SE-Enhanced Representation Learning for Constellation Diagram Classification in Optical Networks

Accurate anomaly detection is essential for reliable operation of optical communication networks, particularly under complex interference conditions. This paper proposes a squeeze-and-excitation enhanced convolutional autoencoder (SE-CAE) framework for representation learning and classification of predefined 16QAM signal conditions. The SE mechanism is incorporated into the convolutional autoencoder to adaptively enhance anomaly-relevant channel features and suppress redundant information. The learned low-dimensional representations are subsequently classified using random forest (RF) and support vector machine (SVM) classifiers. Experiments are conducted on five representative signal conditions, including no interference, adjacent-channel interference, excessive interference power, reduced channel spacing, and mixed frequency-offset and power-imbalance interference. The proposed SE-CAE reduces the feature dimensionality by 99.82% while retaining discriminative information for anomaly classification. The SE-CAE-based features improve the RF accuracy from 86% to 89%, while the combination of SE-CAE and SVM achieves an accuracy of approximately 91%. The results demonstrate the effectiveness of attention-enhanced feature representation for the classification of predefined optical signal conditions.