Abstract: The application of fiber-based quantum key distribution (QKD) depends on the compatibility with existing optical networks. The use of wavelength division multiplexing (WDM) technology for cofiber transmission of classical data and quantum signals has the advantages of security, economy, and practicality. Aiming at the prominent problems that affect the application of the classical-quantum signal co-fiber transmission system, such as the difficulty in optimizing and calculating the average photon number of signal states, the number of decoy states and other parameters, and the slow running speed, the channel noise components are modelel and analyzed, and the finite-length effect and the decoy state method are evaluated considering the statistical fluctuations. Furthermore, based on the experimental data set, the back propagation (BP) neural network is trained to predict the system parameters such as the average photon number of signal states under different channel noise conditions. The results show that the predicted average photon number by the BP network is basically consistent with the original curve, and the training error is less than 10−3. It is indicated that the network can be used as an effective model for practical prediction of the parameters of the decoy-state classical-quantum co-fiber transmission system, which is of great practical significance for the development of quantum-secure communication towards high speed, large capacity and intelligence.

Key words: quantum optics, quantum key distribution, wavelength division multiplexed, machine learning, neural network