<p>Massive multiple-input and multiple-output (M-MIMO) is the prime technology of fifth-generation (5 G) communication systems. Antenna diversity and multiplexing gain in M-MIMO systems are achieved through feedback of precise downlink channel state information (CSI) to the base station (BS). The knowledge of CSI at the BS benefits dynamic scheduling, interference management, precoding, detection, and adaptive modulation. However, transmitting CSI to the BS is expensive in frequency division duplexing (FDD) due to the absence of the principle of channel reciprocity and limitations in the bandwidth of the feedback link. In this paper, we designed a deep learning (DL)-based framework DSCNN-AttNet applicable to the 5 G new radio clustered delay line (nrCDL) channel model that conforms with 3GPP specifications. The framework exploits depthwise separable convolution and an attention mechanism and is termed DSCNN-AttNet. For this purpose, a set of synthetic data is generated. The performance is competitive with baseline networks CRNet, CsiNet, MRFNet, CsiNet+, and DFECsiNet. The experimental results exhibit that DSCNN-AttNet demonstrates superior performance in terms of cosine similarity (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\rho $</EquationSource> <EquationSource Format="MATHML"><math> <mi>ρ</mi> </math></EquationSource> </InlineEquation>) and normalized mean square error (NMSE) for four different lengths of codeword (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$N_s$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>N</mi> <mi>s</mi> </msub> </math></EquationSource> </InlineEquation>) while maintaining computational complexity. The designed network provides a 5 dB improvement and a 33.5<InlineEquation ID="IEq3"> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> reduction in complexity. The applicability of the work is extended to the millimeter-wave (mmWave) channel to exhibit the efficacy of the framework in diverse scenarios. To verify the generalization of DSCNN-AttNet, it is trained on a mixed dataset, which shows the feasibility of the architecture in real-time communication scenarios.</p>

DSCNN-AttNet: A low complexity deep learning framework for CSI feedback in mmWave massive MIMO systems

期刊论文
发表日期：

2025年8月22日

查看全文

Anusaya Swain,
Shrishail M. Hiremath,
Sarat Kumar Patra

摘要

Massive multiple-input and multiple-output (M-MIMO) is the prime technology of fifth-generation (5 G) communication systems. Antenna diversity and multiplexing gain in M-MIMO systems are achieved through feedback of precise downlink channel state information (CSI) to the base station (BS). The knowledge of CSI at the BS benefits dynamic scheduling, interference management, precoding, detection, and adaptive modulation. However, transmitting CSI to the BS is expensive in frequency division duplexing (FDD) due to the absence of the principle of channel reciprocity and limitations in the bandwidth of the feedback link. In this paper, we designed a deep learning (DL)-based framework DSCNN-AttNet applicable to the 5 G new radio clustered delay line (nrCDL) channel model that conforms with 3GPP specifications. The framework exploits depthwise separable convolution and an attention mechanism and is termed DSCNN-AttNet. For this purpose, a set of synthetic data is generated. The performance is competitive with baseline networks CRNet, CsiNet, MRFNet, CsiNet+, and DFECsiNet. The experimental results exhibit that DSCNN-AttNet demonstrates superior performance in terms of cosine similarity ( $\rho $ $ρ$ ) and normalized mean square error (NMSE) for four different lengths of codeword ( $N_s$ $N_{s}$ ) while maintaining computational complexity. The designed network provides a 5 dB improvement and a 33.5 $\%$ $%$ reduction in complexity. The applicability of the work is extended to the millimeter-wave (mmWave) channel to exhibit the efficacy of the framework in diverse scenarios. To verify the generalization of DSCNN-AttNet, it is trained on a mixed dataset, which shows the feasibility of the architecture in real-time communication scenarios.