Source-linked AI summary
DeepHSIC: Deep Learning-based Signal Detector for Hybrid Downlink IM-NOMA
Dung Nguyen Tran, Toan D. Gian, Tien-Hoa Nguyen, Mai Xuan Trang, Tien-Cuong Nguyen, Thien Van Luong
TL;DR
Hybrid IM-NOMA detection must balance reliable recovery with the growing complexity of ML and the error propagation of SIC. DeepHSIC replaces computationally demanding SIC operations with DNN-based inference over preprocessed received-signal features. It achieves near-optimal BER under perfect and imperfect CSI while substantially reducing detection time.
Problem
ML detection becomes computationally expensive as users and subcarriers increase, while SIC can propagate early detection errors and fall short of ML performance.
Method
DeepHSIC uses zero-forcing-preprocessed received features and sequential DNN blocks that estimate user symbols while replacing the computational SIC decision block.
Results
DeepHSIC achieves near-optimal BER versus conventional detectors under perfect and imperfect CSI while requiring lower runtime, reducing detection time by up to 7 times versus ML and 3 times versus SIC.
Takeaways & Limitations
DeepHSIC provides a scalable learned receiver design for hybrid downlink IM-NOMA systems and can be extended to larger user configurations and time-varying channels.
Takeaways & Limitations
The detector is evaluated for hybrid IM-NOMA detection with conventional-detector comparisons, while ML complexity grows with the number of users and subcarriers and SIC can suffer error propagation.
Abstract
from arXiv · showhide
DeepHSIC is introduced as a neural receiver for hybrid downlink IM-NOMA transmission. The considered scheme combines power-domain NOMA with a composite OFDM/OFDM-IM waveform, so that user information is mapped jointly onto constellation symbols, subcarrier-index patterns, and different power levels. Although maximum-likelihood detection can achieve strong reliability for this model, its search space grows rapidly with the number of users and subcarriers. Conventional SIC reduces part of this burden, but its sequential cancellation may still accumulate errors and does not fully exploit the structure of IM-NOMA signals. To address this limitation, the proposed detector embeds dedicated deep neural network modules into the receiver and replaces the most computationally demanding SIC operations with learned inference blocks. The receiver is trained for Rayleigh fading channels and uses preprocessed channel-output features to recover user symbols. Simulation results show that DeepHSIC reaches BER performance close to model-based detectors under both perfect and imperfect CSI while requiring substantially lower detection time. These results indicate that learned SIC-style detection is a practical candidate for scalable hybrid downlink IM-NOMA receivers.
1 Introduction
Hybrid IM-NOMA combines OFDM-IM’s selective subcarrier activation with NOMA’s power-domain multiplexing, but detection becomes challenging as system size grows. DeepHSIC addresses this challenge by replacing computationally intensive SIC operations with learned deep-neural-network blocks.
- OFDM-IM activates only a portion of subcarriers, using their indices to convey additional information while improving spectrum usage and lowering peak-to-average power ratio.
- Power-domain NOMA superimposes signals for multiple users on shared time-frequency resources using distinct power allocations and SIC for separation.
- Hybrid IM-NOMA integrates OFDM and OFDM-IM through NOMA superposition, jointly distinguishing streams by subcarrier indices and power levels.
- Maximum-likelihood detection is computationally prohibitive for many users and subcarriers, while SIC remains vulnerable to detection-error propagation.
- DeepHSIC replaces the computational block in SIC with dedicated DNN blocks to provide near-optimal detection with lower complexity for hybrid IM-NOMA signals.
2 Hybrid IM-NOMA system model and baseline detectors
Hybrid IM-NOMA combines conventional OFDM, OFDM-IM, and power-domain NOMA so information is conveyed through symbols, active-subcarrier indices, and user power levels. ML offers strong detection but scales poorly, while SIC lowers complexity yet remains vulnerable to error propagation.
- Traditional OFDM Block: OFDM transmits constellation symbols across all N subcarriers, providing the conventional full-subcarrier baseline.Each subcarrier carries an M-ary constellation point such as a QAM symbol.
- OFDM-IM Block: OFDM-IM activates selected subcarriers and maps information to both constellation symbols and active-subcarrier indices.Its signal carries modulation bits and index bits; with K active subcarriers, there are C(N,K) possible active-index choices.
- NOMA Block: Hybrid IM-NOMA superposes multiple user streams over shared frequency resources using distinct power coefficients.The far-user stream receives higher power, while the near-user stream is recovered after subtracting the reconstructed far-user component.
- System Model: The composite transmission is modeled over Rayleigh fading channels, with users recovering target signals from a broadcast superposition.The channel and noise vectors capture complex channel gains and additive white Gaussian noise.
- Maximum Likelihood Detection: ML selects the feasible transmitted signal minimizing Euclidean distance but becomes computationally expensive as users and subcarriers increase.The method searches the signal space using the channel model and possible transmitted signals.
- Successive Interference Cancellation: SIC reduces computation relative to ML by sequentially estimating and removing user components, but early detection errors can propagate to later users.Its performance also depends on separation between user power levels and may fall short of ML under challenging channel conditions.
3 Proposed DeepHSIC detector
DeepHSIC replaces computationally intensive SIC decision processing with a preprocessing stage and dedicated DNN blocks for user detection. The network is trained on one-hot symbol targets using adaptive optimization and variable-SNR data.
- Network architecture: DeepHSIC uses preprocessing followed by two DNN blocks to detect transmitted symbols for the two-user case.The preprocessing converts received signals and channel information into features before DNN-based detection.
- Preprocessing: Zero-forcing equalization converts the received vector into a feature vector by concatenating the equalized signal’s real and imaginary parts.The equalized signal is produced by applying the inverse channel matrix to the received vector.
- User detection: DNN1 estimates the first user, while DNN2 receives that estimate with the preprocessed features to detect the next user and mitigate inter-user interference.Detected symbols are obtained by applying argmax to the DNN outputs.
- DNN structure: Each DNN block uses fully connected layers, ReLU activation, batch normalization, and Softmax to produce probabilities over feasible transmitted symbols.The final detected symbol corresponds to the largest Softmax output and its one-hot index.
- Training procedure: Training minimizes mean-squared error between predicted and ground-truth one-hot vectors, with Adam updating parameters and SNR variation exposing the model to different channel conditions.The training procedure uses adaptive learning rates and momentum through Adam.
4 Simulation results
The experiments evaluate DeepHSIC against ML and SIC using BER and per-sample detection runtime under perfect and imperfect CSI. DeepHSIC achieves near-optimal or comparable BER while reducing detection time substantially.
- Evaluation design: The experiments compare DeepHSIC with ML and SIC using BER and detection runtime to assess reliability and implementation feasibility.The evaluation targets the hybrid IM-NOMA system.
- Simulation setup: The simulation uses two users, (N, K, M) = (4, 1, 4), and power coefficients P1 = 0.8 and P2 = 0.2.These settings define the principal system configuration reported for the experiments.
- Channel conditions: Channel-estimation errors are modeled with a Gaussian error term of zero mean and variance σ2_e = 0.01, enabling evaluation under ideal and nonideal channel knowledge.The estimated channel is formed by adding the error term to the true channel.
- BER performance: Under perfect CSI, DeepHSIC achieves near-optimal BER relative to ML and SIC, with a small gap from SNR = 15 to high-SNR values.The reported deviation may be associated with the training parameter λtrain under high SNRs.
- BER performance: Under imperfect CSI, DeepHSIC maintains BER comparable to traditional detectors while decoding soft information for each symbol.The result is reported as evidence of robustness to CSI degradation.
- Runtime complexity: DeepHSIC reduces detection time by up to 7 times compared with ML and 3 times compared with SIC.Runtime is measured during testing for each detected sample.
5 Conclusions
DeepHSIC replaces the computationally intensive SIC decision block with DNN modules for hybrid downlink IM-NOMA. It achieves near-optimal BER with lower runtime under both perfect and imperfect CSI.
- DeepHSIC replaces the computationally intensive SIC decision block with DNN modules.
- The detector jointly addresses OFDM-IM and OFDM signal components without modifying the mathematical system model.
- Near-optimal BER is achieved compared with conventional detectors under both perfect and imperfect CSI.
- Lower runtime supports a scalable receiver design for modern communication systems.