Optical transmission and reception device and computer program
The proposed optical transceiver device uses a multi-layered symbol determination unit with normalization and area selection circuits to address waveform distortion in optical communication, reducing calculation time and memory usage while maintaining signal accuracy.
Patent Information
- Application Number
- PCT/JP2023/019017
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-22
- Publication Date
- 2024-11-28
AI Technical Summary
In optical communication using digital coherent technology, the received waveform is distorted due to dispersion and nonlinear characteristics of optical fibers, leading to increased calculation time in multilevel signal determination, even with reduced machine learning circuits.
A received symbol determination unit with multiple layers, including normalization and area selection circuits, is employed to select and process regions based on machine learning results, reducing the number of internal variables in each circuit and thereby decreasing calculation time.
This approach enables efficient symbol determination of multilevel signals in less calculation time while maintaining accuracy, with a slight deterioration in Q value but significant reduction in memory and computational requirements.
Smart Images

Figure JP2023019017_28112024_PF_FP_ABST
Abstract
Description
Optical transmitting and receiving device and computer program
[0001] The present invention relates to an optical transmitting / receiving device and a computer program technology.
[0002] In the field of optical communications, high-capacity multilevel optical communications using digital coherent technology are being widely introduced (see, for example, Non-Patent Document 1). Digital coherent technology applies the elemental technologies of wireless communications to optical communications. However, due to the unique condition of optical communications, which is the use of optical fiber as a transmission path, digital coherent technology is subject to the effects of various dispersions and nonlinear characteristics of optical fiber. This results in the problem of distortion of the received waveform.
[0003] To address this issue, methods of applying machine learning to compensate for distortion have been studied (e.g., Non-Patent Document 2 and Non-Patent Document 3). Multi-class classification is required to determine whether a multi-value signal, such as 16QAM or 64QAM, is received. Therefore, as the number of values increases, the number of internal variables in the machine learning circuit increases, which tends to lengthen the calculation time. Non-Patent Document 2 and Non-Patent Document 3 employ a method that uses a support vector machine (hereinafter referred to as "SVM") for machine learning and reduces the number of SVM circuits used.
[0004] Fig. 21 is a diagram showing an example of a conventional circuit configuration using an SVM circuit. Fig. 22 is a diagram showing an example of a signal point constellation for 16QAM. Note that Figs. 21 and 22 are drawings quoted from Non-Patent Document 2. In the following explanation, a configuration for determining a 16QAM signal consisting of signal points 0 to 15 will be described as an example. Signal points 0 to 15 can be expressed in binary as four bits, 0000 to 1111. In the case of the signal point constellation in Fig. 22, the most significant two bits indicate the position in the x-axis direction, and the least significant two bits indicate the position in the y-axis direction.
[0005] First, consider the two most significant bits. If the signal points are divided according to whether the most significant bit of the two most significant bits is "0" or "1", they are divided into two regions. This is then classified into two regions using an SVM. Next, if the signal points are divided according to whether the least significant bit of the two most significant bits is "0" or "1", they are similarly divided into two regions. This is then classified into two regions using an SVM. The same process is performed for the y-axis. As a result, if a total of four SVM classifiers are used on both the x- and y-axes, 16 signal points can be determined.
[0006] S. Tsukamoto et al., OFC2005, PDP29.M. Li et al., IEEE Photon. J., vol. 5, no. 6, 2013.Y. Han et al., IEEE Photon. J, vol6, no. 5, 2014.
[0007] However, even if the number of machine learning circuits could be reduced, the amount of calculation required for each individual machine learning circuit was not taken into consideration, resulting in an increase in the calculation time.
[0008] In view of the above circumstances, an object of the present invention is to provide a technique that can realize symbol determination of a multilevel signal for a received optical signal in a shorter calculation time.
[0009] One aspect of the present invention includes a received symbol decision unit that performs symbol decision on a multilevel signal for a received optical signal and restores data based on the result of the symbol decision, the received symbol decision unit including region selection layers from 1st layer to Mth layer (M is an integer of 1 or more) and one region decision layer, the mth layer (m is an integer of 1 or more and M or less) region selection layer including normalization circuits and region selection circuits in the same number as the number of region divisions in the m-1th region selection layer ("1" when m=1), the normalization circuits amplify the amplitude of the received multilevel signal, and the mth layer region selection circuit performs symbol decision on the received multilevel signal according to the number of layers (1 to M). an optical transmission / reception device that selects, from a plurality of regions, a region to which a symbol of the received multi-level signal belongs based on a result of machine learning, and outputs the selected region to a next layer; the region selection circuit of the mth layer performs processing according to the region selected by the region selection circuit of the (m-1)th layer; the region determination layer includes region determination circuits according to the total number of outputs of the region selection circuits in the region selection layer of the Mth layer; and the region determination circuit selects, based on a result of machine learning, a region to which a symbol of the multi-level signal received from the region selection layer of the Mth layer belongs, and outputs a determination result indicating the symbol of the multi-level signal according to the selected region.
[0010] One aspect of the present invention includes a received symbol decision unit that performs symbol decision on a multilevel signal for a received optical signal and restores data based on the result of the symbol decision, the received symbol decision unit including region selection layers from 1st layer to Mth layer (M is an integer of 1 or more) and one region decision layer, the mth layer (m is an integer of 1 or more and M or less) region selection layer including the same number of region selection circuits as the number of region divisions in the m-1th region selection layer ("1" when m=1), the mth layer region selection circuit selecting a region to which a symbol of the received multilevel signal belongs from among a plurality of regions according to the number of layers (1 to M) based on the result of machine learning for the received multilevel signal. an optical transmission / reception device according to the present invention, in which an mth layer includes an mth region selection circuit, a nth region selection circuit, and an m-1th region selection circuit, each of which performs processing according to the region selected by the m-1th region selection circuit; the mth region determination layer includes region determination circuits according to the total number of outputs of the region selection circuits in the Mth region selection layer; the region determination circuit selects a region to which a symbol of the multi-level signal received from the Mth region selection layer belongs based on a result of machine learning, and outputs a determination result indicating the symbol of the multi-level signal according to the selected region; and the machine learning result is a result obtained by executing machine learning using appropriate hyperparameter values in each layer.
[0011] One aspect of the present invention is a computer program for causing a computer to function as the optical transmitting and receiving device described above.
[0012] According to the present invention, it is possible to realize symbol determination of a multilevel signal for a received optical signal in a shorter calculation time.
[0013] 1 is a diagram showing an example of a system configuration of an optical communication system 100 of the present invention. FIG. 1 is a diagram showing an example of a detailed configuration of a transmission signal selection unit 121. FIG. 2 is a diagram showing an example of a detailed configuration of a received symbol decision unit 212. FIG. 3 is a diagram showing an example of a detailed configuration of a decision unit 2122. FIG. 4 is a diagram showing an example of a specific configuration of a normalization circuit 41. FIG. 5 is a diagram showing an example of a specific configuration of a region selection circuit 42. FIG. 6 is a diagram showing an example of a configuration of a region decision circuit 51 using an SVM. FIG. 7 is a diagram showing an example of a configuration of a region decision circuit 51 using an NN. FIG. 8 is a diagram showing an example of a configuration in which two-branch region selection circuits 42 are combined in three layers to decide a 16QAM signal. FIG. 9 is a diagram showing an outline of processing in a configuration in which three-branch region selection circuits 42 are combined in three layers to decide a 16QAM signal. FIG. 10 is a diagram showing the principle of SVM. FIG. 11 is a diagram showing the results of a simulation of the operation of a conventional method. FIG. 12 is a diagram showing the results of a simulation of the operation of this embodiment (proposed method). FIG. 13 is a diagram showing the results of a simulation of the operation of the conventional method. FIG. 14 is a diagram showing the results of a simulation of the operation of this embodiment (proposed method). FIG. 15 is a diagram showing the results of a comparison of the operations of the conventional method and the proposed method by simulation. It is a figure which shows the result of comparing the operation of the conventional system and the proposed system by simulation. It is a figure which shows the first modified example. It is a figure which shows the second modified example. It is a figure which shows the conventional circuit configuration example which uses an SVM circuit. It is a figure which shows the example of the signal point arrangement in 16QAM.
[0014] 1 is a diagram showing an example of the system configuration of an optical communication system 100 according to an embodiment of the present invention. The optical communication system 100 includes a transmitting device 10, a receiving device 20, and a control device 30.
[0015] The transmitting device 10 includes a multiplexing device 11 and an optical modulation device 12. The optical modulation device 12 includes a transmission signal selection unit 121 and an optical modulation unit 122. Optical data signals transmitted from multiple terminal devices (not shown) are input to the multiplexing device 11 of the transmitting device 10. The multiplexing device 11 multiplexes the input optical data signals to generate an optical multiplexed data signal, and outputs the optical multiplexed data signal to the optical modulation device 12. The transmission signal selection unit 121 of the optical modulation device 12 outputs either the optical multiplexed data signal or a training signal to the optical modulation unit 122 in accordance with the control of the control device 30. The optical modulation unit 122 modulates the optical signal output from the transmission signal selection unit 121 and transmits it.
[0016] The receiving device 20 includes an optical demodulator 21 and a demultiplexer 22. The optical demodulator 21 includes a dispersion compensator 211 and a received symbol determiner 212. The receiving device 20 receives an optical signal transmitted from the transmitting device 10. The dispersion compensator 211 performs dispersion compensation processing and modulation dimension separation processing, such as X-Y polarization separation, on the received optical signal. The received symbol determiner 212 performs symbol determination on the optical signal output from the dispersion compensator 211 and outputs the determination result to the demultiplexer 22. The received symbol determiner 212 restores the received symbols to digital data. The received symbol determiner 212 sends the restored data value to the demultiplexer 22, which is paired with the multiplexer 11 in the opposing transmitting device 10. When performing symbol determination, the received symbol determiner 212 uses the results of machine learning performed in advance. The demultiplexer 22 outputs a signal according to the determination result to a path according to the destination.
[0017] The control device 30 controls whether the received symbol decision unit 212 is in a mode in which machine learning is performed (hereinafter referred to as the "training mode") or a mode in which communication is performed using the results of machine learning (hereinafter referred to as the "data decision mode"). When operating the optical communication system 100 in, for example, the training mode, the control device 30 outputs a training control signal to the transmission signal selection unit 121 and the received symbol decision unit 212. The training control signal includes information indicating that communication is to be performed in the training mode and information indicating what symbol signal is to be transmitted as the training signal. The control device 30 may be configured as a device separate from the transmitting device 10 and the receiving device 20 as shown in FIG. 1, or may be configured as a part of the transmitting device 10 or the receiving device 20.
[0018] The control device 30 further transmits to the received symbol decision unit 212 a signal (hereinafter referred to as an "HP setting signal") indicating optimal hyperparameters to be used in the machine learning executed in the received symbol decision unit 212. The optimal hyperparameters designated by the control device 30 are, for example, hyperparameters that cause a decrease in calculation accuracy (e.g., Q value) within a predetermined threshold and reduce the amount of calculation to a threshold or more. The control device 30 may designate optimal hyperparameters for each layer (e.g., for each region selection layer described below), or may designate optimal hyperparameters for an individual circuit (e.g., region selection circuit 42 described below) that performs calculations using machine learning.
[0019] The control device 30 further transmits to the received symbol decision unit 212 a normalization setting signal indicating operation setting values for the normalization process executed in the received symbol decision unit 212. Upon receiving the normalization setting signal, the received symbol decision unit 212 executes normalization processing based on the operation setting values indicated by the received normalization setting signal. Specifically, the received symbol decision unit 212 processes the received training signal or data signal so that the output amplitude value and bias level become those indicated by the normalization setting signal.
[0020] For example, when the machine learning executed in the received symbol decision unit 212 is a support vector machine, a radial basis function is used as the kernel function of the support vector machine, and the input amplitude of a signal to the support vector machine in the first layer is A1, the hyperparameter γ of the first layer is γ1, and the input amplitude of a signal to the support vector machine being set in the second layer or later is An, the hyperparameter γn of the support vector machine being set may be expressed by the following formula: γn = γ1 × (A1 / An)^2
[0021] 2 is a diagram showing a detailed configuration example of the transmission signal selection unit 121. The transmission signal selection unit 121 includes a training control circuit 1211 and a selection switch 1212. The training control circuit 1211 receives a training control signal transmitted from the control device 30. The training control circuit 1211 outputs a signal of a symbol indicated by the received training control signal (hereinafter referred to as a "training signal") to the selection switch 1212. Furthermore, when the training control circuit 1211 receives the training control signal, it outputs a selection control signal indicating that the training signal should be output to the selection switch 1212.
[0022] The selection switch 1212 receives either a training signal or a data signal, or both. The selection switch 1212 has the function of switching between the training signal and the data signal and sending it out. The selection switch 1212 sends out either the training signal or the data signal, whichever is indicated by the selection control signal.
[0023] 3 is a diagram showing a detailed example configuration of the received symbol decision unit 212. The received symbol decision unit 212 includes a training control circuit 2121 and a decision unit 2122. The training control circuit 2121 receives the HP setting signal, the normalized setting signal, and the training control signal transmitted from the control device 30. The training control circuit 2121 outputs the received HP setting signal, the normalized setting signal, and the training control signal including information on the symbol indicated by the training signal to the decision unit 2122.
[0024] The determiner 2122 performs normalization processing on the signal (training signal or data signal) output from the dispersion compensator 211 based on the normalization setting signal, whether in training mode or data judgment mode. When the training control signal is received (when in training mode), the determiner 2122 performs machine learning using the hyperparameters indicated by the HP setting signal, using as teacher data the signal obtained by performing normalization processing on the training signal (hereinafter referred to as the "normalized signal") and information indicating the symbol pattern indicated by the training control signal. That is, the determiner 2122 performs learning processing on the assumption that the pattern of the signal output from the dispersion compensator 211 is the pattern indicated by the training signal. In this learning processing, the determiner 2122 learns compensation parameters for nonlinear distortion. When the training control signal is not received (when in data judgment mode), the determiner 2122 performs normalization processing on the data signal output from the dispersion compensator 211, performs nonlinear distortion compensation based on the results of the machine learning, and performs symbol judgment. The signal obtained by performing the normalization process on the data signal in this manner will be described as one aspect of the normalized signal described above. The determining unit 2122 outputs a restored data value based on the determination result.
[0025] 4 is a diagram showing a detailed example configuration of the determination unit 2122. In machine learning training mode, the transmission signal selection unit 121 notifies the receiving device 20, as a training control signal, that a training signal is being transmitted and the pattern of the training signal being transmitted. The training control circuit 2121 in the receiving device 20 generates a correct label according to the pattern of the training signal.
[0026] The determination unit 2122 includes a plurality of region selection layers 40, a region determination layer 50, and a multiplexing circuit 60. In the example of FIG. 4 , M region selection layers 40 are provided, where M is an integer value of 2 or greater. Each region selection layer 40 includes one or more normalization circuits 41 and region selection circuits 42. The region selection layer 40_1 in the first layer includes one normalization circuit 41 and one region selection circuit 42. Each region selection layer 40 in the second to Mth layers includes normalization circuits 41 and region selection circuits 42 corresponding to the total number of outputs in the region selection circuit 42 of the region selection layer 40 immediately preceding it. The region determination layer 50 includes a plurality of normalization circuits 41 and region determination circuits 51.
[0027] The normalization circuit 41 performs normalization processing on the input signal based on the operation setting value indicated by the normalization setting signal output from the control device 30. The normalization circuit 41 processes, for example, the input signal so that it has the output amplitude value and bias level indicated by the normalization setting signal. Each normalization circuit 41 may perform normalization processing so that the input signal has the same amplitude and zero bias as the input signal (input signal to the normalization circuit 41) of the region selection layer 40_1 of the first layer, for example.
[0028] A correct label is input as a training control signal to each region selection circuit 42 and each region determination circuit 51. The region selection circuit 42 uses the hyperparameters indicated by the HP setting signal to set the internal variables of the machine learning circuit so as to classify the received symbols according to the correct label. The region selection circuit 42 outputs the received symbols to the corresponding normalization circuit 41 in the next region selection layer 40 according to the correct label. Similarly, in the next region selection layer 40, the normalization circuit 41 performs normalization processing on the received symbols, and the region selection circuit 42 sets the internal variables of the machine learning circuit and outputs the received symbols to the corresponding normalization circuit 41 in the next region selection layer.
[0029] The number of layers in the determination unit 2122 and the number of branches in each region selection circuit 42 are set in advance so that each region contains one type of symbol as a result of the determination by each region determination circuit 51 in the final layer, region determination layer 50. By inputting training signals one after another in this way and learning all paths, the internal node states of the machine learning circuits provided in all region selection circuits 42 and region determination circuits 51 are learned and determined.
[0030] In the data decision mode, the normalization circuit 41 of the first region selection layer 40_1 receives the received symbol and performs normalization processing on the received symbol. The region selection circuit 42 then uses the normalized received symbol to perform region selection in accordance with the machine learning results. The region selection circuit 42 outputs the received symbol to the corresponding normalization circuit 41 of the next region selection layer 40 (40_2). Similarly, in the region selection layer 40_2, the normalization circuit 41 performs normalization processing on the received symbol, and the region selection circuit 42 uses the normalized received symbol to output the received symbol to the corresponding normalization circuit 41 of the next region selection layer 40 (40_3) in accordance with the machine learning results. The output of the region decision layer 50, which is the final layer, determines a single type of symbol. The multiplexing circuit 60 receives a decision result from only one region decision circuit 51 of the region decision layer 50 via a connection. The number of the connection indicates the region for which the last decision was made. The input value from that connection indicates the final decision result. The data value represented by the received symbol is uniquely determined based on the number of the connection and the decision result. The multiplexing circuit 60 outputs the data value as the recovered data value.
[0031] Next, the processing in each layer will be further described. The first layer, the region selection layer 40_1, includes one normalization circuit 41 and a region selection circuit 42. The processing of the normalization circuit 41 has been described above, so a repeated description will be avoided. In data determination mode, the region selection circuit 42 classifies the received symbol into one of L(1,1) regions using the machine learning results. The region selection circuit 42 outputs the received symbol directly only to the normalization circuit 41 in the next layer (second layer) that is responsible for the corresponding region. Here, L(m,n) is the number of output connections of the nth region selection circuit constituting the mth layer. Furthermore, if the total number of output connections from the mth layer is LS(m), the following equation 1 holds.
[0032]
[0033] The region selection layer 40_2 in the second layer includes L(1,1) normalization circuits 41 and region selection circuits 42. Of these multiple normalization circuits 41 and region selection circuits 42, only those normalization circuits 41 and region selection circuits 42 that receive input of a received symbol from the first layer operate. For example, if only the nth normalization circuit 41 and region selection circuit 42 in the second layer receives input, those normalization circuits 41 and region selection circuits 42 perform normalization processing and classify the received symbol into one of the L(2,n) regions using the machine learning results. The region selection circuit 42 outputs the received symbol as is only to the normalization circuit 41 and region selection circuit 42 in the next layer (third layer) that is responsible for the corresponding region.
[0034] Similarly, region selection proceeds while sequentially sending received symbols to the next layer, and in the Mth layer, the following equation 2 holds.
[0035]
[0036] The output of the Mth region selection layer 40_M is input to the region determination layer 50. The region determination layer 50 includes LS(M) normalization circuits 41 and region determination circuits 51. Only the normalization circuit 41 and region determination circuit 51 that receive an input of a received symbol from the Mth region selection layer 40_M operate. When only the nth normalization circuit 41 and region determination circuit 51 receive an input, the normalization circuit 41 performs normalization processing, and the region determination circuit 51 classifies the received symbol into one of L(x, n) regions using the machine learning results, and outputs only the determination result to the downstream multiplexing circuit 60. Here, the region determination layer 50 is represented by the symbol "x".
[0037] The multiplexing circuit 60 outputs a restored data value that is uniquely determined from the position (number) of the area determination circuit 51 that has received the input and its value.
[0038] The region selection circuit 42 and region determination circuit 51 of each layer described above may be configured to classify received symbols into one of L(m, n) regions through machine learning using a neural network with specified hyperparameters. The region selection circuit 42 and region determination circuit 51 of each layer described above may be configured to classify received symbols into one of L(m, n) regions through one-versus-other machine learning using a support vector machine with specified hyperparameters. Note that both m and n are integers greater than or equal to 1.
[0039] 5 is a diagram showing a specific example of the configuration of the normalization circuit 41. The normalization circuit 41 performs normalization processing on the input signal in accordance with the operation setting value indicated by the normalization setting signal. The normalization circuit 41 outputs the normalized signal generated by performing the normalization processing.
[0040] 6 is a diagram showing a specific example of the configuration of the region selection circuit 42. The region selection circuit 42 includes a bifurcating circuit 421, a region determination circuit 422, and a 1×ns switch 423. The bifurcating circuit 421 bifurcates the normalized training signal or data signal into two, and outputs one to the region determination circuit 422 and the other to the 1×ns switch 423. The region determination circuit 422 is a machine learning circuit. ns=L(m, n).
[0041] First, the operation in training mode will be described. The region determination circuit 422 identifies whether the input signal is a training signal or a data signal based on the training control signal. If the input signal is a training signal, the correct label is also notified based on the training control signal. The region determination circuit 422 performs learning through machine learning using a neural network (hereinafter referred to as "NN"), SVM, or the like. At this time, the region determination circuit 422 performs learning using hyperparameters specified by the HP setting signal. Furthermore, the region determination circuit 422 determines, based on the training control signal, to which region selection circuit 42 or region determination circuit 51 in the subsequent layer (region selection layer 40 or region determination layer 50) the input signal should be output. The region determination circuit 422 outputs a control signal to the 1xns switch 423 based on the determination.
[0042] Next, the operation in the data determination mode will be described. Based on the results of machine learning performed in advance, the region determination circuit 422 determines to which normalization circuit 41 in the subsequent layer (region selection layer 40 or region determination layer 50) the input signal should be output. In accordance with the determination, the region determination circuit 422 outputs a control signal to the 1xns switch 423. The 1xns switch 423 switches the output destination of the received symbol to one of 1 to ns, depending on the output of the region determination circuit 422.
[0043] 7 is a diagram showing an example of the configuration of a region determination circuit 51 using an SVM. The region determination circuit 51 constitutes part of the region selection circuit 42 (region determination circuit 422), and is also used in the region determination layer 50. A training signal or a normalized signal of a data signal is input to the region determination circuit 51. The region determination circuit 51 includes an SVM circuit 511. The region determination circuit 51 may be configured as a single SVM circuit 511 itself. The region determination circuit 51 may also be configured as a round-robin determination circuit of a plurality of SVM circuits 511.
[0044] The training control signal specifies whether the input signal is a training signal or a data signal. If the input signal is a training signal, its correct label is also notified at the same time. In this case, the SVM circuit 511 performs machine learning using the SVM using the training signal and the correct label. At this time, the SVM circuit 511 performs learning using hyperparameters specified by the HP setting signal. If the input signal is a data signal, the SVM circuit 511 makes a judgment using the results of the machine learning and outputs the judgment result. If the area judgment circuit 51 is part of the area selection circuit 42 (area judgment circuit 422), the judgment result is used as a control signal for the 1xns switch 423.
[0045] 8 is a diagram showing an example of the configuration of a region determination circuit 51 using an NN. The region determination circuit 51 constitutes part of the region selection circuit 42 (region determination circuit 422), and is also used in the region determination layer 50. A training signal or a normalized signal of a data signal is input to the region determination circuit 51. The region determination circuit 51 includes an NN circuit 512. The region determination circuit 51 may be configured as a single NN circuit 512 itself.
[0046] The training control signal specifies whether the input signal is a training signal or a data signal. If the input signal is a training signal, its correct label is also notified at the same time. In this case, the NN circuit 512 performs machine learning using a NN using the training signal and the correct label. At this time, the NN circuit 512 performs learning using hyperparameters specified by the HP setting signal. If the input signal is a data signal, the NN circuit 512 makes a judgment using the results of the machine learning and outputs the judgment result. If the area judgment circuit 51 is part of the area selection circuit 42 (area judgment circuit 422), the judgment result is used as a control signal for the 1xns switch 423.
[0047] Fig. 9 is a diagram showing an example of a configuration in which a 16QAM signal is determined by combining three layers of two-branch region selection circuits 42. Fig. 10 is a diagram showing an outline of the processing of a configuration in which a 16QAM signal is determined by combining three layers of two-branch region selection circuits 42. In the region selection layer 40_1, as shown in the first layer of Fig. 10, the 16QAM constellation is divided into two regions, left and right, and machine learning training is performed.
[0048] In the region selection layer 40_2, as shown in the second layer of FIG. 10 , each of the left and right regions is divided into upper and lower regions, and machine learning training is performed. This region division is repeated in the same manner until the region is divided into 16 regions. That is, in the specific example of FIG. 9 , L(m, n) = 2 is set in all region selection circuits 42, and L(x, n) = 2 is set in all region determination circuits 51. Furthermore, the region determination circuit 51 classifies the received symbol into one of two regions by one-to-one machine learning using a support vector machine.
[0049] In this embodiment, the number of required machine learning circuits (the number of area determination circuits 422 and area determination circuits 51) is 15, which is an increase compared to the four required in the conventional method. However, because the number of internal variables in each machine learning circuit is reduced, the total number of internal variables remains approximately the same. On the other hand, when training (machine learning processing) is completed and a symbol is determined, for example, in the first layer (area selection layer 40_1), if the received symbol is determined to belong to the right area, calculations for the left area are not performed in subsequent calculations. This reduces the amount of calculation and the calculation execution time compared to the conventional method. In this example, it is assumed that the area is repeatedly divided into two parts, but multiple divisions of more than two parts may be used, and the number of divisions may vary for each layer. The number of divisions may also vary between different areas within the same layer.
[0050] FIG. 11 illustrates the principle of SVM. A boundary line is drawn that allows discrimination between received symbols indicated by circles and received symbols indicated by triangles with as low an error rate as possible. During machine learning training for the SVM, margin regions are set from the boundary line to the inside of each of the circles and triangles. Only received symbols outside the margin region, as viewed from each region, are used for discrimination. Hereinafter, these received symbols are referred to as support vectors. Received symbols inside the margin region are not used for discrimination. Therefore, if the number of received symbols outside the margin region is reduced during training, the amount of computation required for data discrimination can be reduced. By performing this process, the number of matrix coefficients that can be ignored for discrimination increases, reducing the number of parameters and reducing the amount of computation required for data discrimination. In the case of a neural network, the amount of computation required for data discrimination can be reduced by deliberately deleting internal variables close to zero.
[0051] FIG. 12 shows the results of simulating the operation of the conventional method. FIG. 13 shows the results of simulating the operation of this embodiment (proposed method). FIGS. 12 and 13 plot the positions of support vectors learned and saved in 64QAM classification. In the proposed method, the output of the normalization circuit 41 is set to have the same amplitude and zero bias as the input signal of the first layer. An SVM is used in the region determination circuit 422 (51). The hyperparameter value is the same for all circuits, γ = 0.1. FIG. 12 shows the learning results of the first layer using the conventional method, and FIG. 13 shows the learning results of the proposed method. In this case, since the output of the normalization circuit 41 is equal to the input, the proposed method and the conventional method produce the same results. The number of support vectors (nsv) is also the same, nsv = 302.
[0052] Figure 14 shows the results of a simulation of the operation of the conventional method. Figure 15 shows the results of a simulation of the operation of this embodiment (proposed method). Figures 14 and 15 plot the positions of support vectors learned and saved in 64QAM classification. In the proposed method, the output of the normalization circuit 41 is set to have the same amplitude and zero bias as the input signal of the first layer. An SVM is used in the region determination circuit 422 (51). The hyperparameter value is the same for all circuits, γ = 0.1. Figure 14 shows the learning results of the sixth layer using the conventional method, and Figure 15 shows the learning results of the proposed method. In this case, due to the operation of the normalization circuit 41, the symbols are widely distributed around the origin in the proposed method. The number of support vectors (nsv) in the conventional method is nsv = 150, while in the proposed method it is reduced to nsv = 28, approximately one-fifth of the number in the conventional method.
[0053] Figures 16, 17, and 18 show simulation results comparing the operation of the conventional and proposed methods. A 64QAM signal was used in each simulation. In the proposed method, the output of the normalization circuit 41 was set to have the same amplitude and zero bias as the input signal to the first layer. An SVM was used in the region determination circuit 422 (51). The hyperparameter value was the same for all circuits, γ = 0.1. The results shown in Figure 16 are the Q-values of the classification results. At γ = 0.1, the Q-value of the proposed method was approximately 0.08 dB lower than that of the conventional method. The results shown in Figure 17 are the total number of support vectors learned and saved. In the proposed method, this is one-third of that of the conventional method. The results shown in Figure 18 are the average number of support vectors used during classification. In the proposed method, the number was reduced by approximately 40% compared to the conventional method. As such, in this embodiment, although there is a slight degradation in the Q-value, the reduction in support vectors enables significant reductions in memory and computational costs.
[0054] 19 is a diagram showing a first modified example. As shown in FIG. 19, the values of the hyperparameters may be defined as the same values (fixed values) in all circuits. In this case, the normalization circuit 41 is used in the same manner as in the configuration described above.
[0055] 20 is a diagram showing a second modified example. As shown in FIG. 20, the normalization circuit 41 may not be provided. In this case, the values of the hyperparameters are used in the same manner as in the above-described configuration. For example, the values of the hyperparameters may be set to the reciprocal of the amplification factor.
[0056] Each circuit in this embodiment may be configured using a processor such as a CPU (Central Processing Unit) and a memory. All or part of each circuit in this embodiment may be realized using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array). The above program may be recorded on a computer-readable recording medium. Examples of computer-readable recording media include portable media such as flexible disks, magneto-optical disks, ROMs, CD-ROMs, and semiconductor storage devices (e.g., SSDs: Solid State Drives), as well as storage devices such as hard disks and semiconductor storage devices built into computer systems. The above program may be transmitted via a telecommunications line.
[0057] Although an embodiment of the present invention has been described above in detail with reference to the drawings, the specific configuration is not limited to this embodiment, and includes designs within the scope of the gist of the present invention.
[0058] The present invention is applicable to optical communications.
[0059] 100...optical communication system, 10...transmitting device, 11...multiplexing device, 12...optical modulation device, 121...transmitting signal selection unit, 122...optical modulation unit, 20...receiving device, 21...optical demodulation device, 22...demultiplexing device, 211...dispersion compensation unit, 212...received symbol judgment unit, 30...control device, 1211...training control circuit, 1212...selection switch, 2121...training control circuit, 2122...judgment unit, 40...region selection layer, 41...normalization circuit, 42...region selection circuit, 50...region judgment layer, 51...region judgment circuit, 60...multiplexing circuit, 421...two-branching circuit, 422...region judgment circuit, 423...1xns switch, 511...SVM circuit, 512...NN circuit
Claims
1. A received symbol decision unit performs symbol decision on a received optical signal and restores data based on a result of the symbol decision, the received symbol decision unit comprises region selection layers from a first layer to an Mth layer (M is an integer equal to or greater than 1) and one region decision layer, the mth layer (m is an integer equal to or greater than 1 and equal to or less than M) comprises normalization circuits and region selection circuits in the same number as the number of region divisions in the m-1th layer region selection layer (when m=1, it is "1"), the normalization circuits amplify the amplitude of the received multi-level signal, the mth layer region selection circuit selects a region to which the symbol of the received multi-level signal belongs from among a plurality of regions according to the number of layers (1 to M) based on a result of machine learning for the received multi-level signal, and outputs the selected region to a next layer, the mth layer region selection circuit is processed by a region selection circuit according to the region selected by the m-1th layer region selection circuit, the region decision layer comprises region decision circuits according to the total number of outputs of the region selection circuits in the Mth layer region selection layer, The area determination circuit selects an area to which the symbol of the multi-value signal received from the Mth area selection layer belongs based on a result of machine learning, and outputs a determination result indicating the symbol of the multi-value signal according to the selected area.
2. A received symbol decision unit that performs symbol decision on a multi-level signal for a received optical signal and restores data based on a result of the symbol decision, the received symbol decision unit includes region selection layers from a first layer to an Mth layer (M is an integer equal to or greater than 1) and one region decision layer, the mth layer (m is an integer equal to or greater than 1 and equal to or less than M) includes region selection circuits in the same number as the number of region divisions in the m-1th layer region selection layer (when m=1, it is "1"), the mth layer region selection circuit selects a region to which the symbol of the received multi-level signal belongs from among a plurality of regions according to the number of layers (1 to M) based on a result of machine learning for the received multi-level signal, and outputs the selected region to a next layer, the mth layer region selection circuit performs processing according to the region selected by the m-1th layer region selection circuit, the region decision layer includes region decision circuits according to the total number of outputs of the region selection circuits in the Mth layer region selection layer, an optical transmission / reception device, wherein the area determination circuit selects an area to which the symbol of the multi-value signal received from the Mth area selection layer belongs based on a result of machine learning, and outputs a determination result indicating the symbol of the multi-value signal according to the selected area, and the result of machine learning is a result obtained by executing machine learning using appropriate hyper-parameter values in each layer.
3. The optical transmission / reception device of claim 1 or 2, wherein, when operating in a training mode, the area selection circuit and the area judgment circuit learn which area to select from the multiple areas in accordance with the received multi-value signal by performing machine learning using the received multi-value signal and a correct answer label indicating the symbol possessed by the multi-value signal, and when operating in a data judgment mode, the area selection circuit and the area judgment circuit select the area for the received multi-value signal based on the learning result obtained in the training mode.
4. An optical transmitting / receiving device as described in claim 1 or 2, wherein the machine learning in the area selection circuit and the area judgment circuit is machine learning using a neural network, and the area selection circuit and the area judgment circuit classify the symbol into one of a plurality of areas.
5. An optical transmission / reception device as described in claim 1 or 2, wherein the machine learning in the area selection circuit and the area judgment circuit is machine learning using a support vector machine, and the area selection circuit and the area judgment circuit classify the symbol into one of a plurality of areas by one-to-other type machine learning.
6. A computer program for causing a computer to function as the optical transmitter-receiver according to claim 1 or 2.
Citation Information
Patent Citations
Symbol determination device of input signal, method, and program
JP2021034816A
Optical fiber nonlinearity compensation using neural networks
JP2021513251A
Optical analog matrix multiplier for optical neural networks
US20210135764A1
