Signal decomposition is an important task in Prognostics and Health Management (PHM) of mechanical transmission systems characterized by nonlinear and non-stationary vibration responses. However, conventional signal processing methodologies are often constrained by fixed basis functions and manual parameter tuning, which compromise noise robustness and limit their ability in capturing transient signatures of incipient faults. Leveraging the paradigm of signal processing-informed neural networks, this paper proposes an interpretable Frequency-Aware Residual Decomposition Network (FARCD-Net), which unfolds iterative filtering into a learnable framework to reduce reliance on manually designed optimization objectives. Specifically, a Generalized Adaptive Decomposition Module (GADM) is constructed via residual learning to generalize the conventional filtering operator, preserving the underlying decomposition principle. To enhance feature extraction in noisy environments, a data-driven Frequency Masking Layer (FML) is introduced. The layer learns adaptive spectral weights, enabling the network to dynamically localize transient signatures within resonance bands. Quantitative and qualitative evaluations on a run-to-failure benchmark and a high-speed train bogie system demonstrate effective extraction of fault-induced transients, with physical interpretability supported by intermediate network representations.