Machine Learning-Based Modulation Format Identification Under Varying OSNR
Main Article Content
Abstract
A receiver must frequently determine the format of an incoming signal without previous knowledge because elastic optical networks alter a channel's modulation format to fit the quality of the link. In this work, four forms (QPSK, 8-QAM, 16-QAM, and 64-QAM) at optical signal-to-noise ratios (OSNR) ranging from 4 to 26 dB are identified using machine learning. Three classifiers—a support vector machine (SVM), a random forest (RF), and a tiny convolutional neural network (CNN) using 32 x 32 constellation images—are evaluated using synthetic 32 GBd signal frames with residual phase error. At OSNR ? 20 dB, all three models achieve 100% accuracy, whereas at 4 dB, they approach roughly 71–74%. At 12 dB, the CNN achieves 95% accuracy before the SVM and RF, which require 14 dB. The accuracy is 91.0% (SVM), 91.0% (RF), and 92.0% (CNN) on average across all OSNR values. Confusion between 16-QAM and 64-QAM is the primary cause of error below 12 dB. The CNN requires roughly 32 seconds to train, compared to less than a second for the SVM and RF. This means that the CNN's increased accuracy comes at a significantly higher training cost. The findings clearly illustrate how OSNR restricts format identification.