Supercontinuum generation (SCG) is a nonlinear phenomenon that is highly sensitive to noise. Noise can cause spectral fluctuations that degrade coherence and stability in practical applications. While recent machine learning approaches have mainly focused on regression-based spectral prediction, classifying noise-induced degradation has received comparatively little attention. In this work, we introduce a deep learning framework for multi-class classification of noise levels in supercontinuum spectra generated in a dispersion-engineered silicon nitride (Si3 N4) waveguide. A dataset of 1920 spectra is created using generalized nonlinear Schrödinger equation (GNLSE) simulations under varying pump wavelengths, pulse widths, and peak powers with controlled phase noise of different levels. Noise strength is quantified through average spectral deviation about the -40 dB spectral bandwidth and divided into five discrete classes ranging from No Noise to High Noise as per the magnitudes. The proposed hybrid architecture combines a one-dimensional convolutional neural network (1D-CNN) for spectral feature extraction with a multilayer perceptron (MLP) that ensembles the input laser parameters. The optimized model achieves an accuracy of 94.27% and a macro-averaged F 1 -score of 0.94, reliably classifying all noise levels. These results demonstrate that classification-based noise evaluation offers an interpretable and scalable approach for monitoring spectral quality in integrated broadband sources.