JP7428893B2 - Learning device, abnormal eddy current loss estimation device, learning method, abnormal eddy current loss estimation method, and program - Google Patents

Description

The present invention relates to a learning device, an abnormal eddy current loss estimation device, a learning method, an abnormal eddy current loss estimation method, and a program, and is particularly suitable for use in estimating abnormal eddy current loss in soft magnetic materials. .

It is important to estimate the loss of electrical equipment in order to improve its efficiency. For example, losses occurring in motors mainly include copper loss, iron loss, and mechanical loss. When the motor is operated at high speed, the proportion of iron loss in the total loss of the motor increases. Therefore, in order to estimate the efficiency of the motor when the motor is operating at high speed, it is necessary to estimate the core loss of the motor with high accuracy.

Iron loss is classified into hysteresis loss, classical eddy current loss, and abnormal eddy current loss. Under high-frequency excitation conditions, the proportions of classical eddy current loss and abnormal eddy current loss in iron loss become dominant. Classical eddy current loss is theoretically determined by Faraday's law of electromagnetic induction. On the other hand, it has been hypothesized that abnormal eddy current loss is a loss due to induced current that occurs due to the movement of domain walls, but a physical model for quantitatively evaluating abnormal eddy current loss has not been established. do not have.

Against this background, Patent Document 1 discloses that when the soft magnetic material is excited, the maximum value of the magnitude of the magnetic flux density generated in the soft magnetic material in one cycle is obtained from the results of numerical analysis, and the It is described that an abnormal eddy current loss coefficient corresponding to the maximum value of magnetic flux density is derived. Further, Non-Patent Document 1 describes that the abnormal eddy current loss coefficient is expressed using the time differential value of the magnetic flux density of the electromagnetic steel sheet. The abnormal eddy current loss coefficient is a coefficient by which the classical eddy current loss is multiplied when expressing the total eddy current loss. By using the abnormal eddy current loss coefficient, the abnormal eddy current loss of the soft magnetic material can be derived.

Patent No. 6206608

Ryuhei Mitsuoka and two others, "Study on AC hysteresis model of electrical steel sheets using play model and finite element eddy current analysis", Institute of Electrical Engineers of Japan Research Group Materials Rotating Machine Research Group, SA-12-17/RM-12- 17, January 26, 2012, p. 95-98 Takayoshi Nakata and Norio Takahashi, "Finite Element Method for Electrical Engineering", 2nd edition, Morikita Publishing Co., Ltd., April 1986.

However, in the technique described in Patent Document 1, the abnormal eddy current loss coefficient is determined from the result of measuring the magnetic flux density excited by a sine wave. Therefore, even if the time waveform of the magnetic flux density of the soft magnetic material to be estimated is different from the time waveform of the magnetic flux density when calculating the abnormal eddy current loss coefficient, if the maximum value of their magnetic flux densities is the same, then , the same value is derived as the abnormal eddy current loss coefficient.

Further, in the technique described in Patent Document 2, the abnormal eddy current loss coefficient is derived using the effective value of the time differential value of the magnetic flux density over one period, and the magnetic flux density is not treated as time-series data. Therefore, for example, if the shape of the time waveform of the magnetic flux density is the same, the same value will be derived as the abnormal eddy current loss coefficient whether or not a bias is superimposed on the time waveform.
As described above, in the technologies described in Patent Documents 1 and 2, for example, when the time waveform of the magnetic flux density contains harmonic components or when bias is superimposed, the time waveform of the magnetic flux density corresponds to the time waveform of the magnetic flux density. It is not easy to accurately derive abnormal eddy current loss.
Further, in the techniques described in Patent Documents 1 and 2, a representative value of the abnormal eddy current loss in one period is derived, but it is not possible to derive the abnormal eddy current loss at each time within one period.

The present invention has been made in view of the above problems, and an object of the present invention is to enable accurate derivation of abnormal eddy current loss in soft magnetic materials.

The learning device of the present invention is a learning device for learning parameters of a neural network for estimating abnormal eddy current loss of a soft magnetic material, and the learning device is a learning device for learning parameters of a neural network for estimating abnormal eddy current loss of a soft magnetic material. an acquisition means for acquiring, as measured values for the material, a measured value at each time within one cycle of a physical quantity obtained from the magnetization property of the soft magnetic material, and a measured value of iron loss of the soft magnetic material; a first electromagnetic field analysis means for deriving representative values in one period of hysteresis loss and classical eddy current loss of the soft magnetic material when the material is excited under the predetermined excitation conditions, based on Maxwell's equation; The abnormal eddy current in the soft magnetic material is determined based on the measured value of iron loss acquired by the acquisition means and the representative values of hysteresis loss and classical eddy current loss in one cycle derived by the first electromagnetic field analysis means. abnormal eddy current loss deriving means for deriving a representative value of loss in one cycle; a representative value of abnormal eddy current loss in one cycle of the soft magnetic material derived by the abnormal eddy current loss deriving means; and a magnitude of the physical quantity. a teacher data creating means for creating teacher data based on a value at each time within one cycle of the magnitude of the physical quantity and a value at each time within one cycle of the time derivative of the magnitude of the physical quantity; a learning means for learning parameters of a neural network using the teacher data created by the creation means, and an input layer of the neural network includes the magnitude of the physical quantity at a certain time and the At least a time differential value of the magnitude of the physical quantity is input, at least an abnormal eddy current loss at the relevant time is output from the output layer of the neural network, and the learning means Using an objective function to at least obtain an evaluation value of the difference between a representative value of abnormal eddy current loss in one cycle and a representative value of abnormal eddy current loss at each time in one cycle output from the output layer of the neural network. The method is characterized in that the parameters of the neural network are learned.
The abnormal eddy current loss estimating device of the present invention is an abnormal eddy current loss estimating device that estimates an abnormal eddy current loss of a soft magnetic material using the neural network whose parameters have been learned by the learning device, A derivation means for deriving the magnitude of the physical quantity of the soft magnetic material whose current loss is to be estimated and a time differential value of the magnitude of the physical quantity when the soft magnetic material whose current loss is to be estimated is excited under predetermined excitation conditions; Based on the value output from the output layer by giving the derived magnitude of the physical quantity at a certain time and the time differential value of the magnitude of the physical quantity at the time to the input layer of the neural network, The method is characterized by comprising an estimating means for estimating the abnormal eddy current loss of the soft magnetic material at a given time.

The learning method of the present invention is a learning method for learning parameters of a neural network for estimating abnormal eddy current loss of a soft magnetic material, which an acquisition step of acquiring, as measured values for the material, measured values at each time within one cycle of a physical quantity obtained from the magnetization characteristics of the soft magnetic material and a measured value of iron loss of the soft magnetic material; a first electromagnetic field analysis step of deriving representative values in one period of hysteresis loss and classical eddy current loss of the soft magnetic material when the material is excited under the predetermined excitation conditions, based on Maxwell's equation; The abnormal eddy current in the soft magnetic material is calculated based on the measured value of iron loss obtained in the acquisition step and the representative values in one cycle of hysteresis loss and classical eddy current loss derived in the first electromagnetic field analysis step. An abnormal eddy current loss derivation step for deriving a representative value in one cycle of loss, a typical value in one cycle of abnormal eddy current loss in the soft magnetic material derived by the abnormal eddy current loss derivation step, and the magnitude of the physical quantity. a teacher data creation step of creating teacher data based on a value at each time within one cycle of the magnitude of the physical quantity and a value at each time within one cycle of the time derivative of the magnitude of the physical quantity; a learning step of learning parameters of a neural network using the teacher data created in the creation step, and an input layer of the neural network includes the magnitude of the physical quantity at a certain time and the At least a time differential value of the magnitude of the physical quantity is input, at least the abnormal eddy current loss at the relevant time is output from the output layer of the neural network, and the learning step Using an objective function to at least obtain an evaluation value of the difference between a representative value of abnormal eddy current loss in one cycle and a representative value of abnormal eddy current loss at each time in one cycle output from the output layer of the neural network. The method is characterized in that the parameters of the neural network are learned.
The abnormal eddy current loss estimation method of the present invention is an abnormal eddy current loss estimation method for estimating abnormal eddy current loss of a soft magnetic material using the neural network whose parameters have been learned by the learning method, A derivation step of deriving the magnitude of the physical quantity of the soft magnetic material and the time differential value of the magnitude of the physical quantity when the soft magnetic material whose current loss is to be estimated is excited under predetermined excitation conditions; and the derivation step Based on the value output from the output layer by giving the derived magnitude of the physical quantity at a certain time and the time differential value of the magnitude of the physical quantity at the time to the input layer of the neural network, The method is characterized by comprising an estimating step of estimating an abnormal eddy current loss of the soft magnetic material at a given time.

A first example of the program of the present invention is for causing a computer to function as each means of the learning device.
A second example of the program of the present invention is for causing a computer to function as each means of the abnormal eddy current loss estimating device.

According to the present invention, it is possible to accurately derive abnormal eddy current loss in a soft magnetic material.

It is a figure showing an example of the functional composition of an abnormal eddy current loss estimating device. FIG. 3 is a diagram showing an example of a time waveform of an instantaneous value of magnetic flux density, a time waveform of the magnitude of magnetic flux density, and a time waveform of a time differential value of the magnitude of magnetic flux density. FIG. 2 is a diagram conceptually showing an example of a neural network. 3 is a flowchart illustrating an example of a method for learning parameters of a neural network. It is a flow chart explaining an example of an abnormal eddy current loss estimation method. It is a figure which shows the result of comparing the estimated value of abnormal eddy current loss and the measured value of abnormal eddy current loss in an invention example and a comparative example.

<Idea>
Before describing embodiments of the present invention, an idea obtained by the present inventor will be described.
Abnormal eddy current loss is eddy current loss that occurs due to movement of domain walls. In the magnetization process of a ferromagnetic material, first, movement of the domain wall occurs, and the 180° domain wall expands before the 90° domain wall. After such movement of the domain wall, a process occurs in which the direction of magnetization within the domain wall changes. Therefore, the degree of occurrence of abnormal eddy current loss is considered to vary depending on the magnetization state, and the magnitude is considered to be determined by the magnitude of the magnetic flux density and the temporal change in the magnetic flux density. Furthermore, since the magnetic field strength corresponds to the magnetic flux density, the magnitude of the abnormal eddy current loss is considered to be determined by the magnitude of the magnetic field strength and the temporal change in the magnetic field strength. In addition, since magnetic permeability and differential permeability also change depending on the magnetization state, the magnitude of abnormal eddy current loss is determined by the magnitude of magnetic permeability and time change in permeability, as well as the magnitude of differential permeability and time change in differential permeability. It can be thought that it is also determined by Note that magnetic permeability is the ratio of magnetic flux density and magnetic field strength that correspond to each other in magnetization characteristics (the value obtained by dividing the magnetic flux density by the magnetic field strength). Differential magnetic permeability is the slope of the tangent at a point on the magnetization characteristic curve that is determined according to the magnetic flux density.

Magnetic flux density, magnetic field strength, magnetic permeability, and differential magnetic permeability are all physical quantities obtained from magnetization characteristics, and are one of the parameters expressing magnetization characteristics. Magnetization characteristics represent the relationship between magnetic flux density and magnetic field strength, and are also referred to as magnetic hysteresis characteristics. The magnetization characteristic curve is a curve representing magnetization characteristics.

As explained in the background technology section, it is hypothesized that abnormal eddy current loss is a loss due to induced current that occurs due to movement of domain walls, but there is no way to quantitatively evaluate abnormal eddy current loss. A physical model has not been established. Therefore, as described above, the present inventors focused on the fact that the magnitude and time differential value of the physical quantity obtained from the magnetization characteristics correspond to the magnitude of the abnormal eddy current loss. The present inventor also noted that abnormal eddy current loss occurs at each time. As a result of intensive study, the inventors determined that the magnitude of the physical quantity obtained from the magnetization characteristics and the time differential value of the magnitude of the physical quantity at a certain time (instantaneous value) were used as explanatory variables, and the abnormal vortex By constructing a neural network that uses the current loss value at the relevant time as the objective variable, it is possible to accurately estimate abnormal eddy current loss without using a physical model to quantitatively evaluate abnormal eddy current loss. I came up with the idea.
The following embodiments have been made based on the above idea.

Hereinafter, one embodiment of the present invention will be described with reference to the drawings.
<Configuration of abnormal eddy current loss estimation device>
FIG. 1 is a diagram showing an example of a functional configuration of an abnormal eddy current loss estimating device 100. The hardware configuration of the abnormal eddy current loss estimating device 100 is realized by using, for example, an information processing device having a CPU, ROM, RAM, HDD, and various interfaces, or dedicated hardware. An example of the functions of the abnormal eddy current loss estimating device 100 of this embodiment will be described below. In addition, in this embodiment, the case where a soft magnetic material is an electromagnetic steel plate is illustrated. The electrical steel sheet is, for example, a grain-oriented electrical steel sheet or a non-oriented electrical steel sheet.

<<Acquisition unit 101>>
The acquisition unit 101 acquires a measured value at each time within one cycle of a physical quantity obtained from the magnetization characteristics of the electrical steel sheet, and a measured value of iron loss of the electrical steel sheet. These measured values were measured when the electromagnetic steel sheet was excited under the same excitation conditions. The excitation condition is information that determines the time waveform of the excitation current. When the time waveform of the excitation current is a sine wave, such information is, for example, an effective value and a frequency. Moreover, one period is one period of the time waveform of the excitation current. Further, each time within one cycle is, for example, a time when the first time of the cycle is set as a reference value (for example, 0), and does not necessarily match the time in the real world. This also applies to the following explanations.

As described above, the physical quantities obtained from the magnetization characteristics are magnetic flux density, magnetic field strength, magnetic permeability, and differential magnetic permeability. At least one of magnetic flux density, magnetic field strength, magnetic permeability, and differential magnetic permeability can be used as a physical quantity obtained from magnetization characteristics, but in this embodiment, when magnetic flux density is a physical quantity obtained from magnetization characteristics exemplify.

The measured value of magnetic flux density and the measured value of iron loss of electrical steel sheets are, for example, JIS C 2556:2015 "Method for measuring the magnetic properties of electrical steel strip using a veneer tester" and JIS 2550-1:2011 "Method of measuring magnetic properties of electrical steel strip". Since it can be obtained by the known method described in "Strand Test Method - Part 1: Method for Measuring Magnetic Characteristics of Electromagnetic Steel Strip Using Epstein Tester," detailed explanation thereof will be omitted here.

Note that the measured value of iron loss is measured only once in one period using the measured value at each time within one period of magnetic flux density. That is, the measured value of iron loss is a representative value (average value) in one cycle. Further, each measurement value (measurement value at each time within one cycle of magnetic flux density and measurement value of iron loss) acquired by the acquisition unit 101 is a measurement value for the entire electrical steel sheet. Therefore, when a certain electromagnetic steel sheet is excited under certain excitation conditions, one set of the measured value at each time within one period of the magnetic flux density and the measured value of iron loss is obtained.

In the following explanation, pairs of measured values at each time within one cycle of magnetic flux density and measured values of iron loss when the same type of electrical steel sheet is excited under the same excitation conditions are explained as necessary. This is called measurement data. Note that the measurement data also includes information indicating the excitation conditions used to obtain the measurement data and information indicating the type of electromagnetic steel sheet to be measured in the measurement data. The type of electrical steel sheet is information that specifies the classification of the electrical steel sheet according to the attributes of the electrical steel sheet. The attributes of the electrical steel sheet include, for example, the components contained in the electrical steel sheet and the amounts of the components.

By exciting an electromagnetic steel sheet of the same type as the electromagnetic steel sheet whose abnormal eddy current loss is to be estimated under a plurality of mutually different excitation conditions, measurement data corresponding to each excitation condition is obtained. The acquisition unit 101 acquires a plurality of measurement data obtained in this way. Further, the acquisition unit 101 may acquire measurement data regarding multiple types of electrical steel sheets.
Forms of acquiring measurement data include, for example, communication with an external device, reading measurement data stored in a portable storage medium, and inputting measurement data values to the user interface of the abnormal eddy current loss estimating device 100. At least one of these can be adopted.

<<First electromagnetic field analysis unit 102>>
The first electromagnetic field analysis unit 102 derives representative values of hysteresis loss and classical eddy current loss in one cycle of the electromagnetic steel sheet when the electromagnetic steel sheet is excited under predetermined excitation conditions based on Maxwell's equation. A specific example of the processing of the first electromagnetic field analysis unit 102 will be described below.

The first electromagnetic field analysis section 102 analyzes each minute region (mesh) using the finite element method based on Maxwell's equations, in accordance with the electromagnetic field analysis conditions including the excitation conditions included in the measurement data acquired by the acquisition section 101. The magnetic flux density vector B and the eddy current vector J _e are calculated at each time within one period. The magnetic flux density vector B and the eddy current vector J _e are calculated at calculation timings that arrive at every predetermined time step. If the time steps are short, values can be derived at fine time steps, but the processing load increases. The time step is predetermined in consideration of such trade-offs.

Note that as long as the magnetic flux density vector B and the eddy current vector J _e in each minute region can be calculated, the electromagnetic field analysis may be performed using a method other than the finite element method (such as the finite difference method).
The basic equations for calculating the magnetic flux density vector B and the eddy current vector J _e are generally given by the following equations (1) to (4).

In equations (1) to (4), μ is magnetic permeability [H/m], A is vector potential [T・m], and σ is electrical conductivity [S/m], J ₀ is the excitation current density [A/m ² ], and φ is the scalar potential [V].
After solving equations (1) and (2) simultaneously to obtain vector potential A and scalar potential φ, from equations (3) and (4), magnetic flux density vector B and eddy current vector J _e can be calculated. Calculated.

As a solution method using vector potential A and scalar potential φ as unknown variables, there is, for example, the Newton-Raphson method.
The method of performing electromagnetic field analysis using the finite element method is a general method, as described in detail in Non-Patent Document 2, and therefore a detailed explanation thereof will be omitted here.

The first electromagnetic field analysis unit 102 calculates the magnetization characteristic (magnetic hysteresis characteristic) in a certain minute region from the time waveform in one period of the magnetic flux density vector B in a certain minute region and the magnetic field strength vector H corresponding to the magnetic flux density vector B. Derive. Then, the first electromagnetic field analysis unit 102 derives the area of the derived magnetization characteristic as the hysteresis loss W _h in the minute region using the following equation (5).

The first electromagnetic field analysis unit 102 executes the above calculations for all micro regions. The first electromagnetic field analysis unit 102 derives the sum of the hysteresis loss W _h in all the micro regions as a representative value of the hysteresis loss of the electrical steel sheet in one cycle. This hysteresis loss of the electrical steel sheet is a hysteresis loss for the entire electrical steel sheet.

In addition, the first electromagnetic field analysis unit 102 calculates the micro-region based on the eddy current vector J _e in a certain micro-region, the size of the micro-region, and the electrical conductivity σ of the electromagnetic steel sheet using equation (6) below. Derive the classical eddy current loss W _e in the region.

In equation (6), T is the period of the eddy current vector J _e .
The first electromagnetic field analysis unit 102 performs the above calculations for all micro regions. The first electromagnetic field analysis unit 102 derives the sum total of the classical eddy current loss W _e in all minute regions as a representative value of the classical eddy current loss of the electrical steel sheet in one cycle. The classical eddy current loss of this electrical steel sheet is the classical eddy current loss for the entire electrical steel sheet.

The first electromagnetic field analysis unit 102 calculates the representative values of the hysteresis loss and classical eddy current loss of the electrical steel sheet in one cycle for each of the excitation conditions included in the measurement data in the manner described above. Execute by type.
Note that the method by which the first electromagnetic field analysis unit 102 derives the representative values of the hysteresis loss and classical eddy current loss of the electrical steel sheet in one cycle is not limited to the method using equations (5) and (6), Other known methods may also be employed. For example, the Steinmetz equation using the Steinmetz coefficient may be used to derive the representative values of the hysteresis loss and classical eddy current loss in one cycle of the electrical steel sheet.
In the following explanation, representative values of hysteresis loss and classical eddy current loss in one cycle of an electrical steel sheet, which are derived by the first electromagnetic field analysis unit 102 as described above, will be described as needed. It is abbreviated as the representative value of current loss in one cycle.

<<Abnormal eddy current loss deriving unit 103>>
The abnormal eddy current loss deriving unit 103 calculates the iron loss measurement value included in the measurement data acquired by the acquisition unit 101 and the hysteresis loss and classical eddy current loss in one cycle derived by the first electromagnetic field analysis unit 102. Based on the representative values, a representative value in one cycle of abnormal eddy current loss in the electrical steel sheet is derived. A specific example of the processing of the abnormal eddy current loss deriving unit 103 will be described below. As described above, the measured value of iron loss included in the measurement data acquired by the acquisition unit 101 is an average value in one cycle.

The abnormal eddy current loss deriving unit 103 selects one of the measurement data acquired by the acquisition unit 101. The abnormal eddy current loss deriving unit 103 extracts the measured value of iron loss included in the selected measurement data. The abnormal eddy current loss deriving unit 103 calculates the hysteresis loss and classical eddy current loss corresponding to the selected measurement data from among the representative values in one cycle of the hysteresis loss and classical eddy current loss derived by the first electromagnetic field analysis unit 102. Extract the representative value in one period. The typical values of hysteresis loss and classical eddy current loss in one cycle corresponding to the measurement data are the same as the excitation conditions included in the measurement data for the same type of electrical steel sheet as the type of electrical steel sheet included in the measurement data. These are representative values for one period of hysteresis loss and classical eddy current loss when excited under excitation conditions.

The abnormal eddy current loss deriving unit 103 adds the representative value of the extracted hysteresis loss in one cycle and the representative value of the extracted classical eddy current loss in one cycle, and calculates the extracted hysteresis loss and classical eddy current loss. Derive the sum of representative values in the period.
The abnormal eddy current loss deriving unit 103 calculates the value obtained by subtracting the sum of the representative values of the extracted hysteresis loss and classical eddy current loss in one cycle from the extracted measured value of iron loss, as the abnormal eddy current loss of the electrical steel sheet. Derived as a representative value in the period.

The abnormal eddy current loss deriving unit 103 derives the representative value of the abnormal eddy current loss in one cycle of the electrical steel sheet as described above for each type of electrical steel sheet for each excitation condition included in the measurement data. Execute. The typical value of the abnormal eddy current loss of the electrical steel sheet in one cycle is the abnormal eddy current loss for the entire electrical steel sheet. In the following description, the representative value of the abnormal eddy current loss of the electrical steel sheet in one cycle, which is derived by the abnormal eddy current loss deriving unit 103 as described above, will be referred to as learning objective variable data as necessary.

<<Teacher data creation unit 104>>
The teacher data creation unit 104 uses the learning objective variable data (representative value in one cycle of the abnormal eddy current loss) derived by the abnormal eddy current loss derivation unit 103 and the magnitude of the magnetic flux density of the electrical steel sheet within one cycle. Teacher data is created based on the value at each time and the value at each time within one period of the time differential value of the magnitude of the magnetic flux density of the electromagnetic steel sheet. Here, the magnitude of the magnetic flux density is assumed to be an absolute value. Further, the value at each time within one cycle of the magnitude of the magnetic flux density of the electrical steel sheet and the value at each time within one cycle of the time differential value of the magnitude of the magnetic flux density of the electrical steel sheet are acquired by the acquisition unit 101. It is obtained based on the measured value at each time within one cycle of the magnetic flux density included in the measured data. A specific example of the processing of the teacher data creation unit 104 will be described below.

The teacher data creation unit 104 selects one of the measurement data acquired by the acquisition unit 101. The teacher data creation unit 104 derives the value of the magnetic flux density at each time within one cycle based on the measured value at each time within one cycle of the magnetic flux density included in the selected measurement data. For example, the measured value at each time within one cycle of magnetic flux density may be used as is as the value at each time within one cycle of the magnitude of magnetic flux density. In addition, for example, smoothing processing (smoothing processing) is performed on the measured values at each time within one cycle of magnetic flux density to remove noise, and The value may be a value at each time within one cycle of the magnitude of the magnetic flux density.

Further, the teacher data creation unit 104 generates a value at each time within one cycle of the time differential value of the magnitude of the magnetic flux density, based on the measured value at each time within one cycle of the magnetic flux density included in the selected measurement data. Derive. As described above, for example, a time differential value may be derived for the value at each time within one cycle of the magnetic flux density that has been subjected to the smoothing process.

FIG. 2 is a diagram showing an example of a time waveform of an instantaneous value of magnetic flux density, a time waveform of the magnitude of magnetic flux density, and a time waveform of a time differential value of the magnitude of magnetic flux density. In FIG. 2, the period T from time t1 to time t2 is one cycle.
A time waveform 201 is an example of a time waveform of an instantaneous value of magnetic flux density obtained from a measured value of magnetic flux density. The time waveform 202 is an example of the time waveform of the magnitude of the magnetic flux density obtained from the time waveform 201. The time waveform 202 is obtained by taking the absolute value of each time value of the time waveform 201. The time waveform 203 is an example of a time waveform of the time differential value of the magnitude of the magnetic flux density obtained from the time waveform 201. The time waveform 203 is obtained by time-differentiating the time waveform 202. Note that time differentiation refers to first-order differentiation with respect to time. Time differentiation is realized by numerical differentiation. The numerical differentiation method may be a forward difference method, a backward difference method, or a central difference method.

At times t3, t4, and t5 when the magnetic flux density in the instantaneous value of the time waveform 201 of the magnetic flux density is 0, the time waveform 202 of the magnitude of the magnetic flux density becomes non-differentiable. Therefore, as a time differential value of the magnitude of the magnetic flux density at such times t3, t3, and t5, for example, the magnitude of the magnetic flux density at two times temporally preceding and following the times t3, t4, and t5. The arithmetic mean value of the time differential values of can be adopted. Further, at times t3, t4, and t5, there may be no time differential value of the magnitude of the magnetic flux density.
In the following explanation, the value at each time within one cycle of the magnitude of magnetic flux density and the time differential value of the magnitude of magnetic flux density within one cycle, which are derived by the teacher data creation unit 104 as described above, will be described. The values at each time are referred to as learning explanatory variable data as necessary.

Further, the teacher data creation unit 104 creates data in which explanatory variable data for learning and objective variable data for learning are correlated with each other, which corresponds to the selected measurement data, as teacher data. The teacher data also includes information indicating the type of electrical steel sheet included in the selected measurement data.

The learning explanatory variable data corresponding to the selected measurement data is the magnitude of the magnetic flux density derived by the teacher data creation unit 104 based on the measured values at each time within one cycle of the magnetic flux density included in the measurement data. These are the values at each time within one cycle of the magnetic flux density and the values at each time within one cycle of the time differential value of the magnitude of the magnetic flux density.
Further, the learning objective variable data corresponding to the selected measurement data refers to the abnormal eddy current loss of the electrical steel sheet derived by the abnormal eddy current loss deriving unit 103 based on the measured value of iron loss included in the measurement data. This is a representative value in one cycle.

The teacher data creation unit 104 selects the measurement data one by one and creates the teacher data of the electrical steel sheet as described above. As a result, training data for each type of electrical steel sheet is created.

<<Learning Department 105>>
The learning unit 105 uses the teacher data created by the teacher data creation unit 104 to learn the parameters of the neural network. An example of processing by the learning unit 105 will be described below.

The learning unit 105 performs supervised learning to learn the parameters of the neural network. In the teacher data, the learning objective variable data becomes the label (correct data).
FIG. 3 is a diagram conceptually showing an example of the neural network 300.
The input layer 301 of the neural network 300 receives at least the magnitude of the magnetic flux density |B| at a certain time and the time differential value d|B|/dt of the magnitude of the magnetic flux density at that time. Furthermore, the output layer 302 of the neural network outputs at least the value W _ae of the abnormal eddy current loss at the relevant time. The number of intermediate layers (hidden layers) of the neural network 300 is not particularly limited.

The learning unit 105 uses learning explanatory variable data included in the teacher data to calculate the magnitude of the magnetic flux density |B| at a certain time and the time differential value d|B|/ of the magnitude of the magnetic flux density at that time. dt is input to the input layer 301, the abnormal eddy current loss value W _ae output from the output layer 302 at that time is acquired at each time within one cycle. The representative value of the abnormal eddy current loss value W _ae obtained in this way is the representative value in one period of the abnormal eddy current loss obtained from the neural network 300. As the representative value, for example, an average value, a median value, or an effective value can be used.

An objective function is used when learning the parameters of the neural network 300. In this embodiment, the objective function is based on the learning objective variable data (representative value in one cycle of abnormal eddy current loss) included in the teaching data and the representative value in one cycle of abnormal eddy current loss obtained from the neural network 300. is a function that derives at least the evaluation value of the difference between

The learning unit 105 learns the parameters of the neural network 300 so that the value of the objective function becomes the optimal value. For example, the learning unit 105 combines the learning objective variable data (representative value in one cycle of abnormal eddy current loss) included in the teacher data and the representative value in one cycle of the abnormal eddy current loss obtained from the neural network 300. The neural network parameters that minimize the mean squared error are derived as the optimal parameters. In this case, the aforementioned evaluation value is the mean square error.

In this case, the mean square error between the learning objective variable data (representative value in one cycle of abnormal eddy current loss) included in the training data and the representative value in one cycle of abnormal eddy current loss obtained from the neural network 300 is calculated. The derived function can be used as an objective function. A specific example of the objective function J is the following equation (7).

Here, n is a variable that specifies teacher data. N is the number of training data. t _n is learning objective variable data (representative value in one cycle of abnormal eddy current loss) included in the teacher data specified by the variable n. y _n is a representative value in one period of the abnormal eddy current loss obtained from the neural network 300 based on the learning explanatory variable data included in the teacher data specified by the variable n. That is, y _n is the value at each time within one cycle of the magnitude of magnetic flux density included in the teacher data specified by variable n, and the time differential value of the magnitude of magnetic flux density at each time within one cycle. This is a representative value of the abnormal eddy current loss value W _ae at each time within one cycle, which is output from the output layer 302 by inputting the value in the input layer 301 of the neural network 300 .
In equation (7), the optimal value of the objective function J is the minimum value. Note that when the objective function is obtained by multiplying the right side of equation (7) by (-1), the optimal value of the objective function J is the maximum value.

Neural network parameters include weighting coefficients. Depending on the structure of the neural network, other parameters may also be learned. A specific learning method can be realized by a known method such as an error backpropagation method or a method using a metaheuristic algorithm. Therefore, a detailed explanation of the learning method will be omitted here. Note that the optimal value (minimum value, maximum value) of the objective function is the value of the objective function when the value of the objective function is optimal (minimum or maximum) in an algorithm for learning parameters of a neural network. For example, it is possible to determine that the value of the objective function is optimal on the condition that the number of times the objective function J has been derived has reached a predetermined number, or that the amount of change in the objective function J has become less than or equal to a predetermined value. can.

The learning unit 105 learns the parameters of the neural network 300 as described above using teacher data that includes the same information as information indicating the type of electrical steel sheet. Furthermore, if there is training data that includes information indicating mutually different types as information indicating the types of electromagnetic steel sheets, the learning unit 105 learns the parameters of the neural network 300 for each type. In this way, neural network parameters are derived for each type of electrical steel sheet.

As described above, in this embodiment, the parameters of the neural network are learned by using the acquisition unit 101, the first electromagnetic field analysis unit 102, the abnormal eddy current loss derivation unit 103, the teacher data creation unit 104, and the learning unit 105. The functions to do this are realized.
<<Storage unit 106>>
The storage unit 106 stores information indicating the parameters of the neural network 300 learned by the learning unit 105 for each type of electrical steel sheet.

<<Second electromagnetic field analysis section 111>>
The second electromagnetic field analysis unit 111 is activated after information indicating the parameters of the neural network is stored in the storage unit 106.
The second electromagnetic field analysis unit 111 calculates the magnitude of the magnetic flux density and the time differential value of the magnitude of the magnetic flux density of the electrical steel sheet whose abnormal eddy current loss is to be estimated when the electrical steel sheet is excited under predetermined excitation conditions. , derived for each of multiple micro regions (meshes) based on Maxwell's equations. The predetermined excitation conditions here are determined by how to estimate the abnormal eddy current loss when the electromagnetic steel sheet is excited, and are based on the predetermined excitation conditions shown in the section of <<First electromagnetic field analysis section 102>>. The excitation conditions may be different from the excitation conditions. For example, the excitation conditions when driving an electrical device (for example, a motor) to which the electromagnetic steel sheet whose abnormal eddy current loss is to be estimated can be set to be the predetermined excitation conditions. A specific example of the processing of the second electromagnetic field analysis section 111 will be described below.

The second electromagnetic field analysis unit 111 acquires information indicating the excitation conditions of the electromagnetic steel sheet whose abnormal eddy current loss is to be estimated. Forms of acquiring information indicating excitation conditions include, for example, transmission from an external device, reading of measurement data stored in a portable storage medium, and input of measurement data values to the user interface of the abnormal eddy current loss estimating device 100. At least one of the inputs by operation can be adopted.

The second electromagnetic field analysis unit 111 unifies the magnetic flux density vector B and the eddy current vector J _e in each minute region using the finite element method based on Maxwell's equations according to the electromagnetic field analysis conditions including the acquired excitation conditions. Derived at each time in the cycle. The method of calculating the magnetic flux density vector B and the eddy current vector J _e in each minute region can be realized by the same method as the first electromagnetic field analysis section 102 .
The second electromagnetic field analysis unit 111 calculates the magnitude of the magnetic flux density |B| and the magnitude of the magnetic flux density for the microregion at each time in one cycle based on the magnetic flux density vector B in a certain microregion. The time differential value d|B|/dt is derived. The second electromagnetic field analysis unit 111 derives the magnitude of the magnetic flux density |B| and the time differential value of the magnitude of the magnetic flux density d|B|/dt at each time within one cycle, as follows: Execute for each of all microregions. The method of deriving the magnitude of magnetic flux density |B| and the time differential value of the magnitude of magnetic flux density d|B|/dt from the value (instantaneous value) at each time within one cycle of magnetic flux density is as follows. This can be realized by the method described in the section <Teacher data creation unit 104>>. Therefore, a detailed explanation of the method for deriving the magnitude of the magnetic flux density |B| and the time differential value d|B|/dt of the magnitude of the magnetic flux density will be omitted here.

<<Estimation unit 112>>
The estimation unit 112 calculates the magnitude of the magnetic flux density |B| at a certain time derived by the second electromagnetic field analysis unit 111 and the time differential value of the magnitude of the magnetic flux density d|B|/dt using the neural network 300. Based on the value input to the input layer 301 and output from the output layer 302, the abnormal eddy current loss of the electrical steel sheet at the relevant time is estimated. A specific example of the processing of the estimation unit 112 will be described below.

The estimation unit 112 acquires information indicating the type of electrical steel sheet whose abnormal eddy current loss is to be estimated. Examples of methods of obtaining information indicating the type of electrical steel sheet include transmission from an external device, reading of measurement data stored in a portable storage medium, and value of measurement data to the user interface of the abnormal eddy current loss estimating device 100. At least one of the input operations can be adopted. In the following description, the electrical steel sheet whose abnormal eddy current loss is to be estimated will be abbreviated as the electrical steel sheet to be estimated, if necessary.

The estimation unit 112 reads information indicating the parameters of the neural network 300 corresponding to the type of electrical steel sheet to be estimated from the storage unit 106, and sets the parameters of the neural network 300 to the algorithm of the neural network. As a result, the neural network 300 whose parameters have been learned is set.

The estimation unit 112 calculates the magnitude of the magnetic flux density |B| and the time differential of the magnitude of the magnetic flux density, which is derived for a certain minute region at a certain time within one cycle by the second electromagnetic field analysis unit 111. By inputting the value d|B|/dt into the input layer 301 of the neural network 300 where the parameters have been learned, the value W _ae of the abnormal eddy current loss output from the output layer 302 is calculated in the micro region at the time. It is derived as the value W _ae of abnormal eddy current loss.

The estimating unit 112 calculates the abnormal eddy current loss value W _ae in one microregion as described above for each microregion at each time within one cycle. Thereby, at each time within one period, the value W _ae of the abnormal eddy current loss in all the micro regions of the electrical steel sheet to be estimated is derived.

The estimating unit 112 derives the sum of abnormal eddy current loss values W _ae in all minute regions at the same time as the abnormal eddy current loss of the electrical steel sheet to be estimated at the time. The estimating unit 112 calculates the abnormal eddy current loss of the electrical steel sheet to be estimated at one time for each time within one cycle. As a result, the abnormal eddy current loss of the electrical steel sheet to be estimated is derived at each time within one cycle.

Then, the estimating unit 112 derives a representative value in one cycle of the abnormal eddy current loss of the electromagnetic steel sheet to be estimated. The representative value of the abnormal eddy current loss of the electrical steel sheet to be estimated in one cycle is, for example, the average value, median value, or effective value of the values at each time within one cycle of the abnormal eddy current loss of the electrical steel sheet to be estimated. value can be used.

As described above, in this embodiment, the function of estimating abnormal eddy current loss is realized by using the second electromagnetic field analysis section 111 and the estimation section 112.

<<Output section 113>>
The output unit 113 outputs information indicating the abnormal eddy current loss of the electrical steel sheet to be estimated, which has been derived by the estimation unit 112. In this embodiment, the estimating unit 112 derives the abnormal eddy current loss value of the electrical steel sheet to be estimated and its representative value at each time within one period. Therefore, the output unit 113 outputs this information. As the form of outputting the information, for example, at least one of displaying it on a computer display, transmitting it to an external device, and storing it in a storage medium inside or outside the abnormal eddy current loss estimating device 100 is adopted. Can be done.

<Operation flowchart>
Next, an example of a method for learning parameters of the neural network 300 by the abnormal eddy current loss estimating device 100 will be described with reference to the flowchart in FIG. 4. The flowchart in FIG. 4 exemplifies a case where parameters of the neural network 300 for one type of electrical steel sheet are learned. When learning the parameters of the neural network 300 for a plurality of types of electromagnetic steel sheets, the process according to the flowchart of FIG. 4 may be executed for each type of electromagnetic steel sheet.

First, in step S401, the acquisition unit 101 acquires one piece of measurement data. The measurement data includes a set of a measured value of magnetic flux density in one cycle and a measured value of iron loss when the same type of electromagnetic steel sheet is excited under the same excitation conditions. The measurement data also includes information indicating the excitation conditions used to obtain the measurement data and information indicating the type of electrical steel sheet to be measured in the measurement data.

Next, in step S402, the first electromagnetic field analysis unit 102 uses the finite element method based on Maxwell's equations according to the electromagnetic field analysis conditions including the excitation conditions included in the measurement data acquired in step S401. The magnetic flux density vector B and eddy current vector J _e in each minute region are calculated at each time in one cycle. The first electromagnetic field analysis unit 102 calculates representative values of hysteresis loss and classical eddy current loss in one cycle of the electrical steel sheet based on the magnetic flux density vector B and eddy current vector J _e of each minute region at each time in one cycle. Derive.

Next, in step S403, the abnormal eddy current loss deriving unit 103 calculates one period of the hysteresis loss and classical eddy current loss derived in step S402 from the measured value of iron loss included in the measurement data acquired in step S401. The value obtained by subtracting the sum of the representative values in is derived as the representative value (learning objective variable data) in one cycle of the abnormal eddy current loss of the electrical steel sheet.

Next, in step S404, the teacher data creation unit 104 calculates the magnitude of the magnetic flux density within one cycle based on the measured values at each time within one cycle of the magnetic flux density included in the measurement data acquired in step S401. The value at each time and the value at each time within one cycle of the time differential value of the magnitude of the magnetic flux density are derived as explanatory variable data for learning. The teacher data creation unit 104 indicates the type of electromagnetic steel sheet included in the learning explanatory variable data derived in this way, the learning objective variable data derived in step S403, and the measurement data acquired in step S401. Create one piece of teacher data including information.

Next, in step S405, the abnormal eddy current loss estimating device 100 determines whether the necessary number of training data for learning the parameters of the neural network 300 has been created. This number is preset by an operator based on the operation of the user interface of the abnormal eddy current loss estimating device 100, for example. The estimation accuracy of the neural network is improved by using training data under various excitation conditions. Therefore, it is preferable to have a large number of training data. On the other hand, the more training data is created, the more the work burden increases. From this perspective, the number of training data can be determined.

For example, arbitrary teacher data may be prepared without using measurement data. Learning the parameters of the neural network 300 using the teacher data is performed by changing the number of teacher data. Investigate the minimum number of learning data at which the value output from the neural network after learning is approximately constant. The minimum value of the number of learning data can be set to the number necessary for learning the parameters of the neural network 300.

As a result of the determination in step S405, if the number of teacher data necessary for learning the parameters of the neural network 300 has not been created, the process returns to step S401. In step S401, the acquisition unit 101 acquires one measurement data different from the measurement data already acquired in step S401. The measurement data different from the measurement data already acquired in step S401 is measurement data that includes information indicating an excitation condition different from the excitation condition included in the measurement data already acquired in step S401. Then, the processes of steps S402 to S404 are executed based on the newly acquired measurement data.

As described above, the processes of steps S401 to S405 are repeatedly executed until it is determined in step S405 that the necessary number of teacher data for learning the parameters of the neural network 300 has been created. If it is determined in step S405 that the necessary number of teacher data for learning the parameters of the neural network 300 has been created, the process proceeds to step S406.

When the process proceeds to step S406, the learning unit 105 learns the parameters of the neural network 300 using the teacher data created in step S404.
Finally, in step S407, the storage unit 106 stores information indicating the parameters of the neural network 300 learned in step S406. At this time, the storage unit 106 stores information indicating the parameters of the neural network 300 and information indicating the type of electrical steel sheet to be estimated by the neural network 300 in association with each other. When the process in step S407 ends, the process according to the flowchart in FIG. 4 ends.

Next, an example of a method for estimating abnormal eddy current loss by the abnormal eddy current loss estimating device 100 will be described with reference to the flowchart in FIG. 5 . The flowchart in FIG. 5 starts after the process according to the flowchart in FIG. 4 is completed.
First, in step S501, the second electromagnetic field analysis unit 111 acquires information indicating the excitation conditions of the electromagnetic steel sheet to be estimated, and in accordance with the electromagnetic field analysis conditions including the acquired excitation conditions, calculates the finite element based on Maxwell's equation. Using the method, the magnetic flux density vector B and eddy current vector J _e in each minute region are calculated at each time in one cycle. The second electromagnetic field analysis unit 111 calculates the magnitude of the magnetic flux density |B| and the time differential value d| of the magnitude of the magnetic flux density at each time within one cycle based on the magnetic flux density vector B in each minute region. B|/dt is derived for each minute region.

Next, in step S502, the estimation unit 112 calculates the magnitude of the magnetic flux density |B| and the time differential value of the magnitude of the magnetic flux density d|B|/dt derived in step S501 according to the flowchart of FIG. Based on the abnormal eddy current loss value W _ae output from the output layer 302 by inputting it to the input layer 301 of the neural network 300 where the parameters have been learned, the abnormality at each time within one cycle of the electrical steel sheet to be estimated is determined. The eddy current loss and the representative value in one cycle of the abnormal eddy current loss of the electrical steel sheet to be estimated are derived.

The magnitude of the magnetic flux density |B| and the time differential value of the magnitude of the magnetic flux density d|B|/dt, which are input to the input layer 301 of the neural network 300 in step S502, are determined at a certain time and a certain time. This is in two microscopic areas. Therefore, the abnormal eddy current loss value W _ae output from the output layer 302 of the neural network 300 is the value at the time and in the minute area. Therefore, the value W _ae of the abnormal eddy current loss is derived by the neural network 300 for each time and for each minute region.

The estimation unit 112 derives the sum of the abnormal eddy current loss values W _ae at the same time as the abnormal eddy current loss value at the time. The estimation unit 112 derives the abnormal eddy current loss of the electrical steel sheet to be estimated at each time within one cycle by performing such derivation at each time within one cycle. The estimation unit 112 derives a representative value of the values at each time within one cycle of the abnormal eddy current loss of the electrical steel sheet to be estimated as a representative value of the abnormal eddy current loss of the electrical steel sheet to be estimated in one cycle.

Finally, in step S503, the output unit 113 outputs information indicating the abnormal eddy current loss of the electrical steel sheet to be estimated, which was derived in step S502. When the process of step S503 ends, the process according to the flowchart of FIG. 5 ends.

<Example>
Next, an example will be described.
As an example of the invention, the parameters of the neural network 300 were learned by the method described in the present embodiment using measurement data for an electrical steel sheet of the same type as a single electromagnetic steel sheet. Regarding the single electromagnetic steel sheet, a representative value of the abnormal eddy current loss in one cycle was estimated by the method described in this embodiment using the neural network 300 whose parameters were learned.

As a comparative example, abnormal eddy current loss was estimated using the method described in Patent Document 1 for a single electromagnetic steel sheet that is the same as the single electromagnetic steel sheet of the invention example.
In the comparative example, the abnormal eddy current loss was specifically derived as follows.
Regarding the single electromagnetic steel sheet, an abnormal eddy current loss coefficient is derived by the method described in Patent Document 1, and the conductivity is corrected using the abnormal eddy current loss coefficient. Then, for the single electromagnetic steel sheet, the magnetic flux density vector B and eddy current vector J _e in each minute region (mesh) are calculated using the finite element method. Then, the corrected conductivity is given to σ in equation (6), and the entire right side of equation (6) is multiplied by the abnormal eddy current loss coefficient to derive the total eddy current loss. The conductivity before correction is given to derive the classical eddy current loss. Then, the value obtained by subtracting the classical eddy current loss from the total eddy current loss is defined as the abnormal eddy current loss. This abnormal eddy current loss is a typical value in one cycle.

In addition, the value obtained by subtracting the sum of the hysteresis loss and classical eddy current loss of the single electromagnetic steel sheet from the measured value of the iron loss of the single electromagnetic steel sheet was derived as the measured value of abnormal eddy current loss. This abnormal eddy current loss is a typical value in one cycle. The hysteresis loss of a single electromagnetic steel sheet and the classical eddy current loss of a single sheet were derived using equations (5) and (6), respectively.

The abnormal eddy current loss as described above was estimated under different excitation conditions. FIG. 6 shows the results of comparing estimated values of abnormal eddy current loss and measured values of abnormal eddy current loss in the invention example and comparative example.
The values shown in FIG. 6 are the error rate (=estimated value÷measured value−1) of the estimated value of abnormal eddy current loss with respect to the measured value, expressed as a percentage.
Waveforms 1, 2, and 3 indicate time waveforms of magnetic flux density. Waveform 1 has a fundamental frequency of 50 Hz and a maximum value of 1.0 T, indicating that it is a temporal waveform with relatively few harmonics. Waveform 2 has a fundamental frequency of 50 Hz and a maximum value of 1.5 T, indicating that it is a time waveform with relatively many harmonics. Waveform 3 has a fundamental frequency of 1 kHz and a maximum value of 1.5 T, indicating that it is a time waveform with relatively many harmonics.

As shown in FIG. 6, it can be seen that in the invention example, abnormal eddy current loss can be estimated with higher accuracy than in the comparative example. In particular, in the example of the invention, abnormal eddy current loss can be accurately estimated even when the fundamental frequency of the time waveform of magnetic flux density is high or when harmonic components are superimposed on the time waveform of magnetic flux density.
Further, in the comparative example, only a representative value in one cycle can be derived as the abnormal eddy current loss. In contrast, in the example of the invention, a value at each time within one period can be derived as the abnormal eddy current loss.

<Summary>
As described above, in this embodiment, the abnormal eddy current loss estimating device 100 calculates the magnitude of magnetic flux density at a certain time |B| and the time differential value d|B|/dt of the magnitude of magnetic flux density at that time. The parameters of the neural network 300 are learned, which has an input layer 301 to which the values are input, and an output layer 302 to which the abnormal eddy current loss value W _ae at the time is output. The abnormal eddy current loss estimating device 100 uses the learned neural network 300 to estimate the abnormal eddy current loss value W _ae at each time. Therefore, even when harmonics are superimposed on the time waveform of magnetic flux density or when the time waveform of magnetic flux density has a high frequency, abnormal eddy current loss can be derived with high accuracy. Furthermore, the abnormal eddy current loss at each time in one cycle can be derived. Therefore, the abnormal eddy current loss in the electrical steel sheet can be accurately derived.

<Modified example>
In this embodiment, the case where the magnetic flux density is a physical quantity obtained from magnetization characteristics is exemplified. However, as described above, at least one of magnetic flux density, magnetic field strength, magnetic permeability, and differential magnetic permeability can be used as the physical quantity obtained from the magnetization characteristics. However, the magnetic flux density and the magnetic field strength are physical quantities that correspond to each other, and it is thought that using one and not using the other will not have a large effect on the estimation accuracy of abnormal eddy current loss. Furthermore, if both magnetic flux density and magnetic field strength are used, there is a risk that the processing load will increase when learning the parameters of the neural network and estimating abnormal eddy current loss using the neural network. Therefore, when magnetic flux density is used as a physical quantity obtained from magnetization characteristics, it is preferable not to use magnetic field strength, and conversely, when magnetic field strength is used, it is preferable not to use magnetic flux density.

In this embodiment, a learning device that realizes the function of learning neural network parameters (acquisition unit 101, first electromagnetic field analysis unit 102, abnormal eddy current loss derivation unit 103, teacher data creation unit 104, and learning unit 105) and a storage device that realizes the function of storing the neural network parameters (storage unit 106), and a storage device that realizes the function of storing the parameters of the neural network (the storage unit 106), and the abnormal eddy current loss that realizes the function of estimating the abnormal eddy current loss (the second electromagnetic field analysis unit 111 and the estimation unit 112). The case where it is included inside the current loss estimating device is illustrated. However, these functions may be realized by multiple devices.

For example, a learning device that realizes a function of learning the parameters of a neural network (an acquisition unit 101, a first electromagnetic field analysis unit 102, an abnormal eddy current loss derivation unit 103, a teacher data creation unit 104, and a learning unit 105) and a neural A storage device that realizes a function of storing network parameters (storage unit 106), and an abnormal eddy current loss estimation device that realizes a function of estimating abnormal eddy current loss (second electromagnetic field analysis unit 111 and estimation unit 112). An abnormal eddy current loss estimation system may be constructed. Further, a storage device may be included inside the learning device.

In this embodiment, a derivation unit that derives the magnitude of the magnetic flux density and the time differential value of the magnitude of the magnetic flux density of the electrical steel sheet whose abnormal eddy current loss is to be estimated is excited under predetermined excitation conditions. Although the case where the second electromagnetic field analysis section 111 is applied has been described, the derivation section is not limited to the second electromagnetic field analysis section 111. In other words, it is only necessary to derive the magnitude of the magnetic flux density of the electrical steel sheet whose abnormal eddy current loss is to be estimated and the time differential value of the magnitude of the magnetic flux density when the electrical steel sheet is excited under predetermined excitation conditions. These values may be derived from the experimental results. For example, an excitation coil and a detection coil may be wound around an electromagnetic steel plate, and the magnetic flux density may be measured based on the induced electromotive force generated in the detection coil when an excitation current is passed through the excitation coil. In addition, when the second electromagnetic field analysis unit 111 is applied, the magnitude of the magnetic flux density of the electromagnetic steel sheet and the time differential value of the magnitude of the magnetic flux density are divided into multiple micro regions (mesh) based on Maxwell's equation. Although it has been shown that it can be derived for each, it is not limited to this. For example, the magnitude of the magnetic flux density and the time differential value of the magnitude of the magnetic flux density may be derived without mesh division using an equivalent circuit of an electromagnetic steel sheet that is excited under predetermined excitation conditions. As explained at the beginning of the <Modifications> section, it is not necessary to use the magnetic flux density when doing the above. When detecting the magnetic field strength through experiments, for example, an H coil may be used.

Note that the embodiments of the present invention described above can be realized by a computer executing a program. Furthermore, a computer-readable recording medium on which the program is recorded and a computer program product such as the program can also be applied as an embodiment of the present invention. As the recording medium, for example, a flexible disk, hard disk, optical disk, magneto-optical disk, CD-ROM, magnetic tape, nonvolatile memory card, ROM, etc. can be used.
Furthermore, the embodiments of the present invention described above are merely examples of implementation of the present invention, and the technical scope of the present invention should not be construed as limited by these. It is something. That is, the present invention can be implemented in various forms without departing from its technical idea or its main features.

100: Abnormal eddy current loss estimation device, 101: Acquisition unit, 102: First electromagnetic field analysis unit, 103: Abnormal eddy current loss derivation unit, 104: Teacher data creation unit, 105: Learning unit, 106: Storage unit, 111 : second electromagnetic field analysis section, 112: estimation section, 113: output section

Claims (9)

A learning device for learning neural network parameters for estimating abnormal eddy current loss in soft magnetic materials,
Measured values for the soft magnetic material when the soft magnetic material is excited under predetermined excitation conditions include measured values at each time within one period of a physical quantity obtained from the magnetization characteristics of the soft magnetic material, and the soft magnetic material. an acquisition means for acquiring a measured value of iron loss;
A first electromagnetic field analysis means for deriving representative values in one cycle of hysteresis loss and classical eddy current loss of the soft magnetic material when the soft magnetic material is excited under the predetermined excitation conditions based on Maxwell's equations. and,
Based on the measured value of iron loss acquired by the acquisition means and the representative values of hysteresis loss and classical eddy current loss in one cycle derived by the first electromagnetic field analysis means, abnormal eddy in the soft magnetic material is determined. Abnormal eddy current loss derivation means for deriving a representative value of current loss in one cycle;
A typical value of the abnormal eddy current loss in the soft magnetic material derived by the abnormal eddy current loss deriving means in one cycle, a value of the magnitude of the physical quantity at each time within one cycle, and a value of the magnitude of the physical quantity at each time within one cycle. a value at each time within one cycle of the time differential value, and a teacher data creation means for creating teacher data based on the time differential value at each time within one cycle;
a learning means for learning parameters of a neural network using the teacher data created by the teacher data creation means,
At least the magnitude of the physical quantity at a certain time and the time differential value of the magnitude of the physical quantity at the time are input to the input layer of the neural network,
At least the abnormal eddy current loss at the time is output from the output layer of the neural network,
The learning means is configured to calculate a representative value of the abnormal eddy current loss of the soft magnetic material in one period included in the training data and an abnormal eddy current loss at each time in one period output from the output layer of the neural network. A learning device, characterized in that the parameters of the neural network are learned using an objective function that at least obtains an evaluation value of a difference between a representative value and a representative value. The learning device according to claim 1, wherein the physical quantity includes at least one of magnetic flux density, magnetic field strength, magnetic permeability, and differential magnetic permeability. An abnormal eddy current loss estimating device that estimates abnormal eddy current loss of a soft magnetic material using the neural network whose parameters have been learned by the learning device according to claim 1 or 2,
Derivation means for deriving the magnitude of the physical quantity of the soft magnetic material and the time differential value of the magnitude of the physical quantity when the soft magnetic material whose abnormal eddy current loss is to be estimated is excited under predetermined excitation conditions;
The magnitude of the physical quantity at a certain time derived by the derivation means and the time differential value of the magnitude of the physical quantity at the time are given to the input layer of the neural network to obtain the value output from the output layer. an estimation means for estimating the abnormal eddy current loss of the soft magnetic material at the relevant time based on the above information. The derivation means maximizes the magnitude of the physical quantity of the soft magnetic material and the time differential value of the magnitude of the physical quantity when the soft magnetic material whose abnormal eddy current loss is to be estimated is excited under predetermined excitation conditions. The abnormal eddy current loss estimating device according to claim 3, characterized in that the second electromagnetic field analysis means derives the electromagnetic field for each of a plurality of micro regions based on Well's equation. The estimating means provides the magnitude of the physical quantity at a certain time derived by the deriving means and the time differential value of the magnitude of the physical quantity at the time to the input layer of the neural network, thereby estimating the magnitude of the physical quantity from the output layer. Obtaining the output abnormal eddy current loss at the relevant time is executed at each time within one cycle, and based on the abnormal eddy current loss at each acquired time, the abnormal eddy current loss of the soft magnetic material is calculated. The abnormal eddy current loss estimating device according to claim 3 or 4, wherein the abnormal eddy current loss estimating device estimates a representative value in one cycle. A learning method for learning neural network parameters for estimating abnormal eddy current loss in soft magnetic materials, the method comprising:
Measured values for the soft magnetic material when the soft magnetic material is excited under predetermined excitation conditions include measured values at each time within one period of a physical quantity obtained from the magnetization characteristics of the soft magnetic material, and the soft magnetic material. an acquisition step of acquiring a measured value of iron loss;
A first electromagnetic field analysis step of deriving representative values of hysteresis loss and classical eddy current loss in one cycle of the soft magnetic material when the soft magnetic material is excited under the predetermined excitation conditions based on Maxwell's equations. and,
Based on the measured value of iron loss acquired in the acquisition step and the representative values in one cycle of hysteresis loss and classical eddy current loss derived in the first electromagnetic field analysis step, abnormal eddy in the soft magnetic material is determined. An abnormal eddy current loss derivation step of deriving a representative value of current loss in one cycle;
A representative value in one cycle of the abnormal eddy current loss in the soft magnetic material derived by the abnormal eddy current loss deriving step, a value at each time within one cycle of the magnitude of the physical quantity, and a value of the magnitude of the physical quantity at each time within one cycle. A teacher data creation step of creating teacher data based on the value at each time within one cycle of the time differential value;
a learning step of learning parameters of a neural network using the teacher data created in the teacher data creation step;
At least the magnitude of the physical quantity at a certain time and the time differential value of the magnitude of the physical quantity at the time are input to the input layer of the neural network,
At least the abnormal eddy current loss at the time is output from the output layer of the neural network,
The learning step includes determining a representative value of the abnormal eddy current loss of the soft magnetic material in one cycle included in the teacher data and an abnormal eddy current loss at each time in one cycle output from the output layer of the neural network. A learning method characterized in that the parameters of the neural network are learned using an objective function that at least obtains an evaluation value of the difference between the 1 and the representative value. An abnormal eddy current loss estimation method for estimating abnormal eddy current loss of a soft magnetic material using the neural network whose parameters have been learned by the learning method according to claim 6,
a derivation step of deriving the magnitude of the physical quantity of the soft magnetic material and the time differential value of the magnitude of the physical quantity when the soft magnetic material whose abnormal eddy current loss is to be estimated is excited under predetermined excitation conditions;
The magnitude of the physical quantity at a certain time derived in the derivation step and the time differential value of the magnitude of the physical quantity at the time are given to the input layer of the neural network to obtain the value output from the output layer. an estimation step of estimating the abnormal eddy current loss of the soft magnetic material at the relevant time based on the above. A program for causing a computer to function as each means of the learning device according to claim 1 or 2. A program for causing a computer to function as each means of the abnormal eddy current loss estimating device according to any one of claims 3 to 5.

Images