SCIENCE CHINA Information Sciences, Volume 63 , Issue 6 : 160406(2020) https://doi.org/10.1007/s11432-019-2848-3

Deep belief network-hidden Markov model based nonlinear equalizer for VCSEL based optical interconnect

More info
  • ReceivedDec 15, 2019
  • AcceptedMar 18, 2020
  • PublishedMay 11, 2020



This work was supported by National Key RD Program of China (Grant No. 2019YFB1802904) and Joint Fund of the Ministry of Education (Grant No. 6141A02033347).


[1] Xie C. Datacenter optical interconnects: requirements and challenges. In: Proceedings of IEEE Optical Interconnects Conference, Santa Fe, 2017. 37--38. Google Scholar

[2] Tatum J, Landry G, Gazula D, et al. VCSEL-based optical transceivers for future data center applications. In: Proceedings of Optical Fiber Communication Conference, San Diego, 2018. M3F6. Google Scholar

[3] Meloni G, Malacarne A, Fresi F, et al. 6.27 bit/s/Hz spectral efficiency VCSEL-based coherent communication over 800 km of SMF. In: Proceedings of Optical Fiber Communication Conference, Los Angeles, 2015. Th2A.30. Google Scholar

[4] Rubsamen M, Winzer P J, Essiambre R-J. MLSE receivers for narrow-band optical filtering. In: Proceedings of Optical Fiber Communication Conference, Anaheim, 2006. Google Scholar

[5] Kottke C, Caspar C, Jungnickel V, et al. High speed 160 Gb/s DMT VCSEL transmission using pre-equalization. In: Proceedings of Optical Fiber Communications Conference and Exhibition, Los Angeles, 2017. 1--3. Google Scholar

[6] Karinou F, Prodaniuc C, Stojanovic N, et al. Experimental performance evaluation of equalization techniques for 56 Gb/s PAM-4 VCSEL-based optical interconnects. In: Proceedings of the 41th European Conference on Optical Communication, Valencia, 2015. 1--3. Google Scholar

[7] Tan Z, Yang C, Zhu Y. High Speed Band-Limited 850-nm VCSEL Link Based on Time-Domain Interference Elimination. IEEE Photon Technol Lett, 2017, 29: 751-754 CrossRef ADS Google Scholar

[8] Szczerba K, Lengyel T, Karlsson M. 94-Gb/s 4-PAM Using an 850-nm VCSEL, Pre-Emphasis, and Receiver Equalization. IEEE Photon Technol Lett, 2016, 28: 2519-2521 CrossRef ADS Google Scholar

[9] Tan Z W, Yang C C, Zhu Y X, et al. A 70 Gbps NRZ optical link based on 850 nm band-limited VCSEL for data-center intra-connects. Sci China Inf Sci, 2018, 61: 080406. Google Scholar

[10] Lavrencik J, Kota Pavan S, Thomas V A. Noise in VCSEL-Based Links: Direct Measurement of VCSEL Transverse Mode Correlations and Implications for MPN and RIN. J Lightwave Technol, 2017, 35: 698-705 CrossRef ADS Google Scholar

[11] Estaran J, Rios-Müller R, Mestre M A, et al. Artificial neural networks for linear and non-linear impairment mitigation in high-baudrate IM/DD systems. In: Proceedings of the 42th European Conference on Optical Communication, Dusseldorf, 2016. M.2.B.2. Google Scholar

[12] Chuang C, Wei C, Lin T, et al. Employing deep neural network for high speed 4-PAM optical interconnect. In: Proceedings of the 43th European Conference on Optical Communication, Gothenburg, 2017. W.2.D.2. Google Scholar

[13] Goodfellow I, Bengio Y, Courville A. Deep Learning. Cambridge: The MIT Press, 2016. Google Scholar

[14] Dahl G E, Dong Yu G E, Li Deng G E. Context-Dependent Pre-Trained Deep Neural Networks for Large-Vocabulary Speech Recognition. IEEE Trans Audio Speech Lang Process, 2012, 20: 30-42 CrossRef Google Scholar

[15] Liang A, Yang C, Zhang C. Experimental study of support vector machine based nonlinear equalizer for VCSEL based optical interconnect. Optics Commun, 2018, 427: 641-647 CrossRef ADS Google Scholar

[16] Rabiner L, Juang B. An introduction to hidden Markov models. IEEE ASSP Mag, 1986, 3: 4-16 CrossRef Google Scholar

[17] Badino L, Canevari C, Fadiga L, et al. Deep-level acoustic-to-articulatory mapping for DBN-HMM based phone recognition. In: Proceedings of IEEE Spoken Language Technology, 2012. Google Scholar

[18] Dahl G, Yu D, Deng L, et al. Large vocabulary continuous speech recognition with context-dependent DBN-HMMS. In: Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing, 2011. Google Scholar

[19] Mohamed A, Sainath T N, Dahl G, et al. Deep belief networks using discriminative features for phone recognition. In: Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing, 2011. Google Scholar

[20] Sheri A M, Rafique A, Pedrycz W, et al. Contrastive divergence for memristor-based restricted Boltzmann machine. Engineering Applications of Artificial Intelligence 2015, 37: 336-342. Google Scholar

[21] Rabiner L R. A tutorial on hidden Markov models and selected applications in speech recognition. Proc IEEE, 1989, 77: 257-286 CrossRef Google Scholar

[22] Newport, 1784 VCSEL Datasheet. 2017. https://www.newport.com.cn/p/1784. Google Scholar

[23] Garrido M, Kallstrom P, Kumm M, et al. CORDIC II: a new improved cordic algorithm. IEEE Trans Circ Syst II, 2016, 63: 186--190. Google Scholar

  • Figure 1

    (Color online) The pre-processing of signals, the current transmitted signal and its ${L}$ adjacent signals are mapped into a new symbol. For the received signal, the current vector after sinc interpolation and its ${L}$ adjacent vectors are wrapped as the final feature vector.

  • Figure 2

    (Color online) (a) The structure of RBM; (b) the structure of DBN, which can be approximated as the stack of RBMs.

  • Figure 3

    (Color online) The training process of DBN-HMM based equalizer. The dataset is divided into ${N}$ sub-datasets, ${\rm~DBN}^{(m)}$ is trained with the $m$th sub-dataset ($m=1,2,\dots,N$).

  • Figure 4

    (Color online) The equalizing process of DBN-HMM based equalizer.

  • Figure 5

    (Color online) Experiment block diagram of VCSEL based optical interconnect link with DBN-HMM. The input data is generated with bit-pattern generator (BPG) using random pattern (56 Gb/s).

  • Figure 6

    (Color online) Measured BER vs. ROP with different equalizing strategies. The adjacent number $L~=~2$, the BER performance of MLSE and ANN are also given. In ANN(8, 2, 20), 8, 2, 20 represent the interpolation multiple, adjacent number, and the number of neurons in hidden layer respectively. Compared with MLSE, the BER performance of DBN-HMM(8, 2, 15) can be greatly improved. DBN-HMM(8, 2, 15) and ANN(8, 2, 20) have the same computational complexity.