In the realm of computational mechanics, reliably detecting, characterizing, and quantifying nonlinear mode interactions in vibrating structures remains a pivotal yet challenging endeavor. This study intro- duces a methodology that synergistically combines post-hoc interpretability with deep learning to analyse response signal time series and to reveal physically meaningful signatures of interaction without relying on handcrafted diagnostics. The methodology builds on our previous work on single-degree-of-freedom (SDOF) [1], and begins with the generation of a comprehensive training dataset obtained by solving parameterized equations of motion for multi-modal structural systems under controlled variation in de-tuning, damping, coupling strength, excitation type, amplitude, and measurement noise. A neural network model, trained for multiclass classification, is then employed to discern the presence of mode interaction and to categorize interaction regimes (e.g., non-interacting, weakly interacting, strongly interacting) directly from raw time-series response histories. The overall workflow is complemented by interpretability, to understand what portions of data the model deemed important for distinguishing the different classes. The results demonstrate a strong capability to detect and categorize nonlinear mode interactions, while providing explanations that are auditable and aligned with physically-interpretable interaction signatures, marking a practical contribution to trustworthy AI-assisted diagnostics in structural dynamics.