is obtained and fed back to the base station. Base station end processing: The base station end processes the received signal according to the traditional processing branch and the learning processing branch in parallel, and then fuses them; S2. The received signal is processed according to a conventional processing branch, and the base station end performs the processing of the received signal character-by-character differential demodulation, inverse mapping, despreading, and dequantization processing to recover the modulated UL-US and downlink CSI recovery information S3. Learning processing branch: input the received signal into the learning branch network Infeature_Net to obtain CSI initial features S4. Will With The fusion network Fushion_Net, to get the recovery accuracy of the enhanced downlink CSI fused with the characteristics of two signals 2. The method of claim 1, wherein, In step S1, the real-valued is to make the downlink CSI vector Real-valued as real-valued CSI information The quantization is a-bit quantization on to obtain bit stream information m∈{0,1} 2aN×1 ; The spreading refers to spreading the matrix and m to obtain a spread signal w of the same length as d by the formula w = Qm. The differential modulation is to s = [s1, s2,..., s P ] T According to the rule The differential modulation is to s = [s1, s2,..., s 3. The method of Claim 2, wherein The mapping processing in step S1 further comprises the following sub-steps: S11. According to the modulation order γ1 adopted by the element digital modulation in UL-US d, the modulation order γ2 of the element symbol that can be mapped in the superimposed signal (d+w) is obtained as: γ2=γ1+1; S12. mapping each element in the superimposed signal (d + w) to a modulation symbol of order γ2, to form digital modulation symbols of order one higher than d 4. The method of Claim 3, wherein S2 comprises the following sub-steps: S21. Base station end pairs receive signals Performing character-by-character differential demodulation, inverse mapping to recover the bit stream form CSI And the modulated UL-US S22. The method of S21 further comprising: performing de-spreading and de-quantization to obtain downlink CSI recovery information The character-by-character differential demodulation is by solving obtaining a differential demodulation signal wherein, represents the i-th decision value; r i , r i-1 represents the i-th, i-1-th element in r respectively; s t represents the attempted constellation point value in the (γ+1)-th order digital modulation constellation set , the superscript * represents the conjugate operation, represents the s selected in the set t , Re[·] represents the real part; The inverse mapping refers to a process of recovering and ​ 5. The method of claim 4, wherein, Step S3 comprises the following sub-steps: S31. Constructing a dataset {r, h label}training the learning branch network Infeature_Net to obtain network parameters of the learning branch network Infeature_Net; S32. Training label is a real-valued vector obtained by real-valuing the downlink CSI vector h; S33. During online operation, input r into the learning branch network Infeature_Net to obtain CSI initial features 6. The method of claim 5, wherein, Step S4 comprises the following sub-steps: S41. The despreading processing refers to using the same spreading matrix as the user end by the formula obtaining the despread CSI in the form of a bit stream S42. The dequantization process refers to performing a-bit dequantization on the to obtain the initial features of the CSI S43. The fusion network Fushion_Net is trained to obtain network parameters by constructing a data set {h T ,h label} on Fushion_Net, h T is a set of training data when Fushion_Net is trained offline. ​

Images