Deep-learning-based tool for anomaly detection and optical signal-to-noise ratio/Channel Spacing estimation in Elastic Optical Networks

Detecting anomalies and estimating system parameters are crucial for maintaining service quality in gridless Elastic Optical Networks (EONs). The absence of fixed Channel Spacing (ChSp) and the need for flexible resource allocation make anomaly detection and parameter estimation particularly challenging, especially when ChSp is close to or smaller than the Channel Bandwidth (ChBw). In this work, we propose a deep learning (DL)-based optical performance monitoring tool built upon a common neural architecture that can be parametrically configured for either classification or regression, depending on the target monitoring task. When configured in classification mode, the model detects anomalies, such as excessive noise levels and spectral overlap, whereas in regression mode, it provides an accurate estimate of ChSp or OSNR. This task-oriented configurability allows the same architectural framework to be applied across multiple monitoring functions without requiring fundamentally different model designs. We experimentally validated our method in 3 × 16- and 3 × 32-GBd 16-Quadrature-Amplitude-Modulation (QAM) Nyquist Wavelength-Division Multiplexing (WDM) systems. Our tool achieved approximately 98% and 89% accuracy in binary spectral-overlap classification for 16- and 32-GBd cases, respectively. ChSp estimation errors were limited to 0.3 GHz (16-GBd) and 0.9 GHz (32-GBd). Detection of low- and high-OSNR levels achieved approximately 96% and 81%–85% accuracy for 16-GBd and all the 32-GBd scenarios, respectively. Estimation errors less than 0.5 dB for the 16-GBd scenario and 0.8–1.6 dB for all the 32-GBd scenarios were achieved. These results demonstrate the robustness and versatility of the proposed OPM framework, which operates exclusively at the symbol level, thereby avoiding access to higher-rate signal samples or modifications to the Digital Signal Processing (DSP) chain of conventional coherent receivers, making it well-suited for future dynamic gridless optical transmission systems.