Vibration signal recognition has become a hot spot in the field of optical fiber sensors. However, the majority of previous research focuses on tackling signal-source recognition. There is often more than one event co-occurring in most realistic scenarios. The aliased signal is generated when multiple unpredictable vibration sources are superimposed at the same fiber receiving point, which is challenging to recognize because the fusion of different event characteristics makes it highly disruptive. One possible way is to separate the mixed signals before identifying them. In this work, we introduce noise robust Transformer Demucs (NRTD), a spectrogram-based mask estimation model, with a spectral Demucs structure and Transformer. Specifically, we replace the innermost layer of Demucs with a spectral-domain feature extraction module based on self-attention to capture the inner spectral feature of the signal. Furthermore, the noise-robust block is designed to suppress the remaining scattered impulse noise in the spectral domain. Experiments on the VibDataset show that NRTD achieves sota results with 12.77 SDR (even surpassing the sophisticated model BSRNN). We also provide qualitative evaluations from both spectral and waveform perspectives, demonstrating a promising performance for multi-source aliased signal separation.