Research map / Architectures and efficiency / RNNs and normalization
Model-Driven Deep Neural Network for Enhanced AoA Estimation Using 5G gNB
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MoD-DNN, which alternates a 1D-CNN spectrum calibrator with a sparsity-constrained conjugate gradient reconstruction to estimate AoA under hardware phase impairments using 5G gNB data.
Problem
AoA estimation with a 5G gNB array is degraded by angular-dependent hardware phase errors that break model-driven estimators such as MUSIC, while purely data-driven DNNs need large datasets.
Approach
Reformulates AoA estimation as reconstruction of a sparse spatial spectrum from the coarray spectrum, and unrolls an iterative optimization in which a 1D-CNN calibrates hardware-induced errors and a sparsity-constrained conjugate gradient (SCG) step recovers the sparse spectrum, with shared CNN weights across iterations.
Key results
- In anechoic chamber experiments, MoD-DNN reduces the 80th-percentile AoA error from about 3° to 0.15°, outperforming MUSIC, DeepMUSIC and CNN.
- In simulation, MUSIC and DeepMUSIC RMSE stays flat versus SNR, while MoD-DNN consistently beats the CNN baseline and its advantage grows at high SNR.
- MoD-DNN training takes 15.6 h versus 30.8 h for DeepMUSIC and 1.6 h for CNN; testing takes 14.9 ms versus 31.6 ms for DeepMUSIC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MoD-DNN instead of MUSIC | AoA estimation · Anechoic chamber, outer angular regions | maximum AoA estimation error | Helped |
| Use MoD-DNN instead of DeepMUSIC | AoA estimation · Anechoic chamber, outer angular regions | median AoA estimation error | Helped |
| Use MoD-DNN instead of CNN | AoA estimation · Simulation, varying SNR | RMSE | Helped |
| Use MoD-DNN instead of MUSIC or DeepMUSIC | AoA estimation · Simulation, increasing impairment degree ρ | RMSE | Helped |
| Use MoD-DNN instead of CNN or DeepMUSIC | AoA estimation · Simulation, training | convergence speed | Helped |
| Use MoD-DNN instead of other methods | AoA estimation · Anechoic chamber, 80th percentile estimation error | AoA estimation error: from around $3^{\circ}$ to $0.15^{\circ}$ | Helped |
| Use MoD-DNN instead of DeepMUSIC | AoA estimation · Training | time (h): 30.8 -> 15.6 | Helped |
| Use MoD-DNN instead of CNN | AoA estimation · Training | time (h): 1.6 -> 15.6 | Hurt |
| Use MoD-DNN instead of DeepMUSIC | AoA estimation · Testing | time (ms): 31.6 -> 14.9 | Helped |
| Use MoD-DNN instead of CNN | AoA estimation · Testing | time (ms): 1.8 -> 14.9 | Hurt |
Surprising findings
- The MoD-DNN is slower to train than the plain CNN (15.6 h vs 1.6 h), though faster than DeepMUSIC.
Limitations
- Evaluation limited to a single-path LoS setting in simulation and an anechoic chamber with no multipath or NLoS.
- Training cost (15.6 h) is higher than the plain CNN baseline.
Experimental setup
Datasets: simulated 5G CSI from link-level 5G simulator (3GPP 38.901 InF LoS) anechoic chamber measurements with commodity 5G gNB
Benchmarks: simulated 5G InF LoS AoA estimation anechoic chamber AoA experiment
Metrics: RMSE CDF of AoA error 80th percentile error boxplot median/IQR/max error training time testing time
Baselines: MUSIC DeepMUSIC CNN
Builds on: MUSIC DeepMUSIC
Closest papers by meaning
- Physics Unrolled Neural Operator for Wireless Field Modeling (NeurIPS 2026 · similarity 0.82)
- Robust Angular Synchronization via Directed Graph Neural Networks (ICLR 2024 · similarity 0.80)
- NeWRF: A Deep Learning Framework for Wireless Radiation Field Reconstruction and Channel Prediction (ICML 2024 · similarity 0.79)
- Batch Normalization Alleviates the Spectral Bias in Coordinate Networks (CVPR 2024 · similarity 0.78)
- Multitarget Device-Free Localization via Cross-Domain Wi-Fi RSS Training Data and Attentional Prior Fusion (AAAI 2024 · similarity 0.78)
- Differentiable and Learnable Wireless Simulation with Geometric Transformers (ICLR 2025 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.