Aya Mostafa Ahmed,Udaya Sampath K.P. Miriya Thanthrige,Aly El Gamal,Aydin Sezgin
Abstract:We present a MUSIC-based Direction of Arrival (DOA) estimation strategy using small antenna arrays, via employing deep learning for reconstructing the signals of a virtual large antenna array. Not only does the proposed strategy deliver significantly better performance than simply plugging the incoming signals into MUSIC, but surprisingly, the performance is also better than directly using an actual large antenna array with MUSIC for high angle ranges and low test SNR values. We further analyze the best choice for the training SNR as a function of the test SNR, and observe dramatic changes in the behavior of this function for different angle ranges.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: How to improve the accuracy of direction - of - arrival (DOA) estimation in MIMO radar systems without increasing hardware complexity and computational cost. Specifically, the author proposes a deep - learning - based method, which uses a small - scale antenna array to simulate the signal reception effect of a large - scale antenna array, thereby achieving more accurate DOA estimation.
### Problem Background
Traditional DOA estimation methods usually rely on large - scale antenna arrays. Although this can improve the estimation accuracy, it also increases the hardware cost and computational complexity. In addition, in the case of low signal - to - noise ratio (SNR) and a high - angle range, the performance of existing methods will decline significantly. To overcome these problems, the author proposes a new strategy, using a deep neural network (DNN) to learn the mapping relationship between small - scale and large - scale antenna arrays, so as to improve the accuracy of DOA estimation without increasing the hardware overhead.
### Main Contributions
1. **Enhancement of DOA Resolution**:
- A DNN - based method is proposed to learn the signal mapping received between small - scale and large - scale antenna arrays.
- Experimental results show that in low - SNR and high - angle ranges, this method can not only approach the performance of the actual large - scale antenna array, but even exceed it.
2. **Analysis of Optimal Training SNR**:
- The relationship between the optimal training SNR and the test SNR is analyzed, and it is observed that there are significant differences in the function behavior in different angle ranges.
- It is found that for different test SNRs, choosing an appropriate training SNR can significantly improve the performance of DOA estimation.
3. **Analysis of Denoising Ability**:
- Based on the experimental results, the author finds that the proposed DNN has a denoising function and can effectively improve the SNR of the input signal under low - SNR conditions, thereby improving the accuracy of DOA estimation.
### Method Overview
- **System Model**:
- Consider a co - located MIMO radar system, which contains \( M \) transmitting antennas and \( N \) receiving antennas.
- The received signal \( r(t) \) can be expressed as the superposition of multiple target echo signals and is affected by additive white Gaussian noise.
- **Deep - Learning Architecture**:
- A feed - forward deep neural network (DNN) is used to learn the signal mapping received between small - scale and large - scale antenna arrays.
- The input layer and the output layer correspond to the received signals of the small - scale and large - scale antenna arrays respectively, and there are three hidden layers in between.
- **DOA Estimation**:
- Use the trained DNN to predict the received signal of the large - scale antenna array.
- Then estimate the DOA from the predicted received signal through the MUSIC algorithm.
### Experimental Results
- **Performance Comparison**:
- In low - SNR and high - angle ranges, the performance of DOA estimation using the DNN - predicted signal is better than that of directly using the small - scale antenna array and the actual large - scale antenna array.
- Especially under low - SNR conditions, the DNN shows stronger denoising ability, thereby improving the accuracy of DOA estimation.
- **Cramér–Rao Bound (CRB) Analysis**:
- Calculate the CRB of MIMO radar DOA estimation and compare it with the DOA estimation results of the DNN - predicted signal.
- The results show that the DOA estimation performance of the DNN - predicted signal is close to the CRB, especially under low - SNR conditions.
### Conclusion
This paper proposes a deep - learning - based method, which significantly improves the DOA estimation accuracy of MIMO radar systems by simulating the signal reception effect of large - scale antenna arrays. Experimental results show that this method performs particularly well in low - SNR and high - angle ranges and has strong denoising ability. In addition, the author also analyzes the selection of the optimal training SNR, providing guidance for practical applications.
### Formula Summary
- Received Signal Model:
\[
r(t)=\sum_{k = 1}^{K}\sum_{p = 1}^{P}\alpha_k^{(p)}a_r(\theta_k)a_t^H(\theta_k)s(t - pT)+n(t)
\]
where \( a_r(\theta_k) \)