Designing of the Submarine Cable Backhaul Optical Transport Network and the Prediction of OSNR using Artificial Neural Network

In the era of the fast exponential increase of internet traffic, thereby, widespread deployment of IP over Optical Transport Network (OTN) necessitates the designing of a robust and stable backhaul network, so that the Services/Clients don’t experience any blackouts and outages from the International Bandwidth. Because of being a terrestrial backhaul transport network, the performance parameter is mostly the Optical Signal to Noise Ratio(OSNR). In this paper, in the 1st phase, motivated by live real network, it’s been fully designed the Submarine Cable Backhaul Transport Network from Cable Landing Station(CLS) to Destination with working, protection, and restoration path prioritizing the fidelity of the network. In the 2nd phase, collecting the real-time OSNR time series data from the live circuit, an Artificial Neural Network (ANN) Model has been proposed to predict the OSNR to monitor the performance quality. The Model ANN result shows the network capability for better OSNR assessment and forecasting.


Introduction
To connect the whole country International Internet Gateway (IIG), International Gateway (IGW), Interconnection Exchange (ICX), Internet Exchange (IX), International Long Distance Telecommunications Services (ILDTS), and International Private Leased Circuit (IPLC) to Singapore, the middle east, and Europe through the submarine cable system, a high-capacity Optical Transport Backhaul Network is a must. The OTN suffers from huge impacts if the signal quality degrades. Therefore, to keep up the integrity of the network, the OSNR plays a vital role because if OSNR falls, the whole network collapses. This paper proposes a design of an OTN network, and for early detection of the OTN failures and anomalies in the OSNR, an ANN has been created there with letting not the signal quality deteriorate by analyzing the ANN forecasted OSNR(Allogba et al., 2022).

100G Line Card
This 100G line card is where the channel originated and terminated. Being compatible with International Telecommunications Union-Telecommunications sector (ITU-T)G.694.1, the (100G Form factor Pluggable) CFP interface of this card supports from 192 THz-195.95THz (Antil, et al., 2012) at 50GHz spacing (192 THz+50GHz*79=195.95 THz). The CFP could be tuned to any of these 80 channels (C band), and in the future, it could be extended to 191.35 THz and 196.1THz (extended C Band) (Mubarakah, et al., 2020) and also tunable to transmit (Tx) power. The Tributary board (Client Card) and the line card work with Cross-connect Boards. The Tributary board receives client signals, performs O-E conversion, maps the services into ODUk(Optical Data Unit k=0,1,2,3,4)containers, and sends the ODUk electrical signals to the line card via a cross-connect board. The line card multiplexes and maps ODUkinto OTUk(Optical Transport Unit)and converts the OTUk signals into standard Dense Wavelength Division Multiplexing (DWDM)lamda/channel. The line card receives power must be within -2dBm to -18dBm (the best is -6dBm), and the OSNR sensitivity (B2B) is 13dB. The 100G transceiver uses Polarization Multiplexed Quadrature Phase Shift Keying (PMQPSK) coherent modulation and being coherent modulation/optics, with SD-FEC (Forward Error Correction) and 60,000 ps/nm dispersion tolerances, the 100G Channel can reach up to 3529km without regeneration.
The embedded high-speed Digital Signal Processor(DSP)technology compensates for dispersion and Polarization Mode Dispersion (PMD). The line card can map 80 port ODU0/40 port ODU1/10 port ODU2/2 port ODU3+2 port ODU2 signal sent by cross-connect board to OTU4 and converts it to DWDM standard lamda.

Erbium Doped Fiber Amplifier (EDFA)
Digital coherent detection with SD-FEC enables 100G lamdas to transmit over 4000km with just the implementation of EDFAs-this is the exquisite side of EDFA.
Erbium Doped Fiber Amplifier (EDFA)has an amplification bandwidth of 4THz (80*50 GHz spaced lamda). The EDFA is used (Sajjan et al., 2015) at three locations in a DWDM System:

Optical Booster Amplifier (OBA)
It boosts the signal at the transmitter end to compensate relatively low output power of the laser.

Optical Pre-amplifier (OPA)
It boosts the signal before the optical receiver to increase receiver sensitivity.

Optical Line Amplifier (OLA)
It's directly inserted into the optical transmission link to amplify signals.
Considering the future fiber loss, the 3121 model, 5 dB noise figure (31 is maximum adjustable gain and maximum output power is 21) EDFA has been chosen for our design.

Mux De-mux Unit (MDU)
The MDU, made of thin fiber filter AWG (Array Waveguide Grating), is a 40-channel unit (even and odd), and it's purely passive. Even and odd channels have space of 100GHz according to G.694.1 frequency grid. Even channels cover from 192 THz to 195.9THz, and odd channels cover from 192.5 to 195.95 THz. The 80-channel, i.e., 80*100G=8Tbps per fiber capacity, is achieved after using an Inter-Leaver Unit or ILU (ILU incorporates both the even and odd channels). The MDU can integrate the line cards from another DWDM network-called alien lamda incorporation. The number of MDU is equal to the number of optical pathsANN

Re-configurable Optical Add-Drop Multiplexer (ROADM)
The ROADM is the heart of an agile network. In a ROADM card (Fig.1), signal is received at common port Rx and broadcast across all the Add/Drop Transmit Tx portswhereas the signal received at Add/Drop Rx ports is selected by Wavelength Selective Switch(WSS) and transmitted over common Tx port.

.1 Colorless
The line card is connected to the fixed frequency MDU port. If the line card is tuned to another frequency, then the patch cord has to shift to the tuned frequency port. For ROADM to be colorless, no matter what frequency the line card is tuned to, the patch cord does not need to be moved, i.e., the port will automatically tune to the line card frequency. Directionless means that any λ can be directed in any direction. In Fig.2, λ1, connected to port-1, can be directed to ROADM common port-2, or λ-common-1 can be directed to λ-common-2.

Contention less
The ROADM is contention less if the same λ can be simultaneously added/dropped in different directions. In Figure 2, λ1 at port-1 and λ1 at port-3 could be added and dropped in λ-common-1 and λcommon-2.

Flexible Grid
A flexible Grid is when λ spectral width is in the multiple of 12.5 GHz. It provides better spectral efficiency. Spectral width = n*12.5 GHz.

Optical Fiber Cable (OFC)
The OFC deployed in the OTN is the single-mode fiber complying with the G.652 specifications (Sajjan et al., 2015). Two main specifications are the attenuation (0.25dB/km) and dispersion coefficient (17ps/nm-km). Hence, if the 100G line card CFP dispersion tolerance is 60,000 ps/nm-km, then, lamda transmission distance D*17=60,000, thereby D=3529.7km. The 100G lamda can travel around 3529km just using the amplifier in the link.

The Network Planning and Designing
The designed network has 8 Tbsp per fiber capacity, but for the time being, it's been configured with only 400G circuits with theworking, protection, and restoration path.
Fig4: Complete design of the Network (blue fiber for working and red fiber for protection and restoration path).  Similarly, (Fig.4) the gain power adjustment is calculated from Destination to Source CLS of the working path.
The protection path ( Fig.3 & Fig.4) has 06 (Six) OLAs, 08 (Eight) OBAs and OPAs of the same3121model between source CLS to destination and vice versa. Whereas, in between there're 03 (three) 3 degree ROADMs (P2, P5 & P9) and they switch and amplify the lamdas of both the protection and restoration paths. The protection path also has 04(Four) channels/lamdas (λ) of 0dBm launching power from 04(Four) line cards at both the ends (Fig.4)  In the same way, the gain power adjustment is calculated from the Destination to the Source CLS of the protection path.
Following (Table4) is the gain adjustment for the restoration path. The protection and restoration paths are the same from source CLS to P2 and P5 to P9. The P2 to P5 and P9 to destination alternate routes are added for the restoration path (Table1 & Fig.3), hence, the restoration path has 09 (Nine) OLAs, 08 (Eight) OBAs & OPAs of the same3121 model between source CLS to destination and vice versa. Whereas, in between there're 03 (three) 3 degree ROADMs (P2, P5 & P9). The restoration path also has 04(Four) channels/lamdas (λ) of 0dBm launching power from 04(Four) line cards at both the ends (Fig.4) and all 4λs are amplified together from source CLS toDestination and vice versa at OBAs, OPAs and OLAs. The lamdas (λ) being fewer in number (4λs), it's been fixed the amplifier output to 2dBm ( The EDFA amplifies all the lamda altogether with the function of automatic gain control, but after amplification together, the per-channel power difference has to be less than 3dB for a healthy network. The bar chart in Fig.5 shows the per channel power:

Network components port connection
The ports are to connect physically with patch cords as well as on the software; the source port and the destination port are configured bi-directionally, viz, the MDU common ports are to connect with the ROADM add/drop ports, the add/drop ports of one ROADM are to connect with another ROADM add/drop ports. The ROADM common ports are to connect (unidirectional) with the OPA/OBA (Fig.2).

The Lamda switching/connections
The lamda can be switched from MDU common port/MDU add/drop ports to ROADM add/drop ports or common ports (directionless ROADM) (Fig.2).

The Network Performance Parameter
The key parameters that indicate the channel as well as the network robustness:

OSNR
The OSNR indicates the channel, circuit's healthiness, and better transmission quality. Higher the OSNR, the better the transmission performance (Gumaste et al., 2003). Factor affecting the OSNR is the Amplified Spontaneous Emission (ASE) noise from the EDFA than the nonlinearities

Bit Error Rate (BER)
The OSNR indirectly reflects the BERproviding the calculation of BER from OSNR as follows:  16.2 2.6*10^-2 0 If the OSNR would be less than the B2B tolerance of the line card (13dB), uncorrectable bit errors would be generated. However, in this case, the OSNR being higher than 13dB, the uncorrectable bits are 0 in number.

Generalized Multi-Protocol Label Switching (GMPLS)and the creation of circuits
GMPLS is the defacto control plane protocol for Wavelength Switched Optical Networks Wavelength Switched Optical Network (WSON). DWDM /OTN networks are WSON, where switching happens based on lamda.On the GMPLS enabledsystem, circuits can be added only while more parallel paths are running between the lamda-originated node and lamda terminated node, and at a time, at least two paths are active.
To create/add the 400G (40*10G) circuits between the source cable landing station and destination, the four 10*10G tributary/client cards, four 100G line cards for the working path, four 100G line cards for the protection path, and four 100G line cards for restoration path have to be arranged into 04(four) different shelves at both the ends (source CLS and destination), where in one shelf there'd be one 10*10G client card and three 100G line cards (one is for the working path, one is for the protection path and another is for the restoration path) (Fig.6).

Adding the GMPLS-Circuits
In a GMPLS-enabled circuit, the configuration can be done at any end, and the optical path (line card) would be selected automatically based on the OSNR received.
Before creating the circuits, the first thing to do is to make all the 10G ports (P1 to P10) of both ends' tributary cards UP and configure the ports as STM-64/10 GE LAN/ODU2/ODU2e as per the plan. Next, it needs to select circuit parameters (Fig.7), including the other end's IP address, and also select the slot number and port number under the ingress work channel option, and the egress work channel option to add/create a 10G circuit from one tributary card to another tributary card (Fig.8)(say, source CLS to destination). Among the three paths (working, protection, and restoration), the working path shows (comparing tables 5, 6, and 7) the best OSNR. Automatically, the working path is selected for the automatic cross-connection of the 10G port with the 100G line card. In the same way, we'd finish adding the 40*10G circuits.

Artificial Neural Network for forecasting the OSNR
To predict the OSNR and any anomalies in it, a Feed-Forward Back Propagation Neural Network (FFBPNN) has been created. The neurons are arranged into 03 (three) layers -the input layer, the hidden layer, and the output layer. The network hidden layer has 20 neurons with "logsig" as a transfer function, and the output layer has 1 neuron with "purelin" as a transfer function (Fig. 10). For the FFBPNN training, the "trainlm" activation function has been used. The Mean Squared Error (MSE) has been used as the default performance function.

Network Parameter
The learning rate controls the weight size and bias change and is set to a value of 0.01. The Epoch determines when the train will stop and is set to 8000. The RMSE or RootMean Squared Error is the square root of the sum of the squared difference between network target and actual output divided by no. of patterns. The goal is set to 1exp (-25 (Collected Data Sample) shows how every 15min the field OSNR data is collected for 125hrs. The time unit 0 to 5 indicates 0 to 125 hrs (Allogba, Et al., 2021).70% of the total collected data is used for training the network, and the left 30% is used to get the network to generate the OSNR (for testing and validation). Hence, from Fig.12 above and Table9, it's apparent that the target OSNR and the ANN generated OSNR have a good co-relation, matching and alignment. There're nominal anomalies in the ANN generated OSNR-thereby predicting the best OSNR (Fig.13), thereby forecasting the best signal quality.

Summary
The system parameters indicate the system as a robust one. However, there is still and always some rooms for improvement. In the first case scenario from the designing phase, it is observed that the working and protection paths are separately running in parallel (there're no connections between them). The protection path gets cut frequently and then the system becomes reliant on the working path (the protection and restoration path shares 439km of Fiber). In case indispensably some maintenance works run on the working path, then the 400G circuits get into a blackout. If both protection and working paths would be connected at the same node (ROADM) at some strategic locations, lamdas (λ) could be rerouted, and circuits would be more stable.
In the second phase of ANN prediction of OSNR, for best forecasting and early detection of OSNR, instead of using the FFBPNN, the OSNR data being sequential; it'd deliver more accurate result if the time series OSNR data could be fed into a Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) or Gated Recurrent Unit(GRU) RNN.

Conflicts of interest
Authors have no conflicts of interest.