Comprehensive eye diagram analysis: A transfer learning approach

Abstract

A deep transfer learning (TL)-based comprehensive eye diagram analysis and diagnosis scheme that can output essential eye diagram parameters, estimate fiber link length, calculate Q-factor, and diagnose device imperfection-induced impairments is proposed. TL can be used to extract system information and optical signal characteristics contained in eye diagrams and apply the learned knowledge and extracted features obtained from source tasks to related target tasks. As a source task, the proposed method estimates the transmission distance of a fiber link using convolutional neural network (CNN)-based eye diagram recognition. The feature extraction layers of the CNN are transferred to six target tasks involving the recognition of cross percentage, levels '0' and '1,' eye height and width, and Q-factor. Using TL reduces the total training times for on-off keying (OOK) and pulse amplitude modulation (PAM4) formats by >95% and 60%, respectively. We also investigated six common PAM4 impairments caused by transmitter imperfection by setting the impairment category identification as source task and the impairment-degree diagnoses as target tasks. The TL methods consistently outperformed non-TL methods, with higher accuracies and significantly reduced training times. The proposed impairment diagnosis technique should be useful in impairment healing and fault correction.