Researchers harness deep learning to analyze optical signals after fiber transmission
About Optica
04 November 2024
Researchers harness deep learning to analyze optical signals after fiber transmission
New approach could enable networks that automatically adjust to evolving requirements
Researchers have developed a low-complexity hybrid convolutional neural network (CNN) that can identify important optical signal properties after fiber transmission. This advance in signal monitoring could make it possible for networks to autonomously and quickly adapt to changing demands, enabling more efficient management of high-volume data transmission over long distances.
Gabriel Saavedra from Universidad de Concepción in Chile will present this research at the Latin America Optics and Photonics Congress, held 10 – 14 November 2024 in Puerto Vallarta, Mexico.
“This work provides a high-precision solution to estimate the quality of a signal transmitted through an optical network,” said Saavedra. “This is a key metric to optimize optical networks during operation, allowing for the connection of more users or an increase in the data rate of existing ones.”
As data traffic continues to grow, networks must handle more complex, high-speed transmissions. Fast, autonomous reconfiguration of optical networks will be necessary to ensure that these networks can dynamically adjust to changing demands, prevent congestion, optimize bandwidth usage and maintain seamless connectivity. However, making reconfiguration decisions requires a reliable way to monitor the modulation format and the signal quality of optical signals traveling through the network.
To address this challenge, the researchers developed a low-complexity CNN to analyze the optical signal’s modulation format and signal-to-noise ratio, which is an important indication of signal quality. Low-complexity CNNs perform effectively while using fewer layers, which lowers their computational demands. The new CNN was used to classify images of constellation diagrams — graphical representations of the modulation scheme — formed at the output of a coherent receiver.
To develop the hybrid CNN model, the researchers used both simulation and experimental data to generate a dataset of received constellations. After showing that the model could be successfully adapted to different inputs and generalized for different types of images outside of training, they tested its ability to classify constellations. They report an overall accuracy of 91.38% and an average error of 0.83 dB for estimating the signal’s modulation format and signal-to-noise ratio.
“The next steps for this research are further improvement in the accuracy and the processing speed,” said Saavedra. “After this, we will look into ways to integrate this solution into existing equipment or infrastructure.”
About the Optica Latin America Optics and Photonics Conference (LAOP)
LAOP is the major international conference sponsored by Optica in Latin America with the explicit objective to promote Latin American excellence in optics and photonics research and support the regional community. LAOP is a peer-reviewed, international meeting with content presented in English, which enables maximum international participation. Featuring a comprehensive technical program with recognized experts in fields critical to Latin America, the conference covers all major areas of optics and photonics and features the latest research results that are making an impact in fundamental research and applications.
About Optica
Optica, Advancing Optics and Photonics Worldwide, is the society dedicated to promoting the generation, application, archiving and dissemination of knowledge in the field. Founded in 1916, it is the leading organization for scientists, engineers, business professionals, students and others interested in the science of light. Optica's renowned publications, meetings, online resources and in-person activities fuel discoveries, shape real-life applications and accelerate scientific, technical and educational achievement. Discover more at: Optica.org
Media Contact
