Existing studies on deep learning-based physical layer identification have mainly exploited raw in-phase/ quadrature (IQ) samples or power spectral density (PSD) samples as inputs independently. The raw IQ and PSD samples represent the information in the time and frequency domains, respectively. It has been observed from the results of existing studies that identification using raw IQ samples outperforms that using PSD in low signal-to-noise ratio (SNR) regimes, and that identification using PSD outperforms that using raw IQ in high SNR regimes. In this paper, we propose to use the fusion of raw IQ and PSD samples to enhance deep learning-based physical layer identification. In particular, we design three general fusion frameworks, i.e., input, feature, and decision fusions, and integrate them with three typical deep neural network architectures of fully connected neural network, convolutional neural network, and recurrent neural network to form fusion identification schemes. We conduct experiments using 50 off-the-shelf Wi-Fi devices to validate the concerned fusion schemes and investigate their performance gains in identification and model training. Our experimental results verify that the proposed fusion identification schemes can achieve comparable or superior identification performances to the state-of-the-art schemes in the entire SNR regime. Moreover, for the considered fusion schemes, we further investigate the impacts of fusion strategies, deep-learning networks, and SNR conditions on the identification performance and training time.