Fujitsu Deep Learning Tech Improves Optical NetworkingCreated March 13, 2018
Three Fujitsu organisations have developed a deep learning technology framework to simplify the building, operating, and managing of optical networks. The technology, pioneered by
Fujitsu Laboratories Ltd, Fujitsu Laboratories of America Inc, and Fujitsu R&D Center Co, Ltd, .is used to estimate optical signal transmission parameters from optical receivers.
The companies report that the technology uses deep learning that can be trained on parameters to avoid the impact of systemic errors in optical signal transmission. It learns to estimate optical transmission signal parameters, which is an issue unique to optical communication systems, including calculating symbol rate and Optical Signal to Noise Ratio (OSNR).
The companies developed an experimental transmission system within Fujitsu Laboratories that emulates an optical network. With about 10,000 pieces of data, Fujitsu verified that this technology could estimate OSNR with a measurement error of 1%, and could estimate modulation format and symbol rate with a measurement error of 5%. Using this technology, it is expected that tasks that took an expert using specialised measurement devices several days to complete can now be estimated remotely in a matter of minutes.
Previously, when building an optical network, or when problems arose in operating a network, it was necessary to send an expert in this field with expensive and specialised measurement equipment to a worksite, and conduct measurements and tests to determine the cause. In optical networks that aim to boost capacity and distance, the increasing complexity of types of optical transmission signals and device parameter settings means that building the network or fixing issues may require several days, leading to significant issues in quickly building and managing fibre optic networks.
The newly developed technology trains a deep neural network by inputting the signals received by optical receivers into the network. By using the results of measurement equipment to provide supervisory labels, this technology trains the deep neural network to recreate the measurement results produced by the equipment, enabling it to estimate the optical signal transmission parameters.
Since systemic errors can arise in signal characteristics such as laser frequency when an optically transmitted signal has been received, if the received data is used for training as-is, the neural network will be trained to specialise on erroneous states. This could increase measurement errors in estimates.
As a way to counter this, the new technology virtually generates signals based on optically transmitted signals in varying states, for example, virtually generating multiple data with different laser frequencies, and then combining these to form the training dataset. In so doing, it becomes possible to reflect a variety of situations in the training results, enabling this technology to minimise measurement errors in estimates.
Going forward, Fujitsu will proceed with trials in an actual network environment, with the goal of commercialising this technology in fiscal 2019 or beyond. The companies will additionally continue investigation aimed at automatic operation of optical networks.