Design And Research Of Optical Neural Network For Modulation Format Recognition In Flexible Optical Networks

In recent years,the channel bandwidth has become an important factor that restricts the transmission rate of optical networks.The current deployed optical cross-connects(OXCs),based on wavelength selective switches,are typically able to switch wavelengths.However,the optical network based on OXCs are limited by its large fixed grid which are determined by fixed channel bandwidth as small as 50 GHz.Which employs smaller grid(12.5 GHz)can effectively compensate for the insufficiency of the fixed grid network and improve the optical network transmission capability and spectrum resource efficiency.In the realization of the flexible optical network,the modulation format is an important direct impact on the allocation of spectrum resources.Therefore,introducing a faster and more accurate modulation format recognition(MFR)into the flexible optical network would further improve the transmission rate.For example,node routing and resource allocation strategy brought by modulation format transformation will become more accurate,and the multi-path transmission resources in flexible optical network can be more efficiency optimized by fast MFR.In addition,this method reduces power consumption of network and effectively improves network resource efficiency while ensuring transmission performance.The traditional methods of MFR mainly include decision theory and neural networks.Decision theory based MFR has the advantages of simple steps,low accuracy and high precision at high signal-to-noise ratio.However,it requires a lot of prior knowledge as support,so it is difficult to make decision in real time.In addition,it is easy to cause large errors because of its poor robustness.With the development of neural network technology,it has been widely applied into MFR,leading to a higher accuracy rate and an increase in processing speed.However,as the dramatic increasing of the data volume in big data era,the computing capability of traditional central processing units(CPU)can't meet the needs.Although many other computing hardwares,including field-programmable gate arrays(FPGAs),application-specific integrated circuits(ASICs)and graphical processing units(GPUs),has been developed to improve the computing speed,their computing speeds are limited by electron mobility and operating frequency.Since light has advantages of large bandwidth,high transmission speed and low crosstalk,it is able to overcome the bottleneck of the speed of traditional electric computing hardware.Therefore,we propose a novel method employing photonics information processing not only to increase the rate of MFR but also to maintain the high recognition accuracy.The MFR is realized by a method called optical neural network(ONN),which realize matrix multiplication by an optical on-chip platform mainly consisted by several Mech-Zender interferences(MZI).In this paper,the modulation formats of the four signals 2PSK,2ASK,4PSK,and 4FSK are identified.The overall recognition rate can reach 96%when the SNR is OdB.At the same time,for the theoretical analysis of the method,the transmission time is limited only by the physical size,the spectral bandwidth(THz)of the dispersion component,and the photoelectric detection rate(100 GHz).Therefore,the ONN is increased by two to three orders of magnitude compared to the frequency of the CPU(3-5 GHz).If the length of the chip is 10 cm,the scattering loss of the waveguide is calculated to be 3 dB,that is,the loss of the entire chip will be about 50%.However,compared with traditional neural networks,ONN energy consumption is several orders of magnitude smaller.After that,some improvements have been made to this method,such as matrix decomposition,new neural network training feature parameters,etc.,and considering the application of ONN to more scenarios.This method has been demonstrated with the characteristics of high recognition rate,strong fabrication tolerance and in principle high recognition speed.It should be noted that the proposed ONN has potential in dealing with numerous mode recognition issues in the field of signal processing.