Quantum channel noise prediction method in multi-core optical fiber based on dataset dimensionality reduction

Abstract

The present invention discloses a method for predicting quantum channel noise in a multi-core optical fiber based on data set dimensionality reduction processing. This method evaluates the time delay introduced by quantum channel noise when a new service arrives when quantum signals and classical signals are transmitted in the same multi-core optical fiber, and establishes an effective noise prediction model based on machine learning. It is applicable to multi-core optical fiber quantum light transmission systems with any number of cores and any number of wavelengths. In response to the problem of too many features in the noise data set, the present invention uses a method to reduce the number of features in the data set based on the noise characteristics. Link noise is evaluated by using a noise prediction model established by two machine learning algorithms, XGBoost and LightGBM, in order to reduce the noise evaluation time. Compared with noise prediction models based on other machine learning algorithms, this method has the advantages of high prediction accuracy and fast prediction speed, and has made contributions to the practical application of quantum key distribution and classical signal fusion networking.

Description

Technical Field

[0001] This patent relates to the field of quantum communications, and in particular to quantum key distribution systems integrated with multi-core fiber optical networks. The invention provides a noise prediction method for multi-core fiber-based quantum key distribution optical networks. This method boasts rapid prediction speed, high accuracy, and a wide range of applicability, potentially promoting the practical application of quantum key distribution technology. Background Art

[0002] Quantum Key Distribution (QKD) provides keys to both communicating parties. Combined with a one-time pad, it theoretically achieves unconditional information security. Currently, point-to-point QKD technology is becoming increasingly sophisticated, significantly improving transmission distances and key generation rates. To reduce construction costs and increase transmission capacity, the integration of QKD networks with classical optical networks is currently under intensive research.

[0003] When classical signals and quantum signals are transmitted in the same optical fiber (hereinafter referred to as co-fiber transmission), due to the extremely low power of the quantum signal (usually below -80dBm), it is easily interfered by the noise generated by the classical signal (typical power value 0dBm), such as spontaneous Raman scattering noise (SpRS) and four-wave mixing noise (FWM). In order to further improve network capacity, the application of weakly coupled multi-core optical fiber based on space division multiplexing technology in co-fiber transmission system is being studied in depth. However, while multi-core optical fiber expands transmission capacity, it also introduces new noise, such as inter-core crosstalk noise (ICXT), which is about -40dB / 10km and will seriously affect the performance of the QKD system.

[0004] In order to obtain a higher key generation rate in a co-fiber transmission system, it is necessary to place the quantum signal on a channel with lower noise. Therefore, in an actual network, every time a service arrives, the channel on the link needs to be evaluated. This evaluation process will introduce additional delays and be reflected in the overall network delay. Currently, machine learning (ML) has been widely used in the field of communications to conduct research on channel estimation, channel detection, etc. to reduce the overall network delay, and can be extended to noise evaluation in co-fiber transmission systems. Therefore, the present invention proposes a quantum channel noise prediction method in a multi-core optical fiber based on data set dimensionality reduction processing to quickly evaluate the link noise in the co-fiber transmission system. Summary of the Invention

[0005] This paper addresses the problem of quantum channel noise assessment in multi-core optical fibers and proposes a method for predicting quantum channel noise in multi-core optical fibers based on dataset dimensionality reduction. This method is suitable for multi-core optical fiber quantum key distribution networks with any number of cores and wavelengths. Specifically, a dataset is established using the link channel state as the characteristic value and the noise level as the target value. An ML algorithm framework is used to fit the data and generate a noise prediction model. When new services arrive, a rapid and accurate prediction of link noise is achieved. The specific implementation steps are as follows:

[0006] Step 1: Create a noise dataset based on the number of core wavelengths of the multi-core fiber and the link noise characteristics of the channel;

[0007] In order for the noise prediction model to more accurately predict the noise value of each idle channel under different link states, a large amount of labeled data is required to train the model. The training data in this patent can be obtained by simulating the establishment of a co-fiber transmission system model, or by measuring in an actual co-fiber transmission system. Based on the number of multi-core optical fiber cores C and the number of wavelengths per core W used in the co-fiber transmission system, the link state is abstracted into a C×W state matrix, where if a channel is occupied by a classical signal, the element at that position in the state matrix is ​​set to 1, and if a channel is idle, the corresponding element in the state matrix is ​​set to 0.

[0008] Traverse the state matrix and calculate the noise matrix. Set the noise value for non-idle channels to 0. Calculate the SpRS, FWM, and ICXT noise photon counts for idle channels based on simulation parameters to generate a noise matrix corresponding to the state matrix. Simulation parameters include fiber parameters such as fiber length, single-core attenuation, and core spacing; signal parameters such as the classical signal power per channel; and single-photon detector (SPD) parameters such as dark counts, detection frequency, and detection efficiency.

[0009] Generate m groups of state matrices with different ratios of the total number of classical signals to the total number of channels (hereinafter referred to as channel occupancy). The number of state matrices in each group is n, and calculate the corresponding noise matrix. All elements in the state matrix are used as eigenvalues, and the corresponding elements of the noise matrix are used as target values ​​to generate a noise data set.

[0010] Step 2: Perform dimensionality reduction on the data set based on the noise characteristics of the quantum signal;

[0011] The dataset is preprocessed to improve the training speed and prediction accuracy of the noise prediction model. For each idle channel, SpRS and FWM are caused only by classical signals with the same fiber core but different wavelengths, while ICXT is caused only by classical signals with the same wavelength but different fiber cores. Therefore, for each channel, the number of channels related to its noise is C+W-1. The eigenvalue of each sample is the value of the element in the state matrix at that channel position in the same row but different columns, the same column but different rows, and the element of the sample itself.

[0012] Step 3: Establish a link noise prediction model based on XGBoost and LightGBM algorithms;

[0013] Select the ML algorithms for noise prediction: XGBoost and LightGBM algorithms, which are two gradient boosting decision tree frameworks with good performance. Among them, the XGBoost algorithm has a higher prediction accuracy and the LightGBM algorithm has a shorter training time. Call the MultiOutputRegressor method in the Python machine learning library sklearn to simultaneously predict the values ​​of all positions in the noise matrix. For each prediction model, define its training objective as regression. Gradually adjust the parameters such as the number of weak learners n_estimators and the decision tree depth max_depth of the model to normalize the mean square error (NMSE) and the coefficient of certainty (r 2 ) is the parameter adjustment indicator. The prediction accuracy under different parameters is evaluated through cross-validation. The model that takes into account both prediction accuracy and prediction time is selected. The data set described in step 2 is trained to obtain the noise prediction model.

[0014] NMSE is often used to evaluate the overall error between prediction results and measured results. The smaller the value, the smaller the error and the more reliable the prediction result. 2 It is obtained by calculating the predicted value and the measured value. It is an indicator used to evaluate the degree of fit. Its value range is between 0 and 1. The larger the value, the better the degree of fit. The expressions are as follows:

[0015]

[0016]

[0017] in is the predicted value, is the mean, n is the number of samples, and ||A||2 represents the 2-norm of A.

[0018] Step 4: Use the noise prediction model to predict link noise.

[0019] Based on the noise prediction model obtained in step 3, when a new service arrives, the current link state is read and used as input to obtain the link noise matrix, where the position with the lowest noise can be used to transmit quantum signals to achieve a higher key generation rate.

[0020] The method described in this patent can be applied to optical networks where quantum signals and classical signals are transmitted over multi-core optical fibers. It can reduce the delay introduced by the evaluation noise when new services arrive, and promote the application of integrated transmission of QKD and classical optical networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is the implementation process of the first embodiment of the present invention.

[0022] Figure 2 This is the result of the noise matrix obtained by calculation and the noise matrix obtained by using the noise prediction model for the same state matrix in the first embodiment of the present invention (the number of rows and columns has been adjusted because the noise matrix has too many columns). DETAILED DESCRIPTION

[0023] In order to make the purpose, technical methods and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0024] The method for predicting quantum channel noise in multi-core optical fiber based on data set dimensionality reduction proposed in this patent can use weakly coupled homogeneous seven-core optical fiber with 80 wavelengths per core. The implementation process of this method is as follows: Figure 1 The fiber channel state can be abstracted as a state matrix with 7 rows and 80 columns, where the elements of the channel state matrix corresponding to the channel occupied by the classical signal are set to 1, and the elements of the idle channel state matrix are set to 0.

[0025] Twenty sets of state matrices were generated, with channel occupancy uniformly distributed from 5% to 95%, each containing 50 state matrices. The corresponding noise matrices were calculated for each of these 10,000 state matrices. The noise parameters required were SpRS, FWM, and ICXT. The parameters of the fiber, signal, and single-photon detector in this transmission system are shown in Table 1. The state matrices and their corresponding noise matrices served as eigenvalues ​​and target values, respectively, to form the noise dataset.

[0026] Table 1. Simulation parameters

[0027] parameter Numerical Fiber length 10km Classical signal power per channel 0dBm Channel attenuation 0.2dB / km SPD dark count 700cps SPD detection efficiency 0.1 SPD detection frequency 1MHz

[0028] The data set is preprocessed. The shape of each state matrix is ​​originally 7×80. Since SpRS and FWM are only caused by classical signals with different wavelengths in the same fiber core, and ICXT is only caused by classical signals with the same wavelength in different fiber cores, only the same wavelength in the same fiber core, the same wavelength in different fiber cores, and the position of the required predicted channel itself are relevant to the noise results. The number of channels related to each channel for which noise needs to be predicted is 86, that is, the number of features of each sample is 86, and there are a total of 560 channels for which noise needs to be predicted. The number of samples in the noise data set is 10,000.

[0029] For each ML algorithm, in order to use only one model to evaluate noise, the present invention calls the MultiOutputRegressor class in the sklearn library to modify the base target regressor for predicting single channel noise. The ML algorithms used by the base target regressor are XGBoost and LightGBM, respectively, and their training objectives are defined as regression. NMSE and r are used. 2 As a performance indicator, considering the model prediction accuracy and prediction time, n_estimators is adjusted to 40. The model is used to train the noise dataset to obtain the final noise prediction model.

[0030] For this co-fiber transmission system, when new services arrive, the link state matrix is ​​used as the input of the prediction model, and the output is the link noise matrix. Figure 2 Figures 1 and 2 show the direct calculation and model prediction results for the noise on a link during a service arrival event. The top figure shows the directly calculated noise matrix, while the bottom figure shows the noise prediction results using the noise prediction model based on the XGBoost algorithm. Compared with directly calculating the noise, the noise evaluation delay can be reduced by up to 98.8% for different channel occupancy rates, and the average prediction accuracy reaches 96.25%.

[0031] It can be seen from the above examples that the quantum channel noise prediction method in multi-core optical fiber based on data set dimensionality reduction processing proposed in this patent can greatly reduce the delay introduced by link noise in the evaluation of common fiber transmission system while ensuring its prediction accuracy, and is applicable to transmission links with any number of fiber cores and any number of wavelengths, promoting the practical application of QKD and classical optical network fusion transmission.

Claims

1. A method for predicting quantum channel noise in a multi-core optical fiber based on data set dimensionality reduction processing, suitable for a multi-core optical fiber quantum key distribution network with any number of cores and any number of wavelengths, characterized in that: The steps include: S1. Establish a noise dataset for link noise characteristics in a multi-core fiber-based quantum key distribution network. This step establishes the noise dataset by using the state matrix describing the channel state as the eigenvalue and the channel noise as the target value. This allows the fiber, signal, and single-photon detector parameters required for noise prediction to be implicitly included in the noise prediction model rather than being used as model inputs, thereby reducing the complexity of using the noise prediction model. S2. Perform dimensionality reduction on the dataset based on the noise characteristics of the quantum signal. The main idea of ​​this step is to analyze the noise characteristics of the quantum channel in the multi-core optical fiber, remove noise-irrelevant features in the dataset, and improve the training speed and prediction accuracy of the noise prediction model. S3. Establish a noise prediction model based on the XGBoost and LightGBM algorithms. The two target regressors for predicting the noise of a single channel based on the XGBoost and LightGBM algorithms are denoted as base target regressors. The training objective of the base target regressor is regression. The base target regressor is modified using the MultiOutputRegrssor method in the Python machine learning library sklearn, so that the trained noise prediction model can simultaneously predict the optical noise of all channels. S4. Predict link noise using a noise prediction model.

2. As described in claim 1, a method for predicting quantum channel noise in a multi-core optical fiber based on data set dimensionality reduction processing is characterized in that: The process of establishing a noise dataset for link noise characteristics in a multi-core fiber-based quantum key splitting network includes: S11. Based on the number of cores C and the number of wavelengths W of the multi-core optical fiber, a state matrix of shape C×W is established, where the elements corresponding to the positions of the classical channels are 1 and the elements corresponding to the idle channels are 0; S12, traversing the state matrix, calculating the total noise photon number of spontaneous Raman scattering noise, four-wave mixing noise and inter-core crosstalk noise on each idle channel according to the optical fiber, signal and single-photon detector parameters, and setting the noise number on the non-idle channel to 0; S13. Generate m groups of state matrices with different ratios of the total number of classical signals to the total number of channels. The number of state matrices in each group is n. For each state matrix, generate a noise matrix as described in S12. Use the generated state matrix as the eigenvalue and the corresponding noise matrix as the target value to establish a noise data set.

3. As claimed in claim 1, a method for predicting quantum channel noise in a multi-core optical fiber based on data set dimensionality reduction processing is characterized in that: The dimensionality reduction process of the data set based on the noise characteristics of the quantum signal includes: For each idle channel, among the noises it is subject to, spontaneous Raman scattering noise and four-wave mixing noise are only caused by classical signals of the same fiber core but different wavelengths, and inter-core crosstalk noise is only caused by classical signals of the same wavelength in different fiber cores. Therefore, for each channel, the number of channels related to its noise is C+W-1, and the eigenvalue of each sample is the value of the element in the same row but different columns at the channel position in the state matrix, the value of the element in the same column but different rows, and the value of the sample itself.

4. As claimed in claim 1, a method for predicting quantum channel noise in a multi-core optical fiber based on data set dimensionality reduction processing, characterized in that: The process of building a link noise prediction model based on the XGBoost and LightGBM algorithms includes: S31. Select the machine learning algorithms for noise prediction: XGBoost and LightGBM. These are two gradient boosting decision tree frameworks with good performance. XGBoost has a higher prediction accuracy, while LightGBM takes less time to train. S32. By adjusting the number of weak learners and the depth of the decision tree, the model prediction accuracy is improved. In order to use a prediction model to simultaneously predict all elements of the noise matrix, the multi-output regression MultiOutputRegrssor method in the machine learning library sklearn in Python is used to modify the base target regressor.

Images