DeepMUSIC: Multiple Signal Classification via Deep Learning
DeepMUSIC employs multiple CNNs on subdivided angular spectra for multi-target DOA estimation, outperforming traditional methods in accuracy and efficiency.
Key Findings
Methodology
The framework utilizes multiple deep CNNs, each dedicated to a specific angular subregion, taking the array covariance matrix as input. The angular spectrum is partitioned into Q non-overlapping segments, with each CNN learning the nonlinear mapping from covariance features to the MUSIC spectrum within its assigned subregion. During training, simulated multi-target data with varying SNRs are used, employing convolutional, normalization, ReLU, and fully connected layers to optimize spectral peak detection. The approach effectively combines spatial partitioning with deep feature extraction, enabling accurate multi-target DOA estimation with reduced computational load.
Key Results
- In simulations with 2 targets, RMSE was below 0.8°, outperforming traditional MUSIC (~1.2°) and single deep network models (~1.5°). For 6 targets, RMSE remained below 1° at SNR≥20dB, showing robustness. Computation time was approximately 0.002 seconds, significantly faster than spectral MUSIC (~0.03s), demonstrating real-time capability.
Significance
This work addresses the limitations of classical spectral methods in multi-target scenarios, introducing a scalable deep learning framework that overcomes target number constraints and reduces computational complexity. It paves the way for high-precision, real-time target localization in complex environments, impacting radar, sonar, and wireless communications industries.
Technical Contribution
The innovative multi-CNN architecture partitions the angular spectrum, enabling parallel processing and reducing training complexity. The use of array covariance matrices as input captures spectral features effectively, and deep convolutional layers enhance feature extraction beyond traditional linear methods. This design significantly improves multi-target estimation accuracy while maintaining low computational costs, setting a new standard for deep learning-based DOA estimation.
Novelty
This is the first work to propose a multi-network deep learning framework that divides the angular spectrum for multi-target DOA estimation, overcoming the limitations of prior models restricted to single or two targets. The approach combines spatial spectrum partitioning with deep feature learning, offering scalable and accurate multi-target localization.
Limitations
- The model relies heavily on simulated training data, which may not fully capture real-world array imperfections and environmental noise. In dense target scenarios, spectral overlap can cause misclassification. The network's complexity demands significant computational resources during training, and its robustness in highly dynamic or cluttered environments remains to be validated.
Future Work
Future research will focus on incorporating real measurement data to improve robustness, developing adaptive spatial partitioning strategies for dense target environments, and optimizing network architectures for deployment on resource-constrained platforms. Additionally, integrating array calibration and environmental awareness could further enhance performance.
AI Executive Summary
The challenge of accurately estimating multiple signal directions in complex environments has long limited the capabilities of traditional array processing methods like MUSIC. While highly effective for single or few targets, these methods struggle with increased target density and computational demands. Recent advances in deep learning have opened new avenues, but most approaches are constrained to simple scenarios. This paper introduces DeepMUSIC, a novel framework that employs multiple convolutional neural networks, each dedicated to a specific angular subregion, to perform multi-target DOA estimation efficiently.
The core innovation lies in dividing the angular spectrum into Q segments, training a CNN for each segment to learn the nonlinear relationship between the array covariance matrix and the MUSIC spectrum. During inference, multiple CNNs simultaneously produce spectral estimates, which are combined to locate multiple targets with high accuracy. Extensive simulations demonstrate that DeepMUSIC achieves RMSE below 0.8° for 2 targets and maintains performance for 6 targets at SNR≥20dB, outperforming classical MUSIC and single-network deep models. Its computational efficiency—taking only 0.002 seconds per estimate—makes it suitable for real-time applications.
This approach significantly advances array signal processing by enabling scalable, accurate, and fast multi-target localization. It addresses key limitations of existing methods, such as target number restrictions and high computational costs, offering a promising solution for radar, sonar, and wireless systems. Future work will focus on real-world validation, robustness enhancement, and adaptive spectrum partitioning, aiming to deploy this technology in dynamic, cluttered environments with high target densities.
Deep Dive
Abstract
This letter introduces a deep learning (DL) framework for direction-of-arrival (DOA) estimation. Previous works in DL context mostly consider a single or two target scenario which is a strong limitation in practice. Hence, in this work, we propose a DL framework for multiple signal classification (DeepMUSIC). We design multiple deep convolutional neural networks (CNNs), each of which is dedicated to a subregion of the angular spectrum. In particular, each CNN is fed with the array covariance matrix and it learns the MUSIC spectra of the corresponding angular subregion. We have shown, through simulations, that the proposed DeepMUSIC framework has superior estimation accuracy and exhibits less computational complexity in comparison with both DL and non-DL based techniques.