Method and device for estimating wavelength dispersion value of optical transmission path

Abstract

Disclosed is a method for estimating the wavelength dispersion value of an optical transmission path. Disclosed is a method for training a neural network, the method including: a step for generating a histogram showing the frequency of each predetermined signal intensity section on the basis of a received time-series signal that has passed through an optical transmission path having a predetermined wavelength dispersion value; and a step for training a neural network using the shape of the histogram and a predetermined wavelength dispersion value. Also disclosed is a method for estimating a wavelength dispersion value, the method including: a step for generating a histogram showing the frequency of each predetermined signal intensity section on the basis of a received time-series signal that has passed through an optical transmission path having an unknown wavelength dispersion value that is an estimation target; and a step for inputting the shape of the histogram to the trained neural network and outputting an estimated wavelength dispersion value.

Description

Method and apparatus for estimating chromatic dispersion value of optical transmission path

[0001] The present disclosure relates to a method and apparatus for estimating a chromatic dispersion value of an optical transmission line.

[0002] As optical communication networks become larger in scale, there is a demand for both higher speeds and lower costs for transmitted signals. In particular, optical communication using Intensity Modulation and Direct-Detection (hereinafter referred to as "IMDD") does not require the expensive transmission and reception systems required for coherent optical communication, and has been used primarily for short-distance optical wiring within racks in data centers and station buildings, or between racks. Meanwhile, with the recent increase in demand for data centers, the application of IMDD communication methods to communications between data centers is also being considered. In the case of long-distance signal transmission using IMDD, a major issue is the compensation of inter-symbol interference due to chromatic dispersion in the optical fiber that constitutes the optical transmission path.

[0003] In general long-distance communications, it is possible to recover the original signal by obtaining the intensity and phase components of the signal light using coherent detection and applying inverse dispersion using digital signal processing (DSP). On the other hand, in the case of the IMDD system, it is difficult to recover the original signal using the above-mentioned digital signal processing because the phase component cannot be extracted. Therefore, various methods have been proposed, such as reducing the degradation of the error rate (BER) due to inter-symbol interference using the maximum likelihood estimation (MLSE) method (Non-Patent Document 1), estimating and compensating for the phase signal from the intensity signal using a phase recovery algorithm (Non-Patent Document 2), and applying inverse dispersion using an optical filter.

[0004] In either method, it is generally assumed that the impulse response, i.e., chromatic dispersion, of the optical transmission line is known. Although a blind processing MLSE method has also been proposed for cases where the impulse response of the transmission line is unknown (Non-Patent Document 3), the method of Non-Patent Document 3 requires very heavy processing, which leads to high costs for digital signal processing circuits (DSPs). Therefore, it is desirable to monitor the chromatic dispersion of the optical transmission line in some way.

[0005] As a method for monitoring chromatic dispersion in an optical transmission line, for example, an optical interference method using a nonlinear element as disclosed in Non-Patent Document 4 has been proposed. In addition, for dispersion compensation applications of femtosecond lasers other than optical communications, frequency-resolved optical gating (FROG) (Non-Patent Document 5) and the like are used.

[0006] However, these monitoring methods require additional equipment dedicated to monitoring chromatic dispersion, making miniaturization and cost reduction of these methods challenging. Therefore, a signal processing-based dispersion estimation method that does not require additional dedicated equipment is desirable.

[0007] As a signal processing-based variance estimation method, a method of estimating variance by digital processing of complex signals has been proposed for coherent communication (Non-Patent Document 6). However, this method cannot be applied to IMDD transmission because phase information is lost.

[0008] For IMDD transmission, a signal processing-based dispersion estimation method has been proposed in which a clock tone is superimposed and the RF spectrum is analyzed (Non-Patent Document 7). However, this method cannot be easily used because it violates optical communication standards due to the superimposition of a clock tone. Furthermore, this method has issues such as insufficient accuracy and dynamic range, and the inability to determine whether chromatic dispersion is positive or negative.

[0009] Visintin, M., P. Poggiolini, and G. Bosco. "Long-haul optically uncompensated IMDD transmission with MLSE using the M-method." IEEE Photonics Technology Letters 19.16 (2007): 1230-1232.Xiang, Yating, et al. "Phase retrieval of complex OFDM signal by solving temporal transport-of-intensity equation." IEEE Photonics Technology Letters 33.18 (2021): 1006-1009.Gorshtein, Albert, et al. "Blind channel estimation for MLSE receiver in high speed optical communications: theory and ASIC implementation." Optics Express 21.19 (2013): 21766-21789.Ng, T. T., et al. "In-band OSNR and chromatic dispersion monitoring using a fibre optical parametric amplifier." Optics Express 13.14 (2005): 5542-5552.DeLong, K. W., et al. "Frequency-resolved optical gating with the use of second-harmonic generation." JOSA B 11.11 (1994): 2206-2215.Xu, Tianhua, et al. "Chromatic dispersion compensation in coherent transmission system using digital filters." Optics express 18.15 (2010): 16243-16257.Petersen, Martin Nordal, et al. "Online chromatic dispersion monitoring and compensation using a single inband subcarrier tone." IEEE Photonics Technology Letters 14.4 (2002): 570-572.

[0010] In order to solve the above-described problems, an object of the present disclosure is to provide a method for estimating the chromatic dispersion value of an IMDD transmission line with high accuracy and in a wide dynamic range.

[0011] To achieve this object, the present disclosure is characterized by generating a histogram based on the intensity of a received time-series signal of an optical signal propagating through an optical transmission line, and estimating the chromatic dispersion value using the relationship between the shape of the histogram and the chromatic dispersion value. By using the histogram generated based on the intensity of the received time-series signal, it becomes possible to monitor the chromatic dispersion value of an IMDD transmission line based on signal processing.

[0012] The relationship between the shape of the histogram and the dispersion value depends on various variables such as the signal format and Baud rate used in the transmission path, but by approximating this relationship in advance using a neural network, it becomes possible to estimate the chromatic dispersion value from the shape of the histogram of the received time-series signal.

[0013] An embodiment of the present disclosure is a training method for generating a neural network for estimating a chromatic dispersion value that approximates the relationship between a histogram shape and a chromatic dispersion value using a neural network. The training method of an embodiment of the present disclosure is a neural network training method for generating a trained neural network for estimating the chromatic dispersion value of an optical transmission line, the neural network training method including the steps of: generating a histogram indicating the frequency of each predetermined signal intensity interval based on a received time-series signal after passing through an optical transmission line having a predetermined chromatic dispersion value; and training the neural network using the histogram shape and the predetermined chromatic dispersion value.

[0014] Another embodiment of the present disclosure is a chromatic dispersion value estimation method for estimating a chromatic dispersion value of an optical transmission line to be estimated using a trained neural network for estimating a chromatic dispersion value that approximates the relationship between the shape of a histogram and the chromatic dispersion value using a neural network. The chromatic dispersion value estimation method of another embodiment of the present disclosure is a chromatic dispersion value estimation method including the steps of: generating a histogram indicating the frequency of each predetermined signal intensity section based on a received time-series signal after passing through an optical transmission line having an unknown chromatic dispersion value to be estimated; and inputting the shape of the histogram into the trained neural network and outputting an estimated chromatic dispersion value of the optical transmission line having the unknown chromatic dispersion value to be estimated.

[0015] According to the method of the present disclosure, it is possible to generate a neural network that enables estimation of chromatic dispersion values ​​based on signal processing. Furthermore, according to the method of the present disclosure, it is possible to estimate chromatic dispersion values ​​based on signal processing using a trained neural network.

[0016] The method disclosed herein has the excellent advantage of eliminating the need to superimpose a clock tone or the like on the optical transmission signal, as in the conventional techniques described above, and thus not violating communication standards. Furthermore, the method disclosed herein enables estimation of chromatic dispersion values ​​over a wide dynamic range by varying the Baud rate of the transmission signal during neural network training. Furthermore, by inserting a chromatic dispersion element before the transmission signal receiver, it becomes possible to determine whether the chromatic dispersion value is positive or negative.

[0017] FIG. 1 is a schematic configuration diagram of a model of a target optical transmission line. FIG. 2 is a flowchart of a method for estimating a chromatic dispersion value of an optical transmission line according to a first embodiment of the present disclosure. FIG. 3 is a functional block diagram of a chromatic dispersion value estimation device for implementing the method for estimating a chromatic dispersion value of an optical transmission line according to the present disclosure. FIG. 4 is a diagram showing an outline of a neural network model. FIG. 5 is a diagram showing changes in a histogram when a time-series signal is passed through optical transmission lines of multiple distances while changing the Baud rate of the signal. FIG. 6 is a diagram showing the results of estimating, using a simulation, a chromatic dispersion value for a signal rate of 10 GBaud. FIG. 7 is a flowchart of a method for estimating a chromatic dispersion value of an optical transmission line according to a second embodiment of the present disclosure. FIG. 8 is a schematic configuration diagram of an optical transmission line model 800 for explaining a wavelength estimation method according to a third embodiment of the present disclosure.

[0018] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. Note that the following description is an example, and some configurations may be changed without departing from the gist of the present disclosure. The same or similar reference numerals indicate the same or similar elements, and repeated explanations may be omitted. Numerical values ​​in the following description are examples, and other numerical values ​​may be used in the present disclosure without departing from the gist of the present disclosure.

[0019] (Embodiment) FIG. 1 is a schematic diagram of a model of an optical transmission path that is the subject of this disclosure. The optical transmission path model 100 includes an optical transmitter 101, an optical transmission path 102 that transmits a signal from the optical transmitter, a receiver (PD) 103 that receives the signal that has passed through the optical transmission path, an analog-to-digital converter (ADC) 104 that converts the analog signal received by the receiver 103 into a digital signal, and a digital signal processor (DSP) 105 that processes the digital signal. The optical transmitter 101 includes an optical intensity modulator and generates an optical intensity-modulated transmission time-series signal x. The transmission time-series signal x generated by the optical transmitter 101 is introduced into an optical transmission path with an unknown chromatic dispersion value D, and a reception time-series signal y is acquired by the receiver (PD) 103. In this disclosure, a function for estimating the unknown dispersion D is realized by performing signal processing on the digital signal processor (DSP) based on the reception time-series signal y, which is made up of an optical intensity signal acquired by the PD 103. Therefore, it is possible to monitor the chromatic dispersion value even in an intensity modulation / direct reception (IMDD) transmission path.

[0020] Although the model 100 in FIG. 1 illustrates a single-wavelength transmission, it may also be a wavelength-multiplexed WDM (Wavelength Division Multiplexing) transmission path or a spatially multiplexed SDM (Space Division Multiplexing) transmission path. In this case, the optical transmitter 101, PD 103, ADC 104, and DSP 105 exist in the same number as the number of WDM or SDM channels. Furthermore, the chromatic dispersion value estimation method disclosed herein can estimate the chromatic dispersion value D of the optical transmission path independently for each wavelength and spatial channel using the respective DSPs, even if the chromatic dispersion value D differs for each wavelength and spatial channel. This has the excellent characteristic of being applicable to transmission paths that utilize wavelength and spatial routing, such as multiplexed reconfigurable optical add / drop multiplexers (ROADMs).

[0021] 2 is a flowchart of a method for estimating a chromatic dispersion value of an optical transmission line according to a first embodiment of the present disclosure. The method according to the first embodiment of the present disclosure is roughly divided into two methods: a method for training a neural network as a learning phase shown in FIG. 2( a), and a method for estimating a chromatic dispersion value using a trained neural network as an inference phase shown in FIG. 2( b).

[0022] (Learning Phase) The learning phase, which is a first method of the first embodiment of the present disclosure, inputs a transmitted time-series signal x generated by intensity modulation using a predetermined random number sequence to optical transmission lines having various known chromatic dispersion values ​​D, acquires a received time-series signal y, which is the signal intensity of the optical time-series signal that has passed through the optical transmission line, and generates a histogram, which is a frequency distribution of the signal intensity.

[0023] The generated histogram shape and chromatic dispersion value are then used to train and learn a neural network. That is, the learning method of this embodiment aims to generate a neural network model by using a neural network to learn the relationship between the histogram shape and the chromatic dispersion value, so that when input to the input nodes (i.e., when the frequency of each signal strength segment is input to each input node corresponding to each signal strength segment), the value output from the output node will be close to the known chromatic dispersion value D.

[0024] Step 201 is a process of acquiring a received time-series signal y that has passed through an optical transmission line having a known chromatic dispersion value D and generating a histogram.

[0025] The received time-series signal y is a signal obtained after a transmission time-series signal x, input to an optical transmission line having a known chromatic dispersion value as a learning target, passes through the optical transmission line. The transmission time-series signal x is generated by intensity-modulating a predetermined random number sequence using, for example, the intensity modulator shown in FIG. 1 . The predetermined random number sequence used to generate this transmission time-series signal x may be discretized based on the multi-level modulation index supported by the optical transmitter (e.g., in the case of PAM4, the random number sequence is discretized using four values). The predetermined random number sequence used here may be uniform random numbers or Gaussian-distributed random numbers. However, the transmission time-series signal x generated here is generated using a method similar to that described in the inference phase. Furthermore, if the statistical distribution of signals actually flowing through the optical transmission line is known, the transmission time-series signal x generated by intensity-modulating the random number sequence based on that distribution function can be used in the learning phase to estimate dispersion in real time from waveforms during actual operation in the inference phase described below.

[0026] In the learning phase, it is necessary to repeatedly input a transmission time-series signal x to optical transmission lines having various chromatic dispersion values ​​D to obtain a reception time-series signal y after passing through the optical transmission line. The operation of obtaining this reception time-series signal y may be performed using a computer simulation, or may be performed through a real-device experiment using an actual optical transmission line. If the transmitter and receiver of the optical transmission line model shown in Figure 1 can be modeled, there is an advantage that learning can be performed at low cost and high speed through computer simulation. If learning is performed through a real-device experiment, there is an advantage that learning can be performed without modeling the transmitter and receiver.

[0027] The histogram generated here shows the distribution of the signal strength of the received time series signal, and specifically, it shows a predetermined number of bins (light intensity intervals on the x-axis) divided into light intensities, and the frequency of each bin is shown as a bar (frequency on the y-axis).

[0028] Step 202 is a process of training a neural network using the histogram shape and a predetermined chromatic dispersion value D. The generated histogram shape is input to the neural network model, which outputs an estimated chromatic dispersion value De.

[0029] The frequency of each intensity interval of the histogram is input to a corresponding input node of the neural network model, which outputs an estimated chromatic dispersion value De. The neural network model used in the present disclosure may be, for example, a fully connected or convolutional neural network model.

[0030] The neural network model is trained by comparing the estimated chromatic dispersion value De obtained at this time with the known chromatic dispersion value D. A general-purpose method such as the backpropagation method can be used to train the neural network model.

[0031] Step 203 is a process for determining the number of repetitions of the learning phase. Until it is determined in step 203 that a predetermined number of repetitions has been reached, step 204 changes the optical transmission path used for learning and repeats steps 201 and 202. It is desirable to use optical transmission paths with as many different chromatic dispersion values ​​D as possible as the optical transmission paths with known chromatic dispersion values ​​D used for learning. When the predetermined number of repetitions has been reached, the process proceeds to step 205, where the learning phase is terminated.

[0032] It is desirable that the transmitted time series signal x (and therefore the received time series signal y) used in the method of the present disclosure is composed of a signal sequence of at least 1,000 points for generating a histogram, and that the number of bins (the number of received signal strength intervals on the x-axis) of the generated histogram is 10 or more.

[0033] The estimated error ||De-D|| of the obtained estimated chromatic dispersion value 2 If the specification conditions are not met, the problem can be addressed by changing the neural network model used or increasing the number of parameters.

[0034] As described above, the learning phase of this embodiment can be performed by processing the received time-series signal y based on the signal intensity, and therefore it is possible to generate a neural network for estimating chromatic dispersion values ​​even in an intensity modulation / direct reception (IMDD) transmission path.

[0035] (Inference Phase) Next, the inference phase of this embodiment will be described. The inference phase shown in Fig. 2B is a method for estimating the chromatic dispersion value of an unknown optical transmission line by using a trained neural network model.

[0036] The chromatic dispersion value estimation method, which is the inference phase of this embodiment, generates a histogram indicating the frequency of each predetermined signal strength interval based on a received time-series signal y that has passed through an optical transmission line to be estimated and has an unknown chromatic dispersion value. Then, by inputting the shape of this histogram into a trained neural network model (inputting the frequency of each signal strength interval into each input node corresponding to each signal strength interval), it is possible to obtain an estimated chromatic dispersion value De of the optical transmission line to be estimated, which is output from the neural network.

[0037] Step 206 is a process of generating a histogram showing the frequency of each predetermined signal strength section based on the received time-series signal y that has passed through the optical transmission line to be estimated and has an unknown chromatic dispersion value.

[0038] The received time-series signal y is a signal obtained after the transmitted time-series signal x input to the optical transmission line to be estimated has passed through the optical transmission line. The transmitted time-series signal x is generated using the same random number sequence as that used to generate the transmitted time-series signal x in the learning phase.

[0039] The histogram indicates the distribution of the signal strength of the received time-series signal, and is generated in the same bins as the histogram generated in the learning phase.

[0040] Step 207 is a process of inputting the shape of the histogram to the input node of the trained neural network model, and obtaining an estimated chromatic dispersion value De, which is the output of the neural network.

[0041] By inputting the shape of the histogram into a trained neural network that approximates the relationship between the shape of the histogram and the chromatic dispersion value, it is possible to estimate the chromatic dispersion value of the optical transmission line to be estimated.

[0042] In this case, when using a trained neural network that has been trained using a transmission time-series signal x generated by intensity modulation using a uniform random number sequence, a possible usage pattern is that the chromatic dispersion value is estimated using the chromatic dispersion estimation method disclosed herein as an initialization operation when connecting an optical transmission line, and then the estimated chromatic dispersion value is used to perform actual operation with a normal signal.

[0043] On the other hand, when estimating the chromatic dispersion value using a signal that has the same statistical properties as the optical signal during actual operation, that is, when using a trained neural network that has been trained using a signal that has the same statistical properties as the optical signal during actual operation, it becomes possible to estimate the chromatic dispersion value in real time using the signal during actual operation.

[0044] As described above, the inference phase of this embodiment can be performed by processing based on the signal strength of the received time-series signal y, so it is possible to estimate the chromatic dispersion value even in an intensity modulation / direct reception (IMDD) transmission path.

[0045] Fig. 3 is a functional block diagram for explaining the functions of a chromatic dispersion value estimation device for realizing a method for training a neural network used to estimate a chromatic dispersion value according to the first embodiment of the present disclosure described in Fig. 2, and a method for estimating a chromatic dispersion value of an optical transmission line using a trained neural network. This chromatic dispersion value estimation device is implemented, for example, as the digital signal processing device (105) in Fig. 1.

[0046] The chromatic dispersion value estimation apparatus of this embodiment includes, as functional blocks, a received signal acquisition unit 301, a histogram generation unit 302, an input unit 303, a machine learning unit 304, a chromatic dispersion value estimation unit 305, and a neural network 307 stored in a storage device 306 or the like.

[0047] The received signal acquisition unit 301 has a function of acquiring a received time-series signal y after passing through an optical transmission line. The histogram generation unit 302 has a function of generating a histogram indicating the frequency of each predetermined signal strength section from the received time-series signal y acquired by the received signal acquisition unit. The input unit 303 receives as input a predetermined chromatic dispersion value D of the optical transmission line from which the received time-series signal y was obtained.

[0048] The machine learning unit 304 uses the shape of the histogram generated by the histogram generating unit 302 based on the received time-series signal y after passing through an optical transmission line having a predetermined chromatic dispersion value and the predetermined chromatic dispersion value to train and learn the neural network 307. Specifically, the learning phase of FIG. 2( a) is executed.

[0049] The chromatic dispersion value estimation unit 305 inputs the shape of the histogram generated by the histogram generation unit 302 based on the received time-series signal y after passing through the optical transmission path to be estimated to the trained neural network 307, and outputs the chromatic dispersion value of the optical transmission path to be estimated. Specifically, the inference phase of FIG. 2B is executed.

[0050] Each of the functional blocks of the chromatic dispersion value estimation apparatus 300 shown in Fig. 3 can be implemented as hardware using an integrated circuit or the like equipped with at least one processor, such as the digital signal processing device in Fig. 1. Furthermore, these functional blocks can also be implemented as software by reading a program onto a main storage device and executing it using one or more processors.

[0051] 3 has been described as a chromatic dispersion value estimation device that includes a machine learning unit and a chromatic dispersion value estimation unit, it may also be configured as a learning device that executes only the learning phase without including the chromatic dispersion value estimation unit 305. Similarly, it may also be configured as a chromatic dispersion value estimation device that executes only the inference phase without including the machine learning unit 304.

[0052] As a specific example of chromatic dispersion value estimation using the method for estimating a chromatic dispersion value according to the first embodiment of the present disclosure, the results of a study using a simulation will be shown.

[0053] In this simulation, the signal transmission method was the intensity modulation / direct reception (IMDD) method. The transmission time-series signal x was generated by discretizing uniform random numbers in the [0:1] interval into four values, and this signal was transmitted from the optical transmitter as a PAM4 signal. The power of the transmission time-series signal x input to the optical transmission line was 0 dBm, the wavelength was 1.3 μm, and the S / N ratio was 30 dB. A study was conducted to estimate the transmission distance when it was unknown using an optical transmission line with a chromatic dispersion value of 17 ps / nm / km.

[0054] The optical transmission path simulation used a split-step Fourier transform. The signal sequence length M was set to 2^13 (8192). The baud rate was varied from 1 to 25 GBaud, and the signals were received at a 4x oversampling ratio at the receiving end.

[0055] FIG. 4 shows an outline of the neural network model used in the simulation. This neural network model uses a fully connected network with a single hidden layer. As shown in FIG. 4, in the present disclosure, the shape of the histogram of the signal strength of the received time-series signal y, i.e., the frequency value stored in each bin, is introduced into the neural network as the value of the input node. The data input to the input node is converted into an estimated output value De (output node) by a fully connected network with a single hidden layer.

[0056] Figure 5 shows how each histogram changes when the Baud rate of a time-series signal is changed and the signal passes through optical transmission paths of multiple distances. In each histogram in Figure 5, the horizontal axis represents the received field strength of the received time-series signal y, and the vertical axis represents the frequency of each received field strength. As mentioned above, Figure 5 shows the results of a simulation assuming a PAM4 signal format, so the transmitted time-series signal x input to the optical transmission path is discretized into four values. Therefore, the shapes of the histograms for a transmission distance of 1 km in each of Figures 5(a) to 5(c) are all concentrated around four similar values ​​because the influence of chromatic dispersion in the optical transmission path is small.

[0057] In the case of 1 GBaud in Figure 5(a), the histograms for 1 km, 10 km, and 50 km almost overlap, with a change in the shape of the histogram at 200 km. On the other hand, in the case of 25 GBaud in Figure 5(c), the shape of the histogram at 10 km is significantly different from that at 1 km, indicating that a significant change in the shape of the histogram occurs after 10 km transmission.

[0058] As described above, there exists a transmission distance, i.e., a magnitude of the chromatic dispersion value, at which the sensitivity of the dispersion estimation value increases for each Baud rate of the signal used. From the results shown in Figures 5(a) to 5(c), it can be seen that when estimating the chromatic dispersion value (the range of the chromatic dispersion value corresponding to that distance range) of an optical transmission line of a distance of several kilometers to 80 kilometers that is the target for communications using the IMDD system, it is desirable to use a signal of about 10 GBaud.

[0059] For example, when the signal format is NRZ transmission, the Baud rate can be effectively reduced by transmitting the same signal at multiple sample points (for example, in the case of 30 GBaud, a pseudo 10 GBaud signal waveform can be generated by setting the same symbol value at three consecutive points). Therefore, when executing the chromatic dispersion value inference method, it is possible to set a time series signal with an appropriate Baud rate relatively freely. In this way, changing the Baud rate does not pose a problem in actual operation.

[0060] Next, with reference to FIG. 6, the results of estimating chromatic dispersion values ​​by simulation with the signal rate fixed at 10 GBaud will be described.

[0061] In this simulation, the transmission time-series signal x was generated by discretizing uniform random numbers in the [0:1] interval into four values, and this signal was set to be transmitted from the optical transmitter as a PAM4 signal. The power of the transmission time-series signal x input to the optical transmission line was 0 dBm, the wavelength was 1.3 μm, and the S / N ratio was 30 dB. We investigated the case of estimating chromatic dispersion when the transmission distance was unknown in an optical transmission line with a chromatic dispersion value of 17 ps / nm / km.

[0062] The optical transmission path simulation used a split-step Fourier transform. The signal sequence length M was set to 2^13 (8192). The baud rate was 10 GBaud, and the signal was received with a 4x oversampling ratio at the receiving end.

[0063] The neural network model used a fully connected network with one intermediate layer. The number of bins in the input histogram was set to 50, the number of hidden layer nodes was 300, and the number of output nodes was 1. The activation function f used was tanh, and the error backpropagation method was used for learning, the mean squared error (MSE) was used as the cost function, and Adam was used as the update rule.

[0064] The change in MSE during learning is shown in Figure 6(a). Here, Train indicates the MSE during learning, and Test indicates the results of inference performed after each learning cycle. The MSE monotonically decreases, demonstrating that the learning method of the present disclosure can perform machine learning of a neural network to estimate the chromatic dispersion value of an optical transmission line.

[0065] 6(b) is a diagram showing the relationship between the actual chromatic dispersion value and the estimated chromatic dispersion value. Here, the case under consideration is one in which the chromatic dispersion value of an optical transmission line having the above-mentioned 17 ps / nm / km is estimated, and the transmission distance is unknown. Therefore, the horizontal and vertical axes are plotted in units of km, which is the transmission distance. As shown in the figure, a linear function relationship centered on the actual value is obtained, demonstrating that the chromatic dispersion value estimation method of the present disclosure can estimate the unknown chromatic dispersion value of an optical transmission line.

[0066] In Figure 6(b), 10,000 data points are plotted, which makes it appear that there are many outliers, but the estimation error was ±1 km or less for transmission distances of 80 km or less, and approximately ±3 km for transmission distances of 80 km or more. For example, at a Baud rate of 25 G, the unit symbol length corresponds to 40 ps, ​​and the dispersion of 1 km corresponds to 17 ps, so the impact of an estimation error of approximately ±1 km on compensation is sufficiently small.

[0067] The calculation speed for estimating chromatic dispersion values ​​according to the first embodiment of the present disclosure strongly depends on the performance and implementation method of the digital signal processing circuit (DSP) used for the calculation and the size of the neural network, but for example, the illustrated model is a relatively lightweight neural network model with approximately 15,000 parameters, and only approximately 30,000 calculations are required for one inference. This is significantly lower than the theoretical calculation speed (>100 GFLOPs / s) of a typical processor (CPU), and calculations can be completed in milliseconds or less.

[0068] As described above with reference to Figure 6(b), when the transmission distance via the optical transmission line is extended (exceeding 80 km in the above example), the estimation accuracy of the neural network model deteriorates because the signal baud rate is fixed at 10 GBaud. Also, in Figure 6(b), the error becomes large even at distances of a few kilometers or less.

[0069] 6(b), an optical transmission line with a transmission distance of several kilometers to 80 kilometers or less, or a range of chromatic dispersion values ​​corresponding to the transmission distance n, can be set as a section in which the dispersion estimation value is valid when a Baud rate of 10 G signal is used (hereinafter referred to as an "effective section"). As a problem of dynamic range, when the Baud rate is fixed, the estimation accuracy may deteriorate if the chromatic dispersion value of the optical transmission line to be estimated is larger (or smaller) than expected.

[0070] Second Embodiment As a second embodiment of the present disclosure, a method for solving the problem of the dynamic range causing deterioration in estimation accuracy will be proposed.

[0071] 7 is a flowchart of a method for estimating a chromatic dispersion value of an optical transmission line according to a second embodiment of the present disclosure. The method of this embodiment also includes two methods: a learning phase shown in FIG. 7( a), in which a neural network is trained for machine learning, and an inference phase shown in FIG. 7( b), in which a trained neural network is used to estimate a chromatic dispersion value.

[0072] In this embodiment, in the learning phase, machine learning of a neural network model is performed using time-series signals of multiple Baud rates. Then, in the inference phase, it is determined whether the obtained estimated chromatic dispersion value De is within the valid range of the used Baud rate, and if it is determined to be outside the valid range, the Baud rate is changed depending on whether it is above or below the valid range so that the chromatic dispersion value is within the valid range, thereby estimating the chromatic dispersion value. This makes it possible to effectively improve the dynamic range.

[0073] 7A is a flowchart of the learning phase of this embodiment. The learning phase of this embodiment trains the neural network multiple times using signals at multiple predetermined baud rates, thereby generating a neural network for estimating multiple learned chromatic dispersion values ​​trained at multiple different predetermined baud rates.

[0074] Step 701 is a process of setting a Baud rate to be used in the learning phase. Specifically, the Baud rate of the transmitted time-series signal x for obtaining the received time-series signal y in step 201 of FIG. 2A is selected and set from a plurality of predetermined Baud rates. In this embodiment, for example, five Baud rates are prepared: 0.5 Baud, 1 GBaud, 2 GBaud, 5 GBaud, and 10 GBaud.

[0075] Step 702 is a process of executing steps 201 to 203 in FIG. 2A at the Baud rate set in step 701.

[0076] Step 703 is a process of storing the trained neural network model trained in step 702. In step 703, the trained neural network model trained by executing the learning phase of Fig. 2 using a predetermined Baud rate is stored in, for example, a main storage device included in the digital signal processing device of Fig. 1.

[0077] Step 704 is a process for determining whether learning has been completed at all prepared Baud rates. Steps 701 to 704 are repeated until it is determined in step 704 that learning has been completed at all Baud rates. If learning has been completed at all Baud rates, the process proceeds to step 705, where the learning phase ends.

[0078] At this time, as explained in FIG. 6B, by evaluating the learning results at each Baud rate, it is possible to set an effective interval based on the magnitude of the estimated chromatic dispersion value corresponding to each Baud rate.

[0079] At the end of the learning phase, multiple trained neural network models are generated that have been trained using signals at each Baud rate.

[0080] 7B is a flowchart of the inference phase. The inference phase of this embodiment is intended to widen the dynamic range of the estimated chromatic dispersion value by appropriately selecting the baud rate of the signal used when estimating the magnitude of the unknown chromatic dispersion value of the optical transmission line to be estimated.

[0081] Step 706 is a process of selecting and setting one of the multiple Baud rates (e.g., five rates: 0.5 Baud, 1 GBaud, 2 GBaud, 5 GBaud, and 10 GBaud) prepared in the learning phase. Specifically, the Baud rate of the time-series signal when executing step 206 in FIG. 2B is selected from the multiple Baud rates. Here, for example, a transmission signal sequence x with a rate of 1 GBaud is set as the signal to be input to the optical transmission path to be estimated.

[0082] Step 707 is a process of estimating an estimated chromatic dispersion value De by a neural network trained using the time-series signal at the set Baud rate. Specifically, steps 206 to 208 of the inference phase in Fig. 2(b) are executed using the time-series signal at the set Baud rate.

[0083] Step 708 is a process for verifying the reliability of the obtained estimated chromatic dispersion value De. Specifically, it is determined whether the estimated chromatic dispersion value De is within the valid range of the Baud rate used.

[0084] As described above, the effective interval for a 10 GBaud signal is set to, for example, several kilometers to 80 kilometers (the range of the magnitude of the chromatic dispersion value corresponding thereto). As described above, in the chromatic dispersion value estimation method of the present disclosure, when the transmission distance is long, that is, when the estimated chromatic dispersion value is large, a signal with a low Baud rate is more suitable, and when the transmission distance is short, that is, when the estimated chromatic dispersion value is small, a signal with a high Baud rate is more suitable.

[0085] If it is determined that the estimated chromatic dispersion value De is within the valid range, the process proceeds to step 712, where the estimated chromatic dispersion value De is adopted and the inference phase is terminated.

[0086] If it is determined in step 708 that the distance is outside the valid interval, the process proceeds to step 710, where it is determined whether the estimated chromatic dispersion value (distance of the optical transmission path) exceeds the valid interval. If it is determined that the distance exceeds the valid interval, the process proceeds to step 710, where the Baud rate used in the inference phase is set lower by one step. Then, step 707 is executed using the newly set Baud rate.

[0087] If it is not determined in step 709 that the confidence interval is exceeded (i.e., if it is determined that the confidence interval is exceeded), the process proceeds to step 710, where the Baud rate is increased by one step, and step 707 is then executed using the newly set Baud rate.

[0088] As described above, in the method for estimating the chromatic dispersion value of an optical transmission line of this embodiment, the Baud rate of the signal used to estimate the chromatic dispersion value is changed during the estimation, thereby making it possible to widen the range of dispersion values ​​that can be estimated with a predetermined accuracy.

[0089] In this embodiment, as in the first embodiment, both the learning phase and the inference phase can be performed by processing based on the signal strength of the received time-series signal y, and therefore, it can also be used in an intensity modulation / direct reception (IMDD) transmission path.

[0090] When the method of this embodiment is executed using the chromatic dispersion value estimation device of FIG. 3, the learning phase of this embodiment is executed by the machine learning unit, and the inference phase is executed by the chromatic dispersion value estimation unit.

[0091] 8 is a schematic diagram of an optical transmission line model 800 for explaining a wavelength estimation method according to a third embodiment of the present disclosure. In Fig. 8, the same reference numerals as in Fig. 1 are the same as in Fig. 1, and therefore their explanation will be omitted here.

[0092] The chromatic dispersion estimation method disclosed herein performs estimation based on changes in the shape of the histogram of the reception intensity of a reception time-series signal. Therefore, when an ideal transceiver is used, it is not possible in principle to distinguish between positive and negative chromatic dispersion values. For this reason, for example, as shown in FIG. 8 , an offset chromatic dispersion element 801 is placed in front of the receiver (PD) 103 to offset the chromatic dispersion value in the positive direction. Then, the inference phase is executed. That is, in this embodiment, the chromatic dispersion value estimation method, which is the inference phase of FIG. 2B or FIG. 7B, is executed using the reception time-series signal y that has passed through the chromatic dispersion element 801 in addition to the optical transmission path to be estimated, to obtain an estimated chromatic dispersion value De.

[0093] For example, when the estimated chromatic dispersion value obtained is large relative to the chromatic dispersion value of the offset chromatic dispersion element, it can be determined that the chromatic dispersion value estimated for the optical transmission line to be estimated (in this case, calculated as the difference between the estimated chromatic dispersion value and the chromatic dispersion value of the offset chromatic dispersion element) is a positive chromatic dispersion value. On the other hand, when the estimated chromatic dispersion value obtained is small relative to the chromatic dispersion value of the offset chromatic dispersion element, it can be determined that the chromatic dispersion value estimated for the optical transmission line to be estimated is a negative chromatic dispersion value. Thus, according to this embodiment, the excellent effect of being able to estimate both positive and negative dispersions is achieved.

[0094] According to the present disclosure, it is possible to generate a neural network for estimating chromatic dispersion values ​​based on signal processing, and to estimate chromatic dispersion values ​​based on signal processing using a trained neural network.

Claims

1. A neural network training method for generating a trained neural network for estimating the chromatic dispersion value of an optical transmission line, comprising: a step of generating a histogram indicating the frequency of each predetermined signal intensity section based on a received time-series signal after passing through an optical transmission line having a predetermined chromatic dispersion value; and a step of training a neural network using the shape of the histogram and the predetermined chromatic dispersion value.

2. The neural network training method according to claim 1, wherein the optical transmission line having the predetermined wavelength dispersion value is an optical transmission line modeled by simulation or an actual optical transmission line.

3. A neural network training method as described in claim 1, characterized in that the received time series signal is generated by inputting a transmitted time series signal generated from a predetermined random number sequence or a random number sequence corresponding to the distribution function of data used in actual operation into an optical transmission line having the predetermined wavelength dispersion value.

4. A neural network training method as described in claim 1, characterized in that the neural network is trained using the received time series signals at a plurality of predetermined baud rates, and a plurality of trained neural networks trained at the predetermined baud rates are generated.

5. A neural network training method according to any one of claims 1 to 4, wherein the optical transmission line having the predetermined chromatic dispersion value is an intensity modulation / direct reception (IMDD) transmission line.

6. A chromatic dispersion value estimation method comprising: a step of generating a histogram indicating the frequency of each predetermined signal intensity section based on a received time-series signal after passing through an optical transmission line having an unknown chromatic dispersion value to be estimated; and a step of inputting the shape of the histogram into a trained neural network and outputting an estimated chromatic dispersion value of the optical transmission line having an unknown chromatic dispersion value to be estimated.

7. A chromatic dispersion value estimation method according to claim 6, characterized in that the received time series signal is generated by inputting a transmitted time series signal generated from a predetermined random number sequence or a signal during actual operation into an optical transmission line having an unknown chromatic dispersion value that is the object of estimation.

8. The chromatic dispersion value estimation method according to claim 6, characterized in that the received time series signal is passed through an optical transmission line having an unknown chromatic dispersion value to be estimated, as well as a chromatic dispersion element for offset.

9. A method for estimating a chromatic dispersion value using a plurality of trained neural networks that have been trained using the neural network learning method of claim 4 using received time-series signals at a plurality of predetermined baud rates, comprising the steps of: selecting one baud rate from the plurality of predetermined baud rates; generating a histogram indicating the frequency of each predetermined signal strength section based on the received time-series signals at the selected one baud rate after passing through an optical transmission line having an unknown chromatic dispersion value that is the target of estimation; inputting the shape of the histogram into a trained neural network that has been trained using the received time-series signals at the selected one baud rate, and outputting an estimated chromatic dispersion value of the optical transmission line having an unknown chromatic dispersion value that is the target of estimation; determining whether the magnitude of the estimated chromatic dispersion value is within or outside the valid range corresponding to the selected one baud rate; and outputting the estimated chromatic dispersion value if it is determined to be within the valid range, and if it is determined to be outside the valid range, changing the selected one baud rate to another baud rate depending on whether the magnitude of the estimated chromatic dispersion value is above or below the valid range, and estimating the chromatic dispersion value again.

10. The chromatic dispersion value estimation method according to any one of claims 6 to 9, wherein the optical transmission line is an intensity modulation / direct reception (IMDD) transmission line.

11. A chromatic dispersion value estimation device for estimating the chromatic dispersion value of an optical transmission line, comprising: a received signal acquisition unit that acquires a received time-series signal after passing through an optical transmission line; an input unit to which a predetermined chromatic dispersion value is input; a histogram generation unit that generates a histogram indicating the frequency of each predetermined signal strength section based on the received time-series signal acquired by the received signal acquisition unit; a machine learning unit that trains a neural network using the predetermined chromatic dispersion value and the shape of the histogram generated by the histogram generation unit based on the received time-series signal after passing through an optical transmission line having a predetermined chromatic dispersion value; and a chromatic dispersion value estimation unit that inputs the shape of the histogram generated by the histogram generation unit based on the received time-series signal after passing through an optical transmission line having an unknown chromatic dispersion value to be estimated into the neural network trained by the machine learning unit, and outputs an estimated chromatic dispersion value of the optical transmission line having the unknown chromatic dispersion value to be estimated.

Images