Estimation device, learning device, estimation method, and learning method

Abstract

This estimation device comprises: a filter coefficient estimation unit that uses a learning model, in which parameters have been updated to reduce the error between a correct-answer label and a first tap coefficient, to estimate a second tap coefficient on the basis of an input electric field waveform corresponding to a code sequence, the intensity of the input electric field at the input terminal of a transmission route, and a transmission distance; a first filter that generates a simulation signal of an output electric field waveform at the output terminal of the transmission route, on the basis of the input electric field waveform corresponding to the code sequence and the estimated second tap coefficient; and an error rate estimation unit that estimates the error rate of the code sequence at the output terminal of the transmission route, on the basis of the output electric field waveform simulation signal.

Description

[Technical Field]

[0001] The present invention relates to an estimation device, a learning device, an estimation method, and a learning method. [Background technology]

[0002] In an optical communication system using ROADM (reconfigurable optical add / drop multiplexer) (see Non-Patent Document 1), each node in the transmission path forwards the optical signal transmitted from the first communication device to the second communication device using an optical path. Furthermore, in an All Photonics Network (APN), the optical paths are connected end-to-end without any photoelectric conversion being performed on the optical signals.

[0003] Figure 12 shows an example of the configuration of an optical communication system. The optical communication system illustrated in Figure 12 comprises a first communication device, a second communication device, and a transmission line. The transmission line illustrated in Figure 12 comprises a first node, a second node, a third node, a fourth node, and a fifth node. Each node also includes an optical switch (not shown).

[0004] Each node transmits the optical signal without performing photoelectric conversion. As a result, the optical signal transmitted from the first communication device is transmitted to the second communication device while remaining as light (electric field waveform).

[0005] In an all-photonics network, when a signal requesting connection to a second communication device is transmitted from the first communication device, an appropriate optical path from the first communication device to the second communication device is selected from among several optical paths in the transmission line.

[0006] Here, since the modulation method of the optical signal, the transmission rate, the distance (length of the transmission path), the type of optical fiber used in the transmission path, and the gain of the optical amplifier through which the optical signal is transmitted differ for each optical path, the code error rate in the second communication device also differs for each optical path. For this reason, it is necessary to select an optical path capable of error-free transmission (an optical path in which the code error rate is less than a predetermined value) from among multiple optical paths.

[0007] The optical paths may be selected based on the results of verifying whether each optical path is error-free by actually transmitting an optical signal through it. However, in this case, it takes time for multiple optical paths to be covered, so an enormous amount of time is required to activate the optical paths. Therefore, in order to activate the optical paths in a short time, it is effective to select the optical path that is capable of error-free transmission based on the pre-estimated code error rate for each optical path.

[0008] One method for estimating the code error rate in optical communication systems is to have an estimation device estimate the code error rate through propagation simulation. For example, the estimation device estimates the electric field waveform of an optical signal transmitted through a transmission path having an optical fiber using propagation simulation. The estimation device simulates the electric field waveform in a second communication device by adding appropriate noise to the electric field waveform. The estimation device identifies the code sequence received by the second communication device (received code sequence) by performing a threshold judgment process on the simulation result. The estimation device estimates the code error rate based on the difference between the identified received code sequence and the code sequence transmitted from the first communication device (transmitted code sequence).

[0009] The estimation device accurately estimates the change in the electric field waveform of an optical signal transmitted through a transmission path by performing a predetermined algorithmic processing (e.g., the Split Step Fourier Method (SSFM)) on the nonlinear Schrödinger equation. In this case, the estimation device estimates the electric field waveform (simulated signal) that has undergone linear changes (wavelength dispersion) and nonlinear changes (self-phase modulation) due to transmission. This makes it possible to accurately estimate the change in the electric field waveform of an optical signal transmitted through a transmission path even when nonlinear waveform distortion occurs in the electric field waveform.

[0010] However, in methods based on the split-step Fourier method, the fiber section transmitting the optical signal is divided, and the calculation of the electric field waveform for each divided fiber section is repeated sequentially. Therefore, the calculation time increases as the transmission distance increases. Consequently, it is difficult to apply this estimation method to all-photonic networks that require real-time operation.

[0011] To solve these problems, "GNPy" is provided as a method for estimating the code error rate in a short time (see Non-Patent Document 2). In "GNPy", the nonlinear change is approximated by random Gaussian noise (see Non-Patent Documents 3 and 4), which allows for the estimation of the nonlinear change in a short time. [Prior art documents] [Non-patent literature]

[0012] [Non-Patent Document 1] M. Birk et al., “The OpenROADM initiative [Invited]”, Journal of Optical Communications and Networking, vol.12, no.6, pp.C58-67, June 2020. [Non-Patent Document 2] A. Ferrari et al., “GNPy: an open source application for physical layer aware open optical networks”, Journal of Optical Communications and Networking, vol.12, no.6, pp.C31-C40, June 2020. [Non-Patent Document 3] P. Poggiolini et al., “A Detailed Analytical Derivation of the GN Model of Non-Linear Interference in Coherent Optical Transmission Systems” [Non-Patent Document 4] P. Poggiolini et al., “The GN Model of Non-Linear Propagation in Uncompensated Coherent Optical Systems”, Journal of Lightwave Technology, vol.30, no.24, pp.3857-3879, DECEMBER 2012. [Overview of the project] [Problems that the invention aims to solve]

[0013] However, the Gaussian noise approximation for nonlinear changes only holds true when the transmission distance of the optical signal is greater than or equal to a predetermined distance. Therefore, when the transmission distance of the optical signal is less than the predetermined distance, the Gaussian noise approximation for nonlinear changes cannot be applied. Consequently, for example, the Gaussian noise approximation for nonlinear changes cannot be applied to communications within a data center where the transmission distance of the optical signal is relatively short.

[0014] Thus, when the transmission distance of an optical signal is less than a predetermined distance, there is a problem in that the code error rate cannot be estimated in a short time based on the nonlinearly changing electric field waveform.

[0015] In view of the above circumstances, the present invention aims to provide an estimation device, a learning device, an estimation method, and a learning method that can improve the accuracy of estimating the code error rate in a short time based on a nonlinearly changing electric field waveform, even when the transmission distance of an optical signal is less than a predetermined distance. [Means for solving the problem]

[0016] One aspect of the present invention is an estimation device comprising: a filter coefficient estimation unit that estimates a second tap coefficient based on an input electric field waveform corresponding to a code sequence, the light intensity of the input electric field waveform at the input end of a transmission line, and the transmission distance, using a learning model whose parameters have been updated to reduce the error between the correct label and a first tap coefficient; a first filter that generates a simulated signal of the output electric field waveform at the output end of the transmission line based on the input electric field waveform corresponding to the code sequence and the estimated second tap coefficient; and an error rate estimation unit that estimates the error rate of the code sequence at the output end of the transmission line based on the simulated signal of the output electric field waveform.

[0017] One aspect of the present invention is a learning device comprising: a learning unit that generates a first tap coefficient using a learning model based on the input electric field waveform at the input end of a transmission line, the light intensity of the input electric field waveform, and the transmission distance; and an error calculation unit that obtains a correct label generated based on the output electric field waveform at the output end of the transmission line and the input electric field waveform, and calculates the error between the correct label and the first tap coefficient, wherein the learning unit updates the parameters of the learning model to reduce the error.

[0018] One aspect of the present invention is an estimation method performed by an estimation device, comprising the steps of: estimating a second tap coefficient based on an input electric field waveform corresponding to a code sequence and the light intensity and transmission distance of the input electric field waveform at the input end of a transmission line, using a learning model whose parameters have been updated to reduce the error between a correct label and a first tap coefficient; generating a simulated signal of the output electric field waveform at the output end of the transmission line based on the input electric field waveform corresponding to the code sequence and the estimated second tap coefficient; and estimating the error rate of the code sequence at the output end of the transmission line based on the simulated signal of the output electric field waveform.

[0019] One aspect of the present invention is a learning method performed by a learning device, comprising the steps of: generating a first tap coefficient using a learning model based on an input electric field waveform at the input end of a transmission line, the light intensity of the input electric field waveform, and the transmission distance; and an error calculation unit that obtains a correct label generated based on an output electric field waveform at the output end of the transmission line and the input electric field waveform, and calculates an error between the correct label and the first tap coefficient, wherein the calculation step includes updating the parameters of the learning model to reduce the error. [Effects of the Invention]

[0020] The present invention makes it possible to improve the accuracy of estimating the code error rate in a short time based on a nonlinearly changing electric field waveform, even when the transmission distance of the optical signal is less than a predetermined distance. [Brief explanation of the drawing]

[0021] [Figure 1] This figure shows an example of the configuration of the estimation system in the first embodiment. [Figure 2] This figure shows an example of the configuration of an optical communication system in the first embodiment. [Figure 3] This figure shows an example of the configuration of the filter generating device in the first embodiment. [Figure 4] This figure shows an example of a lookup table in the first embodiment. [Figure 5] This figure shows an example of the configuration of the learning device in the first embodiment. [Figure 6] This is a flowchart showing an example of the operation of the learning device in the first embodiment. [Figure 7] This is a flowchart showing an example of the operation of the estimation device in the first embodiment. [Figure 8] This figure shows an example of the configuration of the estimation system in the second embodiment. [Figure 9] This figure shows an example of the configuration of the feature generation unit in the second embodiment. [Figure 10] This figure shows an example of the configuration of the learning device in the second embodiment. [Figure 11] This figure shows an example of the hardware configuration of the estimation system in each embodiment. [Figure 12] This is a diagram showing an example configuration of an optical communication system. [Modes for carrying out the invention]

[0022] Embodiments of the present invention will be described in detail with reference to the drawings. (First Embodiment) Figure 1 shows an example configuration of the estimation system 1a in the first embodiment. The estimation system 1a is a system that estimates the code error rate in real time (within a predetermined delay time) according to the change in the electric field waveform of an optical signal transmitted through the transmission path of an optical communication system. In other words, the estimation system 1a calculates the electric field waveform when it arrives at the second communication device based on the optical signal corresponding to the transmitted code sequence transmitted from the first communication device, and estimates the code error rate in the second communication device in real time based on the calculated electric field waveform.

[0023] Here, when an optical signal corresponding to the transmitted code sequence (input optical signal) is transmitted to the transmission line, waveform distortion (linear waveform distortion and nonlinear waveform distortion) occurs in the input optical signal waveform due to the effects of linear and nonlinear changes caused by propagation in the transmission line. The output optical signal (hereinafter referred to as "output electric field waveform") containing this waveform distortion is output from the output end of the transmission line to the receiving communication device. Furthermore, the effect of self-phase modulation, which is one of the nonlinear changes, is uniquely determined according to a predetermined set of parameters. These predetermined parameters are, for example, the electric field waveform of the optical signal (main signal) input to the transmission line, its intensity (optical intensity), and the transmission distance of the optical signal. A linear change is, for example, a change due to wavelength dispersion.

[0024] Figure 2 shows an example configuration of the optical communication system 100 in the first embodiment. The optical communication system 100 is a system that communicates using optical signals. The optical communication system 100 comprises one or more first communication devices 110, a second communication device 120, and N (where "N" is an integer of 1 or more) transmission lines 130. The span between communication devices in the optical communication system 100 may be single-span (N=1) or multi-span (N≧2). The optical communication system 100 may also be equipped with optical amplifiers 140 for each span.

[0025] Hereinafter, the electric field waveform of the optical signal at the input terminal of a transmission line, etc., will be referred to as the "input electric field waveform". Symbol "E in " represents the input electric field waveform. Symbol "E out " represents the output electric field waveform. Input electric field waveform "E in " and output electric field waveform "E out These are all time waveforms. The symbol "P in " is the light intensity of the input electric field waveform "|E in | 2 This represents the time average of "<>".

[0026] The first communication device 110 (first user terminal) transmits an optical signal corresponding to the transmission code sequence to the second communication device 120 (second user terminal) using an optical path in a transmission line 130 (for example, an optical fiber) with a transmission distance "L". Here, the input electric field waveform of the optical signal "Ein is input to transmission line 130. The optical amplifier 140 amplifies the power of the output electric field waveform “E out ” that has linearly and non-linearly changed in transmission line 130. The second communication device 120 (second user terminal) acquires the output electric field waveform “E out ” that has linearly and non-linearly changed in transmission line 130.

[0027] Returning to FIG. 1, the description of the configuration example of the estimation system 1a is continued. The estimation system 1a includes a filter generation device 2a, a storage device 3, and a learning device 9a as a model generation device 10a. The estimation system 1a includes a sequence generation device 4 and an electric field generation device 5. The estimation system 1a includes electric field estimation devices 6a-n (“n” is an integer from 1 to N) and an error rate estimation device 7a as an estimation device 8a. That is, the estimation device 8a includes an electric field estimation device 6a and an error rate estimation device 7a. The estimation system 1a includes a learning device 9a.

[0028] The electric field estimation device 6a (electric field estimation unit) includes a filter coefficient estimation unit 61 and a filter 62. The error rate estimation device 7a includes a photoelectric conversion unit 71, a noise processing unit 72, a determination unit 73, and an error rate estimation unit 74.

[0029] The sequence generation device 4 pre-generates a transmission code sequence transmitted from the first communication device 110 (transmission-side communication device) toward the second communication device 120 (reception-side communication device). The sequence generation device 4 transmits the generated transmission code sequence to the electric field generation device 5. The electric field generation device 5 generates an input electric field waveform “E in ” of an optical signal corresponding to the transmission code sequence based on the characteristics of the first communication device 110 (for example, modulation method and transmission characteristics).

[0030] The electric field generation device 5 transmits the generated input electric field waveform “E in ” to the electric field estimation device 6a. Here, the electric field estimation device 6a estimates the output electric field waveform “E in ” by applying a filter 62 having the characteristics of the transmission line 130 to the input electric field waveform “E out ”.

[0031] The filter coefficient estimation unit 61 estimates tap coefficients according to the waveform characteristics of linear and nonlinear changes when an optical signal is transmitted through a transmission path 130 with a transmission distance "L", based on the input electric field waveform "E" at the input end of the transmission path 130. in " and the light intensity of the input electric field waveform "P in Based on the transmission distance "L", the tap coefficient is estimated using a trained model. The trained model has, for example, a neural network used in deep learning. The neural network is, for example, a convolutional neural network. The filter coefficient estimation unit 61 sets the tap coefficient estimated using the trained model into the filter 62.

[0032] Filter 62 is, for example, a Volterra filter (Reference 1: NP. Diamantopoulos et al., “On the Complexity Reduction of the Second-Order Volterra Nonlinear Equalizer for IM / DD Systems”, Journal of Lightwave Technology, vol. 37, no. 4, pp.1214-1224, FEBRUARY 15, 2019). Filter 62 may also be, for example, a finite-time impulse response filter (FIR filter), which is a first-order Volterra filter.

[0033] The filter 62 receives the input electric field waveform "E" from the transmission path 130 with a transmission distance "L". in The output electric field waveform "E" is affected by the linear and nonlinear waveform characteristics when "E" is transmitted. out This is output to the next stage electric field estimation device or error rate estimation device 7a.

[0034] The electric field estimation device 6a outputs the electric field waveform "E outThis is transmitted to the error rate estimation device 7a. The photoelectric conversion unit 71 converts the output electric field waveform into an electrical waveform. The photoelectric conversion unit 71 (detection unit) may be, for example, a direct detection type receiver (e.g., a single photodiode) or a coherent receiver.

[0035] The noise processing unit 72 adds a predetermined appropriate noise to the electrical signal. The predetermined appropriate noise is, for example, Gaussian noise (Reference 2: W. Freude et al., “Quality Metrics for Optical Signals: Eye Diagram, Q-factor, OSNR, EVM and BER”, Mo.B1.5, ICTON 2012). Examples of Gaussian noise include thermal noise in the receiving communication device and noise due to spontaneous emission (ASE).

[0036] The determination unit 73 identifies the received code sequence in the output electric field waveform received by the photoelectric conversion unit 71 by performing a threshold determination process on the noisy electrical signal. The error rate estimation unit 74 estimates the code error rate in the second communication device 120 based on the difference between the identified received code sequence and the transmitted code sequence transmitted from the first communication device 110.

[0037] The learning model used by the filter coefficient estimation unit 61 for estimating the tap coefficients is pre-generated by the model generation device 10a. As described above, the learning model is based on the input electric field waveform "E in " and the light intensity of the input electric field waveform "P in The tap coefficients of the filter 62 that gives transmission characteristics to the input electric field waveform in the electric field estimation device 6a are output using the input electric field waveform "E" and the transmission distance "L" as explanatory variables. In other words, in order to train this learning model, a dataset of four types of parameters is required. These four types of parameters are the input electric field waveform "E" in " and the light intensity of the input electric field waveform "P in These are the explanatory variables " and the transmission distance "L", and the tap coefficient of filter 62 (the dependent variable).

[0038] The filter generation device 2a calculates the tap coefficients of the filter section 23 (corresponding to the tap coefficients of filter 62) for each given explanatory variable, which have appropriate transmission path characteristics. The filter generation device 2a records the tap coefficients of the filter calculated for each explanatory variable as a dataset in the storage device 3. The learning device 9a can read the tap coefficients of the filter calculated for each explanatory variable from the storage device 3 and perform training on the learning model.

[0039] The filter generation device 2a (filter generation unit) generates the output electric field waveform "E" at the output terminal of the transmission line 130. out-m The shape of " and the output electric field waveform "E" at the input terminal of the transmission line 130. in-m The output electric field waveform of the filter of the filter generation device 2a, which receives the input "E" out-m To minimize the difference between the shape of '' and the transmission distance "L m " and the light intensity of the input electric field waveform "P in-m For each combination "m" with "", the filter generation device 2a generates the tap coefficients of the filter (corresponding to the tap coefficients of filter 62). The filter generation device 2a generates the tap coefficients of filter 62 and the input electric field waveform "E in " and the light intensity of the input electric field waveform "P in The transmission distance "L" is recorded in the memory device 3.

[0040] The learning device 9a generates, for example, the tap coefficients (target variable) of a Voltera filter as an output of the learning model by inputting explanatory variables into the learning model. The learning device 9a updates the parameters of the neural network of the learning model using machine learning techniques. The learning device 9a outputs the learning model (trained model) with the updated neural network parameters to the filter coefficient estimation unit 61.

[0041] Figure 3 shows an example of the configuration of the filter generation device 2a in the first embodiment. The filter generation device 2a comprises a delay processing unit 21, an error calculation unit 22, and a filter unit 23 (filter processing unit).

[0042] The delay processing unit 21 outputs the electric field waveform "E out This output electric field waveform "E out The waveform may be one actually acquired by coherent reception or a received power receiver (PR receiver), or it may be a waveform generated using high-precision waveform simulation. High-precision waveform simulation refers to, for example, a waveform simulation using the split-step Fourier method.

[0043] The delay processing unit 21 outputs the electric field waveform "E out If a predetermined time has elapsed since the acquisition time of ", the output electric field waveform "E out The output "E" is sent to the error calculation unit 22. This results in the delayed output electric field waveform "E out " and the output electric field waveform "E" output from the filter section 23 out The error calculation unit 22 synchronizes with the '''.

[0044] The error calculation unit 22 calculates the output electric field waveform "E" output from the delay processing unit 21. out Output electric field waveform "E" for the shape of "" out Error in the shape of '" e=|E out -E out The error calculation unit 22 calculates the error "e" and feeds it back to the filter unit 23.

[0045] The filter section 23 filters the input electric field waveform "E in The filter unit 23 obtains the error "e" from the error calculation unit 22. The filter unit 23 uses a Voltera filter with initial tap coefficients to obtain the input electric field waveform "E". in By using it for the input electric field waveform, the light intensity "P inm =<|E inm | 2 For each combination of ">" and transmission distance "L", the output electric field waveform "E out Generates ''.

[0046] The filter section 23 filters the light intensity "P" of the input electric field waveform. inFor each combination of " and the transmission distance "L", an appropriate tap coefficient is generated as the correct label. The filter unit 23 updates the tap coefficient of the Voltera filter of the filter unit 23 using a predetermined algorithm so that the error "e=0". The predetermined algorithm is, for example, the Least Mean Square (LMS) algorithm.

[0047] As a result, the filter unit 23 sets tap coefficients according to the waveform characteristics of linear and nonlinear changes in the transmission line 130, and the optical intensity "P" of the input electric field waveform. in It is generated for each combination of the input electric field waveform "E" and the transmission distance "L". The filter section 23 generates the input electric field waveform "E" in " and the transmission distance "L" and the light intensity of the input electric field waveform "P in The generated tap coefficients are then registered in the lookup table. The generated tap coefficients are used as the correct labels for supervised learning in the machine learning process performed by the learning device 9a.

[0048] Figure 4 shows an example of a lookup table in the first embodiment. In the stage where tap coefficients used as the correct labels for supervised learning are generated (correct label generation stage), the lookup table is used to determine the light intensity "P" of the input electric field waveform. in-m " and transmission distance "L m A lookup table is generated for each combination of "m". The lookup table contains the tap coefficients (correct labels) generated by the filter unit 23, and the light intensity "P" of the input electric field waveform. in-m " and transmission distance "L m Each combination with "m" is registered.

[0049] Figure 5 shows an example of the configuration of the learning device 9a in the first embodiment. The learning device 9a comprises a learning unit 91a and an error calculation unit 92. The learning unit 91a calculates the input electric field waveform "E" at the input terminal of the transmission line 130. in " and the light intensity of the input electric field waveform "P inThe learning unit 91a obtains the transmission distance "L" and the explanatory variables. The learning unit 91a inputs the explanatory variables into the learning model and generates, for example, the tap coefficients (target variable) of the Voltera filter as the output of the learning model. The learning unit 91a updates the parameters of the learning model to reduce the error calculated by the error calculation unit 92 using machine learning methods (for example, supervised learning and backpropagation).

[0050] The error calculation unit 92 calculates the output electric field waveform "E" at the output terminal of the transmission line 130. out " and input electric field waveform "E in The correct labels generated based on this are obtained from the filter generation device 2a via the storage device 3. The error calculation unit 92 calculates the error between the tap coefficients generated by the learning unit 91a and the correct labels.

[0051] Next, we will explain an example of the operation of the learning device 9a. Figure 6 is a flowchart showing an example of the operation of the learning device 9a in the first embodiment. During the learning process, the learning unit 91a calculates the explanatory variable (the input electric field waveform "E" at the input terminal of the transmission line 130). in ", the light intensity of the input electric field waveform "P in Based on the transmission distance "L" and the learning model, the first tap coefficient (target variable) is calculated (step S101).

[0052] The error calculation unit 92 calculates the output electric field waveform "E" at the output terminal of the transmission line 130. out " and input electric field waveform "E in The correct labels generated based on this are obtained from the filter generation device 2a via the storage device 3 (step S102). The error calculation unit 92 calculates the error between the correct labels and the first tap coefficient (step S103). The learning unit 91a updates the parameters of the learning model to reduce the error (step S104).

[0053] Next, we will explain an example of the operation of the estimation device 8a. Figure 7 is a flowchart showing an example of the operation of the estimation device 8a in the first embodiment. In the stage where the error rate estimation process is performed (estimation processing stage), the filter coefficient estimation unit 61 obtains a trained model from the learning device 9a in which the parameters have been updated to reduce the error between the correct label and the first tap coefficient (step S201). The filter coefficient estimation unit 61 calculates the input electric field waveform "E" according to the transmitted code sequence. in " and the optical intensity of the input electric field waveform at the input terminal of the transmission line 130 "P in Based on the transmission distance "L", the second tap coefficient is estimated using the trained model (step S202).

[0054] The filter 62, with the second tap coefficient set, outputs the input electric field waveform "E" according to the transmitted code sequence. in The signal "E" is obtained from the preceding electric field estimation device or electric field generation device 5 (step S203). The filter 62, with the second tap coefficient set, simulates the output electric field waveform at the output terminal of the transmission line 130. out The acquired input electric field waveform "E in The error rate estimation unit 74 generates the simulated signal "E" of the output electric field waveform (step S204). out Based on this, the error rate of the transmitted code sequence (received code sequence in the second communication device 120) at the output terminal of the transmission line 130 is estimated (step S205).

[0055] As described above, during the learning phase, the learning unit 91a determines the input electric field waveform "E" at the input terminal of the transmission line 130. in " and the light intensity of the input electric field waveform "P in Based on the output electric field waveform "E" at the output end of the transmission line 130, the learning model is used to generate tap coefficients. out " and input electric field waveform "E inThe correct labels generated based on this are obtained from the filter generation device 2a via the storage device 3. The error calculation unit 92 calculates the error between the correct labels and the tap coefficients. The learning unit 91a updates the parameters of the learning model to reduce the error using machine learning methods (e.g., supervised learning and backpropagation).

[0056] Furthermore, during the estimation stage, the filter coefficient estimation unit 61 acquires a learned model (trained model) from the learning device 9a, in which the parameters have been updated to reduce the error between the correct label and the first tap coefficient. Using the acquired learned model, the filter coefficient estimation unit 61 calculates the input electric field waveform "E" according to the code sequence. in "and the optical intensity of the input electric field waveform at the input terminal of the transmission line 130 "P in Based on the transmission distance "L", the second tap coefficient is estimated. Filter 62 (first filter) calculates the input electric field waveform "E" according to the code sequence. in Based on the estimated second tap coefficient, the simulated signal "E" of the output electric field waveform at the output terminal of the transmission line 130 is generated. out The error rate estimation unit 74 generates a simulated signal of the output electric field waveform "E out Based on this, the error rate of the transmitted code sequence (received code sequence in the second communication device 120) at the output terminal of the transmission line 130 is estimated.

[0057] This makes it possible to improve the accuracy of quickly estimating the code error rate based on a nonlinearly changing electric field waveform, even when the transmission distance of the optical signal is less than a predetermined distance.

[0058] In the first embodiment, the filter coefficient estimation unit 61 estimates the input electric field waveform "E in " and the light intensity of the input electric field waveform "P in Based on the input electric field waveform "E" and the transmission distance "L", the tap coefficients of filter 62 are estimated using machine learning techniques. On the other hand, without using machine learning techniques, the input electric field waveform "E" is used. in " and the light intensity of the input electric field waveform "P inBased on a combination that approximates the combination of the input electric field waveform "E" and the transmission distance "L", a tap coefficient registered in the lookup table may be selected, and the selected tap coefficient may be assigned to the filter 62. in " and the light intensity of the input electric field waveform "P in This is effective when there are few combinations of " " and the transmission distance "L".

[0059] However, in order to improve estimation accuracy, the input electric field waveform "E" should be used in the lookup table. in " and the light intensity of the input electric field waveform "P in The combination of "" and the transmission distance "L" needs to be calculated with fine granularity. In such cases, the time required for preparation and the increased memory capacity of the lookup table are concerns. For example, even if the span of the optical communication system 100 is single span, the more types of communication devices there are on the transmitting side, the more the input electric field waveform "E" needs to be calculated. in The pattern of "" becomes more frequent. Also, if the span of the optical communication system 100 is multi-span, the input electric field waveform "E" will vary depending on the number of spans. in The pattern of "[...]" becomes more frequent. From the above perspective, this method makes it possible to reduce the amount of data acquired in advance for generating the lookup table and the capacity of the lookup table, while improving the accuracy of estimating the coding error rate in a short time.

[0060] (Second Embodiment) In the second embodiment, the main difference from the first embodiment is that the first feature generation unit is provided by the electric field estimation device, and the second feature generation unit is provided by the learning device. The second embodiment will be explained focusing on the differences from the first embodiment.

[0061] For example, in Figure 5 illustrated in the first embodiment, the input electric field waveform "E" is 10,000 bits. in-n When " is input to the learning unit 91a, for example, if 4x oversampling is performed, a 400,000-bit input electric field waveform "E in-n This is input into the learning model. As you can see, the computational load of machine learning becomes enormous.

[0062] Therefore, in the second embodiment, the feature amount of the input electric field waveform is extracted from the input electric field waveform "E in-n " in the form of the filter coefficient. The number of bits of the feature amount of the input electric field waveform is less than the number of bits of the input electric field waveform. The feature amount extracted in this way is input to the learning model. Thereby, the calculation amount of machine learning is reduced.

[0063] FIG. 8 is a diagram showing a configuration example of the estimation system 1b in the second embodiment. The estimation system 1b includes a filter generation device 2b, a storage device 3, a learning device 9b, and a feature amount generation unit 63-1 as a model generation device 10b. The estimation system 1b includes a sequence generation device 4 and an electric field generation device 5. The estimation system 1b includes an electric field estimation device 6b-n and an error rate estimation device 7b as an estimation device 8b. The electric field estimation device 6b (electric field estimation unit) includes a filter coefficient estimation unit 61, a filter 62, and a feature amount generation unit 63-2.

[0064] FIG. 9 is a diagram showing a configuration example of the feature amount generation unit 63 in the second embodiment. The feature amount generation unit 63 includes a sequence generation device 41, an electric field generation device 51, a delay processing unit 21, an error calculation unit 22, and a filter unit 64 (filter processing unit).

[0065] The sequence generation device 41 pre-generates a transmission code sequence transmitted from the first communication device 110 (transmission-side communication device) toward the second communication device 120 (reception-side communication device). The sequence generation device 41 transmits an electric signal corresponding to the generated transmission code sequence to the electric field generation device 51.

[0066] The electric field generation device 51 plays a role of inputting an input electric field waveform similar to the input electric field waveform from the first communication device 110 into the filter unit 64. The electric field generation device 51 generates a reference electric field waveform "E0" based on the characteristics of the first communication device 110 (for example, modulation method and transmission characteristics). The electric field generation device 51 outputs the reference electric field waveform "E0" to the filter unit 64. Here, it is not necessary to generate an input electric field waveform for each characteristic of the first communication device 110. By outputting the reference electric field waveform "E0" common to the multiple characteristics of one or more first communication devices 110 to the filter unit 64, appropriate coefficients corresponding to the multiple characteristics of one or more first communication devices 110 are output from one learning model (trained model).

[0067] The delay processing unit 21 of the feature quantity generation unit 63 synchronizes the input electric field waveform "E in-n " with the input electric field waveform "E in-n " by applying a predetermined delay to the input electric field waveform "E in-n ". On the other hand, the filter unit 64 outputs the output electric field waveform "E in-n '" generated by applying a predetermined transfer function to the input electric field waveform "E0" to the error calculation unit 22 of the feature quantity generation unit 63.

[0068] The error calculation unit 22 calculates the error "e = |E in-n - E in-n '|" of the shape of the input electric field waveform "E in-n [[ID=२२]]' with respect to the shape of the input electric field waveform "E in-n " output from the delay processing unit 21 of the feature quantity generation unit 63. The error calculation unit 22 feeds back the error "e" to the filter unit 64.

[0069] The filter unit 64 has a filter with predetermined initial tap coefficients. The tap coefficients of the filter unit 64 (e.g., a Voltera filter) are updated using a predetermined algorithm so that the error "e=0". The predetermined algorithm is, for example, the least squares mean algorithm. This allows the filter unit 64 to create a filter that contains information about the difference from the reference electric field waveform "E0". The filter unit 64 of the feature quantity generation unit 63 of the electric field estimation device 6b updates the tap coefficients of this filter to the input electric field waveform "E in-n The input electric field waveform "E" is output to the filter coefficient estimation unit 61 as a feature quantity. in-n The tap coefficients, which are features of the filter coefficient estimation unit 61, are set in the filter 62 of the filter coefficient estimation unit 61. Therefore, the filter of the filter unit 64 corresponds to the filter 62 of the filter coefficient estimation unit 61.

[0070] Thus, even when the characteristics (e.g., chirp amount, signal bandwidth) of one or more first communication devices 110 (transmitting communication devices) differ, the electric field generator 51 deliberately uses a common reference electric field waveform "E0" to cause the filter unit 64 to generate feature quantities. The characteristics of each first communication device 110 (the difference between the input electric field waveform from the first communication device 110 and the reference electric field waveform "E0") are reflected in the tap coefficients of the filter 62. In this way, not only information on waveform distortion caused in the electric field waveform by the transmission line, but also information on the characteristics of each first communication device 110 is reflected in the filter coefficients. As a result, appropriate coefficients corresponding to the multiple characteristics of one or more first communication devices 110 are output.

[0071] Figure 10 shows an example of the configuration of the learning device 9b in the second embodiment. The learning device 9b comprises a learning unit 91b and an error calculation unit 92. The feature quantity generation unit 63-1 generates an input electric field waveform "E in-m The learning unit 91b generates the feature quantities of the input electric field waveform "E". in-m The feature quantities of "P" and the light intensity of the input electric field waveform in The input electric field waveform "E" and the transmission distance "L" are obtained as explanatory variables. Here, the learning unit 91b obtains the input electric field waveform "E" in-mFor example, the feature quantities of the input electric field waveform "E" in-m The input " is obtained from the feature generation unit 63.

[0072] The learning unit 91b inputs explanatory variables into the learning model and generates, for example, the tap coefficients (target variable) of a Voltera filter as the output of the learning model. The learning unit 91b updates the parameters of the learning model to reduce the error calculated by the error calculation unit 92 using machine learning methods (e.g., supervised learning and backpropagation). The learning unit 91b records the learning model (trained machine learning model) with updated neural network parameters in the storage device 3.

[0073] The error calculation unit 92 calculates the output electric field waveform "E" at the output terminal of the transmission line 130. out " and input electric field waveform "E in The correct labels generated based on this are obtained from the filter generation device 2b via the storage device 3. The error calculation unit 92 calculates the error between the tap coefficients generated by the learning unit 91b and the correct labels.

[0074] As described above, during the learning phase, the feature generation unit 63-1 processes the reference electric field waveform "E" that has passed through the filter unit 64. in-n '" and input electric field waveform "E in-n The tap coefficient of the filter section 64 (the second tap coefficient in the learning device 9b) is set so that it matches the input electric field waveform "E" (to minimize the error). in-n The learning unit 91b generates the input electric field waveform "E" at the input terminal of the transmission line 130. in-n The feature quantity of "" (second tap coefficient in learning device 9b) and the light intensity "P" of the input electric field waveform. in-n " and the transmission distance "L n Based on this, a learning model is used to generate tap coefficients (first tap coefficients). The error calculation unit 92 calculates the output electric field waveform "E" at the output terminal of the transmission line 130. out-n " and input electric field waveform "E in-nThe correct labels generated based on this are obtained from the filter generation device 2a via the storage device 3. The error calculation unit 92 calculates the error between the correct labels and the tap coefficients (first tap coefficients). The learning unit 91b updates the parameters of the learning model to reduce the error using machine learning methods (e.g., supervised learning and backpropagation).

[0075] Furthermore, during the estimation stage, the feature generation unit 63-2 generates a reference electric field waveform "E" that has passed through the filter unit 64 (second filter). in-n '" and input electric field waveform "E in-n The tap coefficient of the filter section 64 (the third tap coefficient in the estimation device 8b) is set so that it matches the input electric field waveform "E" (to minimize the error). in-n The filter coefficient estimation unit 61 generates the input electric field waveform "E" according to the code sequence. The filter coefficient estimation unit 61 obtains a trained model (trained model) from the learning device 9b, in which the parameters have been updated to reduce the error between the correct label and the first tap coefficient. The filter coefficient estimation unit 61 uses the obtained trained model (trained model) to generate the input electric field waveform "E" according to the code sequence. in-n The characteristic quantity of "" (third tap coefficient in estimation device 8b) and the optical intensity of the input electric field waveform at the input terminal of the transmission line 130 "P in-n " and transmission distance "L n Based on this, the second tap coefficient in the estimation device 8b is estimated. Filter 62 (first filter) calculates the input electric field waveform "E" according to the code sequence. in-n Based on the estimated second tap coefficient, the simulated signal "E" of the output electric field waveform at the output terminal of the transmission line 130 is generated. out-n The error rate estimation unit 74 generates the simulated signal of the output electric field waveform "E out-n Based on this, the error rate of the code sequence at the output terminal of the transmission line 130 is estimated.

[0076] This improves the convergence characteristics of the least-squares mean in the tap coefficient derivation process, even when the optical signal transmission distance is less than a predetermined distance, and enhances the accuracy of estimating the code error rate in a short time based on the nonlinearly changing electric field waveform. Furthermore, it reduces the computational load of machine learning.

[0077] (Hardware configuration) Figure 11 shows examples of the hardware configuration of the estimation system in each embodiment. The estimation system 1 illustrated in Figure 11 corresponds to the estimation system 1a in the first embodiment and the estimation system 1b in the second embodiment, respectively.

[0078] The estimation system 1 is implemented as software by a processor 101, such as a CPU (Central Processing Unit), executing a program stored in a storage device 103 having a non-volatile recording medium (non-temporary recording medium) and memory 102. For example, in each embodiment, the processor 101 performs a simulation to estimate waveform distortion occurring in an optical signal waveform using a predetermined algorithmic processing for a nonlinear or linear Schrödinger equation. The program may be recorded on a computer-readable recording medium. A computer-readable recording medium is a non-temporary recording medium such as a portable medium such as a flexible disk, magneto-optical disk, ROM (Read Only Memory), CD-ROM (Compact Disc Read Only Memory), or a storage device such as a hard disk or solid-state drive (SSD) built into a computer system. The communication unit 104 performs predetermined communication processing.

[0079] The estimation system 1 may be implemented using hardware (accelerator) including an electronic circuit (or circuitry) such as an LSI (Large Scale Integrated Circuit), ASIC (Application Specific Integrated Circuit), PLD (Programmable Logic Device), or FPGA (Field Programmable Gate Array).

[0080] Although embodiments of this invention have been described in detail above with reference to the drawings, the specific configuration is not limited to these embodiments and includes designs and the like that do not depart from the spirit of this invention. [Industrial applicability]

[0081] This invention is applicable to optical communication systems. [Explanation of Symbols]

[0082] 1, 1a, 1b… Estimation system, 2a, 2b… Filter generation device, 3… Memory device, 4… Sequence generation device, 5… Field generation device, 6a, 6b… Field estimation device, 7a, 7b… Error rate estimation device, 8a, 8b… Estimation device, 9a, 9b… Learning device, 10a, 10b… Model generation device, 21… Delay processing unit, 22… Error calculation unit, 23… Filter unit, 41… Sequence generation device, 51… Field generation device, 61… Filter Coefficient estimation unit, 62...filter, 63...feature generation unit, 64...filter unit, 71...photoelectric conversion unit, 72...noise processing unit, 73...decision unit, 74...error rate estimation unit, 91a, 91b...learning unit, 92...error calculation unit, 100...optical communication system, 101...processor, 102...memory, 103...storage device, 104...communication unit, 110...first communication device, 120...second communication device, 130...transmission line, 140...optical amplifier

Claims

1. A filter coefficient estimation unit estimates a second tap coefficient based on the input electric field waveform corresponding to the code sequence, the optical intensity of the input electric field waveform at the input end of the transmission line, and the transmission distance, using a learning model whose parameters have been updated to reduce the error between the correct label and the first tap coefficient. A first filter generates a simulated signal of the output electric field waveform at the output end of the transmission line based on the input electric field waveform corresponding to the code sequence and the estimated second tap coefficient, An error rate estimation unit estimates the error rate of the code sequence at the output end of the transmission line based on the simulated signal of the output electric field waveform. An estimation device equipped with the following features.

2. The system further includes a feature generation unit that generates the third tap coefficient of the second filter as a feature quantity of the input electric field waveform so that the reference electric field waveform that has passed through the second filter matches the input electric field waveform. The estimation device according to claim 1, wherein when the filter coefficient estimation unit estimates the second tap coefficient based on the input electric field waveform, the light intensity, and the transmission distance corresponding to the code sequence, it uses the learning model to estimate the second tap coefficient based on the third tap coefficient, the light intensity, and the transmission distance.

3. A learning unit that generates a first tap coefficient using a learning model based on the input electric field waveform at the input end of the transmission line, the light intensity of the input electric field waveform, and the transmission distance, The system includes an error calculation unit that obtains a correct label generated based on the output electric field waveform at the output end of the transmission line and the input electric field waveform, and calculates the error between the correct label and the first tap coefficient. The learning unit updates the parameters of the learning model in order to reduce the error. Learning device.

4. The system further includes a feature generation unit that generates the second tap coefficient of the filter as a feature quantity of the input electric field waveform so that the reference electric field waveform that has passed through the filter matches the input electric field waveform. The learning device according to claim 3, wherein when the learning unit generates the first tap coefficient based on the input electric field waveform, the light intensity, and the transmission distance at the input end of the transmission line, it generates the first tap coefficient using the learning model based on the second tap coefficient, the light intensity, and the transmission distance.

5. An estimation method performed by an estimation device, The steps include: using a learning model whose parameters have been updated to minimize the error between the correct label and the first tap coefficient, estimating the second tap coefficient based on the input electric field waveform corresponding to the code sequence, the optical intensity of the input electric field waveform at the input end of the transmission line, and the transmission distance; A step of generating a simulated signal of the output electric field waveform at the output end of the transmission line based on the input electric field waveform corresponding to the code sequence and the estimated second tap coefficient, A step of estimating the error rate of the code sequence at the output end of the transmission line based on the simulated signal of the output electric field waveform. An estimation method that includes [this].

6. A learning method performed by a learning device, A step of generating a first tap coefficient using a learning model based on the input electric field waveform at the input end of the transmission line, the light intensity of the input electric field waveform, and the transmission distance, The steps include obtaining a correct label generated based on the output electric field waveform at the output end of the transmission line and the input electric field waveform, and calculating the error between the correct label and the first tap coefficient. The calculation step includes updating the parameters of the learning model to reduce the error. Learning methods.

Images