Integrated Deep Recognition for Overlapping Spectrum Signals in IoT Edge Monitoring Systems

With the rapid proliferation of Internet of Things (IoT) devices, overlapping spectrum signals (OSS) have become increasingly prevalent due to concurrent transmissions over the same frequency bands. Accurate recognition of OSS, including their categories, positions, and transmission powers, is crucial for spectrum management and interference mitigation in dense wireless environments. This paper proposes an integrated deep signal recognition (IDSR) architecture based on deep non-negative matrix factorization (Deep-NMF) for multi-parameter OSS estimation. To reduce computational complexity and accelerate iterative convergence, we further develop a collaborative multi-layer signal recognition (CMSR) architecture that decomposes OSS features through a three-layer structure, enabling successive estimation of category, position, and power. Additionally, we incorporate a Dempster-Shafer (D-S) based fusion strategy for signal category recognition and introduce a consensus-constrained NMF algorithm for collaborative localization across monitoring nodes. Particle swarm optimization (PSO) is employed to initialize the power and position estimation to enhance accuracy. Experimental results demonstrate that both IDSR and CMSR outperform traditional blind source separation methods (e.g., PCA, ICA, and standard NMF) and classical deep learning models (CNN and Transformer) in OSS recognition. Specifically, the proposed methods achieve up to 95.6% accuracy in signal classification under SNR = 4 dB, with comprehensive performance gains. Moreover, CMSR attains comparable accuracy to IDSR with lower computational complexity, indicating its potential for real-time edge deployment. The impact of monitoring node placement, signal power ratios, and deployment density is also analyzed, providing practical insights for spectrum monitoring in IoT networks.