Stenio is pursuing a Ph.D. in optical communications at the Technical University of Denmark in the Machine Learning in Photonic Systems group. He received his bachelor’s degree at the University of Sao Paulo, Brazil, and in 2016 his master’s degree at the State University of Campinas, Brazil. From 2011 to 2018, he was working at CPqD Foundation, Campinas, Brazil. At CPqD, he developed and implemented state-of-art digital signal processing algorithms to be employed in a commercial physical layer ASIC for high speed (400G) optical communication transceivers. He also helped to found the Brazilian photonic society (SBFoton) in 2017, where he was the administrative director. Currently, he is in the fiscal council of the organization and a Marie-Curie Fellow.
Stenio’s Blog – the home of his blog is here
State-of-play: Update on Stenio’s research in FONTE
- Performance analysis of monitoring techniques based on machine learning
The constant increase in information due to inventions coming from internet-of-things, autonomous cars, and others, forces the network to re-inventing itself to keep demand efficiently. Optical performance monitor (OPM) techniques ensure that communication in an optical link, where most of the information goes through, is reliable. The current OPM techniques require full signal demodulation, which is too complex and expensive. Therefore, it is desirable to create a simple and cheaper mechanism to extract the necessary information. Using only photodetectors (PDs) to measure the power of the signal might be an alternative solution. However, due to the square-law detection, the phase information is lost in the process. In this scenario, machine learning (ML) techniques might be used to overcome this challenge. ML working as regression is a powerful tool that can determine a nonlinear relationship between input and output. Hence, a good promise in learning the relationship between the power signal and the channel parameters.
We show some of the state-of-art solutions to address the topic highlighted previously and report a new approach using an optoelectronic receiver with ML to mitigate the CD in the direct detected system. Although this technique is used for mitigation, the idea of using an optical pre-processing might be extendable to OPM. The optoelectronic receiver consists of slicing the spectrum and detecting each of them with a photodetector, followed by an ML technique to reconstruct the transmitted information. Through this process, our previous works showed an increase in the transmission reach compare to a single PD receiver. (from: FONTE Deliverable D3.3)
- System identification and parameter estimation
System identification is the field that studies techniques to build mathematical models of dynamic systems. It can be used when no previous information is available from the system or when there is a model, but some parameters are unknown. The latter can also be called parameter identification. This is a powerful tool to simulate and understand real complex devices. The basic steps to build a model of a real system starts by collecting information (data) about it. Followed by choosing a structure that will represent the desired system. An error function is defined to measure the difference between the collected information and the estimative from the model. The difference between both systems is used to improve the model and approximate it from the real system as close as possible. One of the most important step of building a model is the choice of the structure that will represent the system. This can be a hard limitation of the model. For example, trying to model a nonlinear system with a linear model. Over the years, many different approaches were developed in the literature for system identification. A possible classification between the varieties of possibilities in structure is dividing them by linear and nonlinear models. We give a general idea in how to apply system identification for linear system using a finite impulse response (FIR) filter and for a nonlinear system using Volterra filter. (from: FONTE Deliverable D3.2)
- Survey of machine learning algorithms for optical performance monitoring
An incredible 4.8 zettabytes of total yearly IP traffic data has been predicted by Cisco for 2022. New applications of internet of things, the increasing number of smartphones, Ultra High Definition (4K) streaming videos and many other applications are driving the necessity for higher transmission rates to support this increasing need for data. Since fiber-optic communications are the backbone of the telecommunications systems, new solutions to cope with this ever-increasing need for transmission speed are critical. Nowadays, nonlinear effects in optical fibers are one of the major limiting factors for optical communications. Different technologies have attempted to address these impairments for optical transmission. Nonlinearity mitigation through digital signal processing (DSP), optical phase conjugation (OPC) and nonlinear frequency division multiplexing (NFDM) are the most known topics in this area. Although great results show the increase of the transmission data rate and reach, the implementation costs of any of them are still prohibitive. It is also worth to mention that all the mentioned techniques require full knowledge of the fiber transmission parameters to work properly. Optical performance monitoring (OPM) techniques estimates the parameter of the optical fiber channel, which is required at multiples point along the link. The multiple uses of this technology imply the necessity of simpler and cheaper solutions.
Machine learning (ML) techniques may help solve the fiber nonlinearities for optical transmission and OPM, potentially reducing implementation costs. ML tools has a broad area of application and are very well known for being extremely effective for classification problems, typically for image classification and speech detection. Nonetheless, ML implementations with central processing units (CPU) are suboptimal in terms of speed and power efficiency. Considering the optical communication field, it also implies the necessity to convert the signal from optical domain to electrical/digital domain. In other words, an expensive solution. Alternatively, several researches using hybrid optical-electronic systems with reservoir computing (RC) and full optical neural networks (ONN) with programmable nanophotonic processor (PNP) have been demonstrated. We review these algorithms and point out potential advantages and disadvantages. (from: FONTE Deliverable D3.1)