Modeling device, modeling method, and program

Abstract

A modeling device (1) comprises: a neural network unit (12) that, when a signal is input, inputs an input signal into a neural network that outputs a signal representing a common behavior that is not dependent on input power in a target electrical circuit; and a polynomial calculation unit (13) that calculates, through calculation using a polynomial for the input signal, a signal representing a behavior that changes in accordance with the input power in the target electrical circuit.

Description

Modeling device, modeling method, and program

[0001] The present disclosure relates to a modeling device, a modeling method, and a program.

[0002] In the development of electric circuits such as rectifier circuits and amplifiers, simulation analysis of the circuit configuration is performed. A behavioral model that models the behavior of the electric circuit is used in the simulation analysis of the circuit configuration. For example, Non-Patent Document 1 describes a technique for generating a behavioral model of a rectifier circuit. A rectifier circuit is a key component of wireless power transmission technology, and is a circuit that converts high-frequency power indicated by an input signal into direct-current power. The behavioral model of the rectifier circuit described in Non-Patent Document 1 uses the input power indicated by a high-frequency sinusoidal signal as an explanatory variable and the direct-current power as an explained variable.

[0003] X. Xu, et al “Simultaneous information and power transfer under a non-linear RF energy harvesting model,” 2017 IEEE International Conference on Communications Workshops (ICC Workshops), Paris, France, 2017, pp. 179-184, doi: 10.1109 / ICCW. 2017.7962654.

[0004] The conventional technology described in Non-Patent Document 1 has a problem in that it is not possible to accurately model the behavior of an electric circuit. For example, a modulated signal is a signal whose power changes from moment to moment. Therefore, when the input signal is a modulated signal, it may be difficult to accurately model the change in circuit behavior in response to the moment-to-moment changing input power. Furthermore, when modeling a rectifier circuit, it is necessary to model the behavior taking into account the influence of the memory effect of the diodes included in the rectifier circuit. However, the conventional technology described in Non-Patent Document 1 uses a sine wave input signal that does not take into account the memory effect of the diodes, and therefore it is not possible to accurately model the behavior of the rectifier circuit.

[0005] The present disclosure is intended to solve the above-mentioned problems, and aims to provide a modeling device that can accurately model the behavior of an electric circuit.

[0006] The modeling device according to the present disclosure includes a neural network (hereinafter referred to as NN) unit that inputs an input signal to the NN, which outputs a signal that represents common behavior independent of the input power in the target electrical circuit when a signal is input, and a polynomial calculation unit that calculates a signal that represents behavior that changes depending on the input power in the target electrical circuit by performing a polynomial calculation on the input signal.

[0007] According to the present disclosure, a modeling device according to the present disclosure uses a neural network to model common behavior that is independent of the input power in a target electric circuit, and further uses a polynomial to model behavior that changes depending on the input power in the target electric circuit, thereby enabling the modeling device to accurately model the behavior of the electric circuit.

[0008] 1 is a block diagram showing an example of the configuration of a modeling device according to a first embodiment. FIG. 2 is a flowchart showing a modeling method according to the first embodiment. FIG. 3 is a flowchart showing a learning method for a neural network. FIG. 4 is a diagram showing examples of raw data, training data, and test data. FIG. 5 is a waveform diagram showing an output voltage waveform of a rectifier circuit and an output waveform of a neural network. FIG. 6 is a waveform diagram showing an output voltage waveform of a rectifier circuit and an output waveform of a modeling device according to the first embodiment. FIG. 7A and FIG. 7B are diagrams showing a modified example (1) of input data according to the first embodiment. FIG. 8A is a diagram showing a modified example (2) of input data according to the first embodiment, and FIG. 8B is a diagram showing data output by a neural network to which the data of FIG. 8A has been input. FIG. 9 is a diagram showing an example of a trend of time-series data. FIG. 10 is a diagram showing a modified example (3) of input data according to the first embodiment. FIG. 11A and FIG. 11B are diagrams showing a hardware configuration realizing the functions of the modeling device according to the first embodiment. FIG. 11B is a block diagram showing an example of the configuration of a modeling device according to a second embodiment. FIG. 12 is a block diagram showing an example of the configuration of a modeling device according to a third embodiment.

[0009] First Embodiment. The modeling device according to the first embodiment models the behavior of a target electrical circuit based on an input signal and has an architecture combining a "neuronal network" and a "polynomial." A neural network (NN) has the characteristic of being able to model any continuous function (the universality theorem). On the other hand, when the average power input to a rectifier circuit changes over time, the behavior of the rectifier circuit changes. To model such changing behavior using a neural network, a neural network corresponding to each input power is constructed. In this case, a huge number of coefficients are required, and it takes a long time to calculate these coefficients. Therefore, the modeling device according to the first embodiment uses a neural network to model the common behavior of the target electrical circuit that is independent of the input power, and uses a polynomial to model the behavior that changes depending on the input power of the target electrical circuit. This allows the modeling device according to the first embodiment to accurately model the behavior of the electrical circuit.

[0010] (Target Electrical Circuit) The target electrical circuit is a circuit that exhibits behavior that varies depending on the input power. Examples include rectifier circuits and amplifiers. Rectifier circuits convert AC signals into DC signals. Rectifier circuits include half-wave rectifier circuits, full-wave rectifier circuits, and bridge rectifier circuits. Half-wave rectifier circuits pass only the positive half-cycle of an AC input signal and cut off the negative half-cycle, thereby outputting a pulsed DC signal. Half-wave rectifier circuits include diodes that conduct only when the input signal is positive and cut off when the input signal is negative. Full-wave rectifier circuits convert both the positive and negative half-cycles of an AC input signal into DC signals of the same polarity. Full-wave rectifier circuits perform full-wave rectification using a center-tapped transformer. Bridge rectifier circuits convert all half-cycles of an AC input signal using four diodes connected in a bridge configuration. Amplifiers include operational amplifiers (OPAs) and transistor amplifiers. OpAps amplify the difference between input signals. A transistor amplifier amplifies an input signal using a field effect transistor, etc. By appropriately setting the operating point of the transistor, linear amplification is possible.

[0011] Common behavior independent of input power: Common behavior in an electrical circuit is a characteristic or behavior that follows a consistent rule or law regardless of the input power to the circuit. Examples include Ohm's law, Kirchhoff's laws, the characteristics of passive components, and the frequency response of filters. Specifically, Ohm's law states that the current through a resistor is proportional to the voltage across it, and it holds true regardless of the input power to the circuit. Kirchhoff's laws include the current law, which states that at any node in a circuit, the sum of the currents flowing into that node is equal to the sum of the currents flowing out of that node, and the voltage law, which states that for any closed loop in a circuit, the sum of all voltages within that loop is zero. The basic characteristics of passive components such as resistors, capacitors, and inductors are also independent of input power. The frequency response of filter circuits (e.g., low-pass filters, high-pass filters, and band-pass filters) remains constant regardless of the power of the input signal. Common behavior independent of input power can be modeled using any continuous function.

[0012] (Behavior that Changes with Input Power) The behavior of electrical circuits that changes with input power varies depending on the circuit. For example, a rectifier circuit converts an AC input signal into a DC signal, and its output changes with the input power. More specifically, a half-wave rectifier circuit passes only the positive half-cycle of the AC input signal and cuts the negative half-cycle. Therefore, when the amplitude of the input voltage changes, the amplitude of the output DC voltage also changes accordingly. In a full-wave rectifier circuit, as the input voltage increases, the output DC voltage also increases accordingly. Similarly, an amplifier amplifies the amplitude of the input signal, and its output changes with the input power. An operational amplifier amplifies small changes in the input signal, and as the input voltage increases, the output voltage also increases. A transistor amplifier uses transistors to amplify a small input signal to a large output signal. As the amplitude of the input signal increases, the amplitude of the output signal also increases proportionally.

[0013] (Diode Memory Effect) To explain the behavior of a rectifier circuit, which changes depending on the input power, it is necessary to consider the diode memory effect. The memory effect is also known as the hysteresis effect. Specifically, the diode memory effect is a phenomenon in which a previously applied current or previously applied voltage affects the current operation. The memory effect has a significant impact, for example, in high-speed switching or high-frequency operation. When a diode switches from forward to reverse direction, the reverse recovery time, during which current temporarily flows in the reverse direction, can cause ripple or noise in the output waveform. Rectifier circuits used in wireless power transmission technology may be input with a modulated signal whose voltage changes from moment to moment. When a modulated signal is input to a rectifier circuit, the output voltage also changes in response to changes in the input voltage of the modulated signal, and this output voltage is affected by the memory effect.

[0014] (Basic Configuration of Modeling Device) Fig. 1 is a block diagram showing an example of the configuration of a modeling device 1 according to the first embodiment. In Fig. 1, the modeling device 1 is a device for modeling the behavior of a target electric circuit A, which is a rectifier circuit, and includes an acquisition unit 11, an NN unit 12, and a polynomial calculation unit 13. The rectifier circuit receives an AC input signal x ｔ is input, and the DC output signal y ｔ where t is time. The modeling device 1 is implemented by a computer. A memory provided in the computer stores programs constituting information processing applications for realizing the functions of the acquisition unit 11, the NN unit 12, and the polynomial calculation unit 13. A processor provided in the computer reads the information processing application from the memory and executes the information processing application, thereby realizing the functions of the acquisition unit 11, the NN unit 12, and the polynomial calculation unit 13.

[0015] (Acquisition Unit) The acquisition unit 11 acquires an input signal. The input signal is time-series data of a signal supplied to an electric circuit. The time-series data may be a voltage waveform or a power waveform applied to the electric circuit. The signal value at each time is a voltage value or a current value. In FIG. 1, the input signal is represented by x ｔ , xｔ－１ , ..., x ｔ－ｎ＋１ , x ｔ－ｎ where n is an integer. For example, if the modeling device 1 includes a communication unit (not shown in FIG. 1 ) and is capable of communicating with an external device (not shown in FIG. 1 ) via this communication unit, the acquisition unit 11 acquires an input signal from the external device by controlling the communication unit. The communication unit communicates with the external device via a communication network.

[0016] Furthermore, the communication method used for communication with the external device may be, for example, Bluetooth (registered trademark), Wi-Fi, Zigbee (registered trademark), LoRa (registered trademark), or NFC (Near Field Communication). The communication unit is, for example, a wireless communication module compatible with any of these communication methods. The acquisition unit 11 controls the communication unit to execute procedures such as pairing, network connection, and address assignment, thereby establishing communication with the external device. Once communication is established, the communication unit requests an input signal related to the target electrical circuit from the external device, and the external device returns the requested signal. The information received by the communication unit is acquired by the acquisition unit 11.

[0017] 1, in which an input signal for each target electric circuit is stored, the acquiring unit 11 may read and acquire the input signal from the storage unit. In this case, the modeling device 1 does not need to include the communication unit.

[0018] (NN Unit) The NN unit 12 inputs the input signal acquired by the acquisition unit 11 to a NN that, when a signal is input, outputs a signal representing common behavior independent of the input power in the target electric circuit A. The NN takes in the input signal as a feature and is trained to output a signal representing common behavior independent of fluctuations in the input power in the target electric circuit A. The common behavior independent of fluctuations in the input power is modeled by a feature of the NN known as the universality theorem, which allows modeling of any continuous function.

[0019] The NN is assumed to be included in the modeling device 1, but may also be included in an external device that can be connected for communication with the modeling device 1. The NN is composed of data including nodes, edges, and their weight information. This data may be included in the modeling device 1 or in the external device. For example, if the modeling device 1 is equipped with a communication unit not shown in FIG. 1 and is capable of connecting for communication with an external device not shown in FIG. 1 via this communication unit, the NN unit 12 controls the communication unit to input an input signal to the NN included in the external device. An output signal from the NN is received by the communication unit and output to the polynomial calculation unit 13.

[0020] For example, if the target electric circuit A is a rectifier circuit, the NN receives time-series data of a modulated signal to be input to the actual rectifier circuit. ｔ＿ｎｎ is a signal that represents a common behavior independent of the input power in the rectifier circuit. ｔ＿ｎｎ Although the modeling accuracy is not sufficient, the input signal x ｔ and the output signal y ｔ＿ｎｎ are assumed to be real numbers, but may also be complex numbers.

[0021] The NN unit 12 may have the functions of the acquisition unit 11. In this case, the NN unit 12 acquires an input signal and inputs the acquired input signal to the NN, so the modeling device 1 does not need to be equipped with the acquisition unit 11.

[0022] (Polynomial Calculation Unit) The polynomial calculation unit 13 calculates a signal representing behavior that changes depending on the input power in the target electrical circuit by performing a polynomial calculation on the input signal output from the NN unit 12. For example, polynomials (1), (2), and (3) are set in the polynomial calculation unit 13. The polynomial calculation unit 13 selects one of these polynomials to use to calculate an output signal depending on the input power. In FIG. 1 , polynomial (1) is a high-power polynomial that is selected when the input power is in the high-power range. Polynomial (1) has coefficients set therein for calculating an output value obtained by the behavior of the target electrical circuit A that changes depending on the input power in the high-power range. The high-power range is, for example, 100 mV. Polynomial (2) is a medium-power polynomial that is selected when the input power is in the medium-power range. Polynomial (2) has coefficients set therein for calculating an output value obtained by the behavior of target electrical circuit A, which changes in response to input power within a medium power range. The medium power range is, for example, 50 mV. Polynomial (3) is a polynomial for low power, and is selected when the input power is within a low power range. Polynomial (3) has coefficients set therein for calculating an output value obtained by the behavior of target electrical circuit A, which changes in response to input power within a low power range. The low power range is, for example, 10 mV.

[0023] The polynomial calculation unit 13 calculates a signal by performing a calculation using a polynomial on the signal output from the NN unit 12. For example, the polynomial calculation unit 13 selects a polynomial according to each input power from among polynomials (1), (2), and (3), and calculates the signal y obtained from the NN unit 12 for the selected polynomial. ｔ＿ｎｎ The polynomial calculation unit 13 calculates an output value y ｔ For example, the polynomial is expressed by the following formula (1): f(x)=y ｔ = a ０ +a １ x + a ２ x ２ +...+a ｎ x ｎ (1)

[0024] The polynomial may be not only the above formula (1) but also, for example, a Volterra series polynomial. ｔ＿ｎｎ When inputting y into the polynomial operation 13, if the polynomial is a function of complex numbers, the output calculated using the polynomial will also be a complex number. If the output calculated using the polynomial is a complex number, the output of the polynomial may be converted to a real number (L2 norm) and the behavior of the rectifier circuit may be modeled using a real number signal. Note that the output y of the NN ｔ＿ｎｎ When is a real number, the output of the polynomial is also a real number. Regarding which polynomial is selected from the multiple polynomials set in the polynomial calculation unit 13 located downstream of the NN unit 12, a polynomial that corresponds one-to-one with the input power to the rectifier circuit is selected.

[0025] Next, a modeling method according to the first embodiment will be described. FIG. 2 is a flowchart illustrating the modeling method according to the first embodiment, showing a series of operations performed by the modeling device 1. The acquisition unit 11 acquires time-series data, which is an input signal (step ST1). The NN unit 12 inputs the input signal acquired by the acquisition unit 11 to a NN that has been trained to output a signal representing a common behavior independent of the input power in the target electric circuit A when the signal is input (step ST2). The polynomial calculation unit 13 calculates a signal representing behavior that changes depending on the input power in the target electric circuit A by performing a polynomial calculation on the signal output from the NN (step ST3). By performing this method with the modeling device 1, the behavior of the electric circuit can be accurately modeled. Note that if the NN unit 12 has the functions of the acquisition unit 11, the processing of step ST1 is included in step ST2 and is executed by the NN unit 12. Therefore, step ST1 is omitted.

[0026] Next, the learning of the NN in the first embodiment will be described. FIG. 3 is a flowchart showing the learning method of the NN. The following description will be given taking the case where the target electric circuit A is a rectifier circuit as an example. The acquisition unit 11 acquires the time waveforms of the modulated signal input to the rectifier circuit and the signal output from the rectifier circuit for each input power using an oscilloscope or the like (step ST1A). The same number of pieces of time-series data of the input signal (input data) and the output signal (output data) are acquired. However, it is not necessary to acquire the same number of pieces of input / output data for each input power. For example, even if 10,000 pieces of input / output data are acquired when the input power is 1 W, it is possible that 20,000 pieces of input / output data are acquired when the input power is 5 W.

[0027] FIG. 4 is a diagram illustrating examples of raw data, training data, and test data. In the example illustrated in FIG. 4, the acquisition unit 11 acquires input (1) and input (3), which are modulated signals input to the rectifier circuit, and acquires output (1) and output (3), which are signals output by the rectifier circuit. Input (1) is input data when the input power is in the high power range, and input (3) is input data when the input power is in the low power range. Output (1) is output data output from the rectifier circuit when the input power is in the high power range, and output (3) is output data output from the rectifier circuit when the input power is in the low power range. Note that a sufficient number of data points is required to fully learn the characteristics of the input signal (modulated signal). Therefore, if the number of data points is extremely small, such as 10 points, additional data acquisition is performed.

[0028] The acquisition unit 11 divides the acquired input / output data of the rectifier circuit into an arbitrary number of pieces (step ST2A). For example, 10,000 pieces of input / output data acquired when the input power is 1 W are divided into 6,000 pieces of input / output data and 4,000 pieces of input / output data. Furthermore, 20,000 pieces of input / output data acquired when the input power is 5 W are divided into 3,000 pieces of input / output data and 7,000 pieces of input / output data. The processing of step ST2A may be performed by the NN unit 12.

[0029] In FIG. 4, the time series data X １ ~X ｔ is the time series data X １ ~X ｍ and time series data X ｍ＋１ ~X ｔ The time series data x of input (3) is divided into １ ~x ｔ is the time series data x １ ~x ｎ and time series data x ｎ＋１ ~x ｔ In addition, the time series data Y １ ~Y ｔ is the time series data Y １ ~Y ｍ and time series data Y ｍ＋１ ~Y ｔ The time series data y of output (3) is divided into １ ~y ｔ is the time series data y １ ~y ｎ and time series data y ｎ＋１ ~y ｔ and is divided into

[0030] Next, the acquisition unit 11 combines the divided data (step ST3A). Here, input data is combined with input data, and output data is combined with output data. For example, 6,000 pieces of input data acquired when the input power is 1 W are combined with 3,000 pieces of input data acquired when the input power is 5 W (input data (1)). 6,000 pieces of output data acquired when the input power is 1 W are combined with 3,000 pieces of output data acquired when the input power is 5 W (output data (1)). 4,000 pieces of input data acquired when the input power is 1 W are combined with 7,000 pieces of input data acquired when the input power is 5 W (input data (2)). 4,000 pieces of output data acquired when the input power is 1 W are combined with 7,000 pieces of output data acquired when the input power is 5 W (output data (2)). Note that the processing of step ST3A may be performed by the NN unit 12.

[0031] In FIG. 4, the time series data x divided from the input (3) １ ~x ｎ and the time series data X divided from input (1) １ ~X ｍ The time series data x divided from the input (3) is combined (input data A). ｎ＋１ ~x ｔ and the time series data X divided from input (1) ｍ＋１ ~X ｔ The time series data y divided from the output (3) is combined (input data B). １ ~y ｎ and the time series data Y divided from the output (1) １ ~Y ｍ The time series data y divided from the output (3) is combined (output data A). ｎ＋１ ~y ｔ and the time series data Y divided from the output (1) ｍ＋１ ~Y ｔ are combined (output data B).

[0032] The NN unit 12 uses a signal formed by combining multiple time-series data, which are input / output data of the target electric circuit A at different input powers, as training data to train the NN (step ST4A). For example, the NN uses input data (1) and output data (1) as training data to learn common behavior in the rectifier circuit that is independent of the input power. In the example of FIG. 4, the NN unit 12 inputs input data A and output data A to the NN as training data.

[0033] Next, the NN unit 12 verifies overfitting of the NN using a signal formed by combining time-series data other than the training data (step ST5A). For example, the NN unit 12 verifies overfitting of the NN using input data (2) and output data (2) as test data. In the example of FIG. 4, the NN unit 12 verifies overfitting of the NN using input data B and output data B as test data. Verification of overfitting is performed, for example, by plotting the training error and validation error for each epoch during the training process. When overfitting occurs, the training error continues to decrease while the validation error begins to increase at a certain point, and verification is performed based on this trend.

[0034] If the verification result indicates that overlearning has not occurred, the NN unit 12 determines that modeling of the common behavior of the rectifier circuit has been completed regardless of the input power (step ST6A). Note that, in order to effectively model the memory effect of the rectifier circuit, it is desirable for the NN to have a structure that learns not only current information but also past information. The NN may be, for example, a long short term memory (LSTM) network or a recurrent neural network (RNN).

[0035] (Determination of polynomial coefficients) The polynomial is determined by the signal y ｔ＿ｎｎ and the signal y actually output by the rectifier circuit ｔ The polynomial coefficients are the elements of matrix A, and the signal output from the NN is expressed as vector Y ｔ＿ｎｎ The signal actually output by the rectifier circuit is expressed as vector Y ｔ When expressed as a vector Y ｔ is expressed by the following formula (2): ｔ =AY ｔ＿ｎｎ (2)

[0036] In order to obtain the matrix A, the polynomial calculation unit 13 calculates Y from the right of the above equation (2). ｔ＿ｎｎ Inverse matrix Y ｔ＿ｎｎ －１This makes it possible to model the behavior of circuits that could not be modeled using NN. Note that this matrix A is calculated for each input power of the rectifier circuit to be modeled. In other words, if the input power (average value) is 1 mW to 10 mW and all of these are to be modeled, a matrix A dedicated to each input power must be prepared for each input power.

[0037] (Effects of Modeling) FIG. 5 is a waveform diagram showing the output voltage waveform of the rectifier circuit and the output waveform of the NN. In FIG. 5, waveform B is the output voltage waveform of the signal actually output by the rectifier circuit, and waveform C is the output waveform of the signal output from the NN. As shown in FIG. 5, waveforms B and C differ in many ways, and the behavior of the rectifier circuit cannot be fully modeled using only the NN. FIG. 6 is a waveform diagram showing the output voltage waveform of the rectifier circuit and the output waveform of the modeling device according to embodiment 1. In FIG. 6, waveform B is the output voltage waveform of the signal actually output by the rectifier circuit, and waveform D is the output waveform of the signal output from the polynomial calculation unit 13. As shown in FIG. 6, waveforms B and D almost overlap, and by using polynomial calculation, the behavior of the rectifier circuit can be accurately modeled.

[0038] For example, the distribution (bias) of modulated signals varies depending on the type (modulation method). Furthermore, when a neural network receives a signal with a different signal distribution from the training data, it may output a signal that differs from the expected signal. Therefore, when different types of modulated signals are input to a rectifier circuit (e.g., QPSK, 16QAM, OFDM), it is necessary to model the rectifier circuit taking these factors into account.

[0039] Therefore, the NN unit 12 trains the NN by adding labels corresponding to the type of modulated signal in addition to the time-series data of the modulated signal, which is the signal input to the NN. FIG. 7A shows input data in which the type of modulated signal is not taken into account, and is composed of a combination of modulated signals E and F of different types. For example, the modulation method of modulated signal E is 16QAM, and the modulation method of modulated signal F is OFDM. FIG. 7B shows a modified example (1) of input data in the first embodiment, which, like FIG. 7A, is composed of a combination of modulated signals E and F of different types, and further has a label unique to the type of modulated signal added. The acquisition unit 11 analyzes the modulation method of the modulated signal in the acquired data and assigns a label corresponding to the method obtained by the analysis. The NN unit 12 trains the NN using the labeled training data. This enables the NN to train while taking the type of modulated signal into account.

[0040] FIG. 8A is a diagram showing a modified example (2) of input data in the first embodiment, and is data in which modulated signals having modulation methods (types) of QPSK, 64QAM, 256QAM, and OFDM are combined in a time series. Labels for QPSK, 64QAM, 256QAM, and OFDM are assigned. Note that FIG. 8A shows time-series data in which the modulation methods are QPSK, 64QAM, 256QAM, and OFDM, respectively, but these are merely examples and the order need not be the same. By inputting the input data shown in FIG. 8A into the NN, a common behavior independent of the input power of the target electrical circuit A is modeled, and the NN outputs the data shown in FIG. 8B. Furthermore, the output data shown in FIG. 8B is data in which modulated signals having QPSK, 64QAM, 256QAM, and OFDM are combined in a time series. The output data may be, for example, time-series data only, and may not have labels corresponding to the types of modulated signals.

[0041] Similarly, in Fig. 8, the acquisition unit 11 analyzes the modulation method of the modulated signal in the acquired data and assigns a label according to the method obtained as a result of the analysis. The NN unit 12 trains the NN using the training data to which the labels have been assigned. This allows the NN to train while taking into account the type of modulated signal. Note that although the case where the acquisition unit 11 assigns the labels has been shown, the NN unit 12 may also assign the labels. Furthermore, labels may be assigned to the data in advance.

[0042] Furthermore, the NN unit 12 may train the NN using, as training data, a data set including either the first-order derivative of the time series data input to the target electric circuit A or the second-order derivative of the time series data input to the target electric circuit A, or both. FIG. 9 is a diagram illustrating an example of the trend of the time series data. By first differentiating the time series data, the trend of the time series data can be grasped. For example, as shown in FIG. 9, the first derivative of time series data that is on an increasing trend is a positive value, and the first derivative of time series data that is on a decreasing trend is a negative value. By taking the trend of the time series data into account in this way, the NN can predict the data at time t that indicates the behavior of the target electric circuit A with higher accuracy than when past data is used directly to predict the data at time t that indicates the behavior of the target electric circuit A. Furthermore, by second-order differentiating the time series data, changes in the rate of change of the time series data can be grasped. By taking into account changes in the rate of change of the time series data in this way, the NN can predict the data at time t that indicates the behavior of the target electric circuit A with higher accuracy than when past data is used directly to predict the data at time t that indicates the behavior of the target electric circuit A.

[0043] FIG. 10 is a diagram illustrating a modified example (3) of input data in the first embodiment. The input data of the modified example (3) is an input signal that combines time series data of modulated signals whose modulation methods (types) are QPSK, 64QAM, 256QAM, and OFDM, the first-order differential values ​​of the time series data of the modulated signals, the second-order differential values ​​of the time series data of the modulated signals, and the labels of the modulation methods. Note that FIG. 10 illustrates time series data whose modulation methods are QPSK, 64QAM, 256QAM, and OFDM, respectively, but these are merely examples and the order does not necessarily need to be the same. For example, by inputting the input data shown in FIG. 10 into a neural network, common behavior that is independent of the input power of the target electrical circuit A is modeled, and data indicating the modeling results is output from the neural network.

[0044] The input signal to the NN may be data obtained by smoothing either the first-order differential of the time-series data input to the target electric circuit A or the second-order differential of the time-series data input to the target electric circuit, or both. For example, if the input signal is time-series data containing noise, noise will still be present even if the time-series data is first- or second-order differentiated. Therefore, the influence of noise can be reduced by treating the first- or second-order differentiated data as input data after smoothing. Examples of smoothing methods include the moving average method.

[0045] Next, we will explain the hardware configuration that realizes the functions of the modeling device 1. The functions of the acquisition unit 11, NN unit 12, and polynomial calculation unit 13 provided in the modeling device 1 are realized by processing circuits. That is, the modeling device 1 includes a processing circuit for executing the processes of steps ST1 to ST3 shown in Fig. 2. The processing circuit may be dedicated hardware, or may be a CPU (Central Processing Unit) that executes a program stored in memory.

[0046] Fig. 11A is a block diagram showing a hardware configuration for realizing the functions of the modeling device 1. Fig. 11B is a block diagram showing a hardware configuration for executing software for realizing the functions of the modeling device 1. In Fig. 11A and Fig. 11B, an acquisition unit 11 acquires an input signal from an external device via an input interface 100. A polynomial calculation unit 13 outputs a signal representing the behavior of the target electric circuit to the outside via an output interface 101.

[0047] 11A , the processing circuit 102 may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof. The functions of the acquisition unit 11, the NN unit 12, and the polynomial calculation unit 13 included in the modeling device 1 may be realized by separate processing circuits, or these functions may be realized together by a single processing circuit.

[0048] 11B, the functions of the acquisition unit 11, the NN unit 12, and the polynomial calculation unit 13 included in the modeling device 1 are realized by software, firmware, or a combination of software and firmware. The software or firmware is written as a program and stored in the memory 104.

[0049] The processor 103 reads and executes programs stored in the memory 104 to realize the functions of the acquisition unit 11, the NN unit 12, and the polynomial calculation unit 13 included in the modeling device 1. For example, the modeling device 1 includes a memory 104 for storing a program that, when executed by the processor 103, results in the execution of steps ST1 to ST3 shown in FIG. 2 . These programs cause a computer to execute the procedures or methods of the processes performed by the acquisition unit 11, the NN unit 12, and the polynomial calculation unit 13. The memory 104 may be a computer-readable storage medium that stores programs for causing a computer to function as the acquisition unit 11, the NN unit 12, and the polynomial calculation unit 13.

[0050] The memory 104 may be, for example, a non-volatile or volatile semiconductor memory such as a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable Read Only Memory), or an EEPROM (Electrically-EPROM) (registered trademark), a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini disk, or a DVD.

[0051] Some of the functions of the acquisition unit 11, NN unit 12, and polynomial calculation unit 13 included in the modeling device 1 may be realized by dedicated hardware, and other functions may be realized by software or firmware. For example, the function of the acquisition unit 11 may be realized by a processing circuit 102 that is dedicated hardware, and the functions of the NN unit 12 and polynomial calculation unit 13 may be realized by the processor 103 reading and executing a program stored in memory 104. In this way, the processing circuit can realize the above functions by hardware, software, firmware, or a combination of these.

[0052] As described above, the modeling device 1 according to the first embodiment includes the NN unit 12 that inputs the input signal acquired by the acquisition unit 11 to the NN that outputs a signal representing common behavior independent of the input power in the target electric circuit when a signal is input, and the polynomial calculation unit 13 that calculates a signal representing behavior that changes depending on the input power in the target electric circuit by performing a polynomial calculation on the signal output from the NN. The NN is used to model the common behavior that does not depend on the input power in the target electric circuit, and further, the polynomial is used to model the behavior that changes depending on the input power in the target electric circuit. This allows the modeling device 1 to accurately model the behavior of the electric circuit.

[0053] In the modeling device 1 according to the first embodiment, the neural network is trained on a signal that is a combination of multiple time-series data representing input and output data of the target electric circuit at different input powers, as training data. This allows the neural network to learn common behavior that is independent of the input power of the target electric circuit.

[0054] In the modeling device 1 according to the first embodiment, the NN is trained using a data set including either or both of the first-order differential of the time-series data to be input to the target electric circuit A as training data. The NN can predict data with higher accuracy than when it does not use first-order or second-order differential data as input data.

[0055] In the modeling device 1 according to the first embodiment, the input signal is data obtained by smoothing either or both of the first-order differential of the time-series data input to the target electric circuit A and the second-order differential of the time-series data input to the target electric circuit A. This makes it possible to reduce the influence of noise contained in the time-series data.

[0056] In the modeling device 1 according to the first embodiment, the target electric circuit A is a rectifier circuit, which allows the modeling device 1 to accurately model the behavior of the rectifier circuit.

[0057] In the modeling device 1 according to the first embodiment, the input signal is composed of a combination of modulated signals of a plurality of modulation methods, and each modulated signal is assigned unique label data. This allows the neural network to learn while taking into account the type of modulated signal.

[0058] The modeling method according to the first embodiment includes step ST2 in which an input signal is input to an NN unit 12, which outputs a signal representing a common behavior independent of the input power in the target electric circuit when the signal is input, and step ST3 in which a polynomial calculation unit 13 calculates a signal representing a behavior that changes depending on the input power in the target electric circuit by performing a polynomial calculation on the signal output from the NN. By having the modeling device 1 execute this method, the behavior of the electric circuit can be modeled with high accuracy.

[0059] The computer that executes the program according to the first embodiment functions as an NN unit 12 that inputs the acquired input signal to an NN that outputs a signal that represents a common behavior independent of the input power in the target electric circuit when a signal is input, and as a polynomial calculation unit 13 that calculates a signal that represents a behavior that changes depending on the input power in the target electric circuit by performing a polynomial calculation on the signal output from the NN. This enables the computer to accurately model the behavior of the electric circuit.

[0060] Second Embodiment In a modeling device according to a second embodiment, an input signal is input to a polynomial calculation unit, and an NN receives the output from the polynomial calculation unit and produces a final output.

[0061] 12 is a block diagram showing an example configuration of a modeling device 1A according to a second embodiment. In FIG. 12, the modeling device 1A is a device for modeling the behavior of a rectifier circuit as a target electric circuit A, and includes an acquisition unit 11A, an NN unit 12A, a polynomial calculation unit 13A, and a memory 14. The modeling device 1A is implemented by a computer. The memory included in the computer stores programs constituting information processing applications for implementing the functions of the acquisition unit 11A, the NN unit 12A, and the polynomial calculation unit 13A. The processor included in the computer reads the information processing application from the memory and executes it, thereby implementing the functions of the acquisition unit 11A, the NN unit 12A, and the polynomial calculation unit 13A.

[0062] (Acquisition Unit) The acquisition unit 11A acquires an input signal and outputs the acquired input signal to the polynomial calculation unit 13A. The input signal is time-series data of a signal supplied to an electric circuit. The time-series data may be a voltage waveform or a power waveform applied to the electric circuit. The signal value at each time is a voltage value or a current value. In FIG. 12, the input signal is expressed as x ｔ , x ｔ－１ , ..., x ｔ－ｎ＋１ , x ｔ－ｎ n is time-series data, where n is an integer. For example, if the modeling device 1A includes a communication unit (not shown in FIG. 12 ) and is capable of communicating with an external device (not shown in FIG. 12 ) via this communication unit, the acquisition unit 11 controls the communication unit to acquire an input signal from the external device. The communication unit communicates with the external device via a communication network. Furthermore, for example, if the modeling device 1A includes a storage unit (not shown in FIG. 12 ) and the storage unit stores input signals for each target electric circuit, the acquisition unit 11A may read and acquire the input signals from the storage unit. In this case, the modeling device 1A does not need to include the communication unit.

[0063] (Polynomial Calculation Unit) The polynomial calculation unit 13A calculates a signal representing a behavior that changes depending on the input power in the target electric circuit by performing a polynomial calculation on the input signal acquired by the acquisition unit 11A. For example, polynomials (1), (2), and (3) are set in the polynomial calculation unit 13A. The polynomial calculation unit 13A selects one of these polynomials to use for calculating the output signal depending on the input power.

[0064] (Memory) The memory 14 is a storage unit that stores a certain number of pieces of output data from the polynomial calculation unit 13A. Because a polynomial has one output signal for each input signal, the output of the polynomial cannot be input directly to the NN. Therefore, the memory 14 stores a certain number of pieces of output data from the polynomial, and the data accumulated in the memory 14 is output to the NN unit 12A. For example, if the NN requires 10 pieces of data as input data, at least 10 pieces of data are stored in the memory 14. After passing the data read from the memory 14 to the NN, the memory 14 discards the oldest piece of data on the time axis and adds and stores the data output from the polynomial calculation unit 13A.

[0065] (NN Unit) The NN unit 12A inputs a signal read from the memory 14 to a NN that, when a signal is input, outputs a signal representing common behavior independent of the input power in the target electric circuit A. The NN takes in the input signal as a feature and is trained to output a signal representing common behavior independent of fluctuations in the input power in the target electric circuit A. The common behavior independent of fluctuations in the input power is modeled by a feature of the NN known as the universality theorem, which allows modeling of any continuous function.

[0066] The NN is assumed to be included in the modeling device 1A, but may also be included in an external device that can be connected for communication with the modeling device 1A. The NN is composed of data including nodes, edges, and their weight information. This data may be included in the modeling device 1 or in an external device. For example, if the modeling device 1A is equipped with a communication unit not shown in FIG. 1 and is capable of connecting for communication with an external device not shown in FIG. 1 via this communication unit, the NN unit 12A controls the communication unit to input an input signal to the NN included in the external device. An output signal from the NN is received by the communication unit and output as a modeling result.

[0067] As described above, the modeling device 1A according to the second embodiment includes a polynomial calculation unit 13A that calculates a signal representing behavior that changes depending on the input power in a target electric circuit by performing a polynomial calculation on an input signal, and an NN unit 12A that inputs the signal calculated by the polynomial calculation unit 13A to an NN that, upon receiving the signal, outputs a signal representing common behavior that is independent of the input power in the target electric circuit. The NN is used to model the common behavior that is independent of the input power in the target electric circuit, and the polynomial is used to model the behavior that changes depending on the input power in the target electric circuit. This allows the modeling device 1A to accurately model the behavior of the electric circuit.

[0068] Embodiment 3 A modeling device according to embodiment 3 includes a plurality of NN units and polynomial calculation units provided between the NNs. The NN units provided on the output side of the polynomial calculation units input signals calculated by the polynomial calculation units to the NNs.

[0069] FIG. 13 is a block diagram showing an example configuration of a modeling device 1B according to a third embodiment. In FIG. 13, the modeling device 1A is a device for modeling the behavior of a rectifier circuit as a target electric circuit A, and includes an acquisition unit 11B, NN units 12B-1 and 12B-2, and a polynomial calculation unit 13B. The modeling device 1B is realized by a computer. A memory included in the computer stores a program constituting an information processing application for realizing the functions of the acquisition unit 11B, NN units 12B-1 and 12B-2, and polynomial calculation unit 13B. A processor included in the computer reads the information processing application from the memory and executes it, thereby realizing the functions of the acquisition unit 11B, NN units 12B-1 and 12B-2, and polynomial calculation unit 13B.

[0070] In the modeling device 1B, as shown in FIG. 13, NN units and polynomial calculation units are alternately connected in multiple stages. The first-stage NN unit 12B-1 performs general modeling, the second-stage polynomial calculation unit 13B performs slightly more accurate modeling, and the third-stage NN unit 12B-2 performs even more accurate modeling. By increasing the number of stages in this way, the number of coefficients that each NN must hold can be reduced. For example, if the NN handled by the first-stage NN unit 12B-1 uses past data such as an LSTM or RNN, it is possible to model behavior that takes into account the memory effect of the diode. Therefore, the NN handled by the subsequent-stage NN unit 12B-2 may be a fully-coupled NN, which is a general NN. Furthermore, there is no need to provide a memory in the output stage of the polynomial calculation unit 13B.

[0071] 13, the NN unit 12B-1 is provided in the first stage, but a polynomial calculation unit may also be provided in the first stage. Even with this configuration, the behavior of the electric circuit can be modeled with high accuracy.

[0072] As described above, the modeling device 1B according to the third embodiment includes NN units 12B-1 and 12B-2, and a polynomial calculation unit 13B provided between NN 12B-1 and NN 12B-2. NN unit 12B-2, which is provided on the output side of polynomial calculation unit 13B, inputs a signal calculated by polynomial calculation unit 13B to the NN. The NN is used to model common behavior that is independent of the input power in the target electric circuit, and the polynomial is used to model behavior that changes depending on the input power in the target electric circuit. This allows the modeling device 1B to accurately model the behavior of the electric circuit.

[0073] It is possible to combine the embodiments, modify any of the components of the embodiments, or omit any of the components of the embodiments.

[0074] The modeling device according to the present disclosure can be used to develop electric circuits such as rectifier circuits.

[0075] 1, 1A, 1B Modeling device, 11, 11A, 11B Acquisition unit, 12, 12A, 12B-1, 12B-2 NN unit, 13, 13A, 13B Polynomial calculation unit, 100 Input interface, 101 Output interface, 102 Processing circuit, 103 Processor, 104 Memory.

Claims

1. A modeling device comprising: a neural network unit that inputs an input signal to a neural network that, when a signal is input, outputs a signal that represents a common behavior independent of the input power in the target electric circuit; and a polynomial calculation unit that calculates a signal that represents behavior that changes depending on the input power in the target electric circuit by calculating the input signal using a polynomial.

2. The modeling device according to claim 1, wherein said polynomial calculation section calculates a signal by performing a calculation using a polynomial on the signal output from said neural network.

3. The modeling device according to claim 1, wherein said neural network section inputs the signal calculated by said polynomial operation section to said neural network.

4. A modeling device as described in claim 1, comprising a plurality of said neural network units and said polynomial calculation units provided between said neural network units, wherein said neural network unit provided on the output side of said polynomial calculation units inputs said calculated signal to said neural network.

5. A modeling device according to any one of claims 1 to 4, characterized in that the neural network is trained on a signal composed of a combination of multiple time-series data representing input and output data of the target electrical circuit at different input powers, as training data.

6. A modeling device according to any one of claims 1 to 4, characterized in that the neural network is trained using a data set including either the first derivative of the time series data to be input to the target electrical circuit or the second derivative of the time series data to be input to the target electrical circuit, or both, as training data.

7. The modeling device according to claim 6, wherein the input signal is data obtained by smoothing either the first derivative of the time series data input to the target electrical circuit or the second derivative of the time series data input to the target electrical circuit, or both.

8. A modeling device according to any one of claims 1 to 7, wherein the target electric circuit is a rectifier circuit.

9. A modeling device according to any one of claims 1 to 8, characterized in that the input signal is composed of a combination of modulated signals of a plurality of modulation methods, and each modulated signal is assigned unique label data.

10. A modeling method using a modeling device, comprising: a step in which a neural network unit inputs an input signal to a neural network that outputs a signal representing common behavior independent of the input power in a target electric circuit when the signal is input; and a step in which a polynomial calculation unit calculates a signal representing behavior that changes depending on the input power in the target electric circuit by calculating the input signal using a polynomial.

11. The modeling method according to claim 10, wherein the polynomial calculation unit calculates a signal by performing a polynomial calculation on the signal output from the neural network.

12. The modeling method according to claim 10, wherein said neural network section inputs the signal calculated by said polynomial operation section to said neural network.

13. A program for causing a computer to function as: a neural network unit that inputs an input signal to a neural network that, when a signal is input, outputs a signal that represents a common behavior independent of the input power in a target electrical circuit; and a polynomial calculation unit that calculates a signal that represents behavior that changes depending on the input power in the target electrical circuit by calculating the input signal using a polynomial.

14. The program according to claim 13, wherein the polynomial calculation unit calculates a signal by performing a polynomial calculation on the signal output from the neural network.

15. The program according to claim 13, wherein said neural network unit inputs the signal calculated by said polynomial operation unit to said neural network.

Images