WO2020217879A1 - Power conversion apparatus, machine learning device, and learned model generation method - Google Patents

Abstract

A machine learning device (10) is configured to carry out machine learning on the basis of a current command value and a current detection value detected by a current detection unit, or a voltage command value which are all included in teacher data, to thereby generate a pattern generation function for determining a switching pattern of a power conversion unit (12). A pattern determination unit (14) receives, as inputs, the pattern generation function provided from the machine learning device (10) together with the voltage command value or the current command and current detection values, carries out an arithmetic operation, and determines the switching pattern of the power conversion unit (12).

Description

Power converters, machine learners, and how to generate trained models

The present application relates to a power converter, a machine learner, and a method of generating a trained model.

Conventionally, in the drive control of a multi-phase AC rotary mechanical device, there is an instantaneous current control that directly determines the switching state of the power converter based on the drive state of the rotary mechanical device, and model prediction control is one of the instantaneous current controls. It has been known.

For example, in the prior art described in Patent Document 1 below, direct torque control based on model predictive control is being studied. This control method is a method of controlling within a predetermined allowable range while predicting the torque of the rotating mechanical device and the stator magnetic flux several steps ahead, and the number of switching of each phase switch is the minimum in the prediction interval of several steps. Search for a switching pattern. Therefore, it is expected to reduce the switching loss in the steady state while maintaining the high-speed torque response time in the transient state by the model prediction control.

Further, in the prior art described in Patent Document 2 below, a feedback control system equipped with a machine learning device is being studied. The machine learner adjusts the time for maintaining the switching state with respect to the switching state of the power converter calculated by the feedback control and the time for maintaining the switching state. This is expected to have an effect of suppressing vibration generated in a transient state.

Japanese Unexamined Patent Publication No. 2011-152038 Japanese Unexamined Patent Publication No. 2012-21258

Direct torque control based on model predictive control as described in Patent Document 1 is expected to reduce steady-state switching loss while maintaining a high-speed torque response time in the transient state. However, if the effect of reducing the switching loss is to be further enhanced, a longer section must be predicted, and if the predicted section is unnecessarily expanded, the amount of calculation increases exponentially. When considering the above, there is naturally a limit to expanding the predicted interval and reducing the switching loss.

Further, in the feedback control system provided with the machine learning device as described in Patent Document 2, the vibration generated in the transient state is suppressed, but the machine learning device only has the time to maintain the switching state of the power conversion unit. It is difficult to consider the switching loss, current harmonics, and drive noise related to the modulation method of the power converter because the switching state of the power converter is not adjusted or determined.

This application discloses a technique for solving the above-mentioned problems, and while considering the performance related to the modulation method of the power conversion unit, it is adjusted to the user's request and the state of the rotating mechanical device to be controlled. It is an object of the present invention to provide a power converter, a machine learner, and a method of generating a trained model capable of determining a switching pattern.

The power conversion device disclosed in the present application includes a plurality of switching elements, a power conversion unit that converts DC power into AC power and supplies it to a rotary mechanical device, and a current detection unit that detects a current flowing through the rotary mechanical device. And, for each set control cycle, the switching state for one cycle in the plurality of switching elements is determined based on the pattern generation function given by the machine learner, and the switching state for one cycle is combined. A pattern determination unit that generates a switching pattern and controls the output of the power conversion unit is provided.
The machine learning device executes machine learning based on either the current command value, the current detection value detected by the current detection unit, or the voltage command value included in the teacher data, and the power conversion unit of the power conversion unit. It generates the pattern generation function that determines the switching pattern.
The pattern determination unit inputs both the pattern generation function given by the machine learning device and either the current command value and the current detection value or the voltage command value to execute arithmetic processing. It determines the switching pattern of the power conversion unit.

Further, the machine learning device disclosed in the present application is a switching state for one cycle of a set control cycle of a plurality of switching elements constituting a power conversion unit that converts DC power into AC power and supplies it to a rotating machine device. A pattern generation function that determines a switching pattern consisting of a combination of is executed and output by machine learning based on the teacher data.
An input data acquisition unit that acquires either the current command value included in the teacher data, the current detection value detected by the current detection unit that detects the current flowing through the rotating mechanical device, or the voltage command value as input data. ,
A label acquisition unit that acquires the switching pattern of the power conversion unit included in the teacher data as a label, and
The switching of the power conversion unit based on either one of the current command value and the current detection value or the voltage command value obtained by the input data acquisition unit and a switching pattern obtained by the label acquisition unit. It includes a learning unit that generates a trained model that determines the switching pattern of the element.

Further, the trained model generation method disclosed in the present application has been learned to determine the switching pattern of the switching element constituting the power conversion unit by performing machine learning using the machine learning device. Generate a model.

According to the power conversion device, the machine learning device, and the trained model generation method disclosed in the present application, the state of the rotating machine device to be controlled and the user's request while considering the performance related to the modulation method of the power conversion unit. The switching pattern can be determined according to the above.

It is a block diagram which shows the structure of the power conversion apparatus by Embodiment 1 of this application. It is a hardware block diagram which realizes the power conversion apparatus by Embodiment 1 of this application. It is a flowchart which shows the operation example of the power conversion apparatus by Embodiment 1 of this application. It is a block diagram which shows the structure of the machine learning apparatus according to Embodiment 1 of this application. It is a hardware block diagram which realizes the machine learning device according to Embodiment 1 of this application. It is a block diagram which shows the structure of the power conversion apparatus by Embodiment 2 of this application. It is a flowchart which shows the operation example of the power conversion apparatus by Embodiment 2 of this application. It is a block diagram which shows the structure of the power conversion apparatus according to Embodiment 3 of this application. It is a flowchart which shows the operation example of the power conversion apparatus according to Embodiment 3 of this application. It is a block diagram which shows the structure of the power conversion apparatus according to Embodiment 4 of this application. It is a hardware block diagram which realizes the machine learning device according to Embodiment 4 of this application. It is a block diagram which shows the structure of the power conversion apparatus according to Embodiment 5 of this application. It is a block diagram which shows the structure of the power conversion apparatus according to Embodiment 6 of this application. It is a figure which shows an example of the switching state of the power conversion part by Embodiment 1 of this application. It is a figure explaining the switching pattern by Embodiment 1 of this application.

Embodiment 1.
Regarding the configuration and operation of the power conversion device and the machine learning device according to the first embodiment, FIG. 1 is a block diagram showing the configuration of the power conversion device, FIG. 2 is a hardware configuration diagram for realizing the power conversion device, and power is supplied. The operation of the conversion device and the machine learning device will be described with reference to FIG. 3, which is a flowchart showing the operation, FIG. 4, which is a block diagram showing the configuration of the machine learning device, and FIG. 5, which is a hardware configuration diagram for realizing the machine learning device.

As shown in FIG. 1, the entire system of the first embodiment is composed of a power conversion device 1, a DC power supply 2, and a rotary mechanical device 3.

The power conversion device 1 is connected between the DC power supply 2 and the rotary mechanical device 3, converts the DC power from the DC power supply 2 into AC power, outputs the AC power to the rotary mechanical device 3, and drives the rotary mechanical device 3. .. The rotary mechanical device 3 converts the AC power output from the power conversion device 1 into power. As the rotary mechanical device 3 used here, various electric motors such as an induction motor and a synchronous motor can be used.

The power conversion device 1 includes a machine learning device 10, a uvw / dq converter 11, a power conversion unit 12, which is a main circuit, a current detection unit 13, and a pattern determination unit 14.

The current detection unit 13 detects the current values iu (k), iv (k), and iw (k) for the three phases output by the power conversion unit 12 to the rotary mechanical device 3. The uvw / dq converter 11 converts the detected current values iu (k), iv (k), and iw (k) into current values id (k) and iq (k) on the dq coordinates.

The pattern determination unit 14 switches the current values id (k), iq (k), current command values idref (k), iqref (k), which are the outputs of the uvw / dq converter 11, and the previous cycle for each control cycle. The switching pattern SWP of the power conversion unit 12 is based on the states SWu (k-1), SWv (k-1), SWw (k-1), and the pattern generation function output by the machine learning device 10 described in detail later. (K) is determined. In the drawings of the first embodiment (FIGS. 1, 3, and 4), and also in the drawings other than the first embodiment, the pattern generation function is represented as PGF for simplification of notation.

Further, the switching pattern SWP (k) is composed of a combination of switching states for one cycle of the control cycle, and the details will be described later. The notations (k-1) and (k) represent discrete-time signals for each control cycle, where (k-1) is the previous value, (k) is the current value, and (k + 1) is the next value. is there. This also applies to FIGS. 1 and 1 and subsequent figures.

When the detected current values iu (k), iv (k), and iw (k) are collectively described, the current detected value iuvw (k) is appropriately described. When id (k) and iq (k), which are the current values on the dq coordinates, are described together, they are appropriately described as the dq coordinate current value idq (k). When the current command values idref (k) and iqref (k) are described together, they are appropriately described as the current command values idqref (k). When the switching states SWu (k-1), SWv (k-1), and SWw (k-1) of the previous cycle (k-1) are described together, the switching state SW (k-1) of the previous cycle is appropriately described. ). These notations are the same in FIGS. 1 and 1 and subsequent figures.

The power conversion device 1 is realized by, for example, the hardware configuration shown in FIG.
The power conversion device 1 is composed of a power conversion unit 12, a current detection unit 13, a processor 20 that controls the power conversion unit 12, and a storage device 21 included in the processor 20.

The power conversion unit 12 is composed of a three-phase inverter circuit that converts the DC power of the DC power supply 2 into three-phase AC power, and drives a rotating mechanical device 3 such as an electric motor, which is a load. The power conversion unit 12 includes a plurality of switching elements Q1 to Q6 in which diodes D are connected in antiparallel. In this example, the U-phase upper and lower arms are provided with switching elements Q1 and Q2, the V-phase upper and lower arms are provided with switching elements Q3 and Q4, and the W-phase upper and lower arms are switching elements Q5. And Q6. Then, from the connection point between the upper arm and the lower arm of each phase, the bus bar is connected to the input terminal of each phase of the rotary mechanical device 3.

The storage device 21 includes a volatile storage device such as a RAM (Random Access Memory) and a non-volatile auxiliary storage device such as an HDD (Hard Disk Drive) and an SSD (Solid State Drive) (all of which are not shown). .. As the non-volatile auxiliary storage device, a flash memory or the like may be used instead of the HDD.

The processor 20 executes the control program input from the storage device 21.
Since the storage device 21 includes an auxiliary storage device and a volatile storage device, a control program is input to the processor 20 from the auxiliary storage device via the volatile storage device.
The processor 20 may output data such as a calculation result to the volatile storage device of the storage device 21, or may store these data in the auxiliary storage device via the volatile storage device.

As described above, the pattern determination unit 14 controls the power conversion unit 12 by outputting a switching pattern SWP (k) composed of a combination of switching states for one cycle for each control cycle.
FIG. 14 is a diagram showing an example of a switching state of the power conversion unit 12. The switching state is a combination of on (: 1) and off (: 0) signals of the switching elements Q1 to Q6. Of the switching elements Q1 to Q6 of the upper arm and the lower arm, eight switching states (SW1 to SW8) in which one is on and the other is off, and all the switching elements Q1 to Q6 are switched when the power conversion device 1 is stopped. There are nine switching states (SW0) that are turned off. In the switching state SW1 (SWu1, SWv1, SWw1) of FIG. 14, the switching elements Q1, Q4, and Q6 are on, and the switching elements Q2, Q3, and Q5 are off.

FIG. 15 is a diagram illustrating a switching pattern SWP (k). The switching pattern SWP (k) is a combination of switching states for one cycle, and one cycle T is divided into a plurality of sections, and the switching state assigned to each section is determined. In this case, the switching pattern SWP (k) is a combination set to switch in the order of the switching states SW3, SW4, SW2, SW2, SW6, and SW7.
It should be noted that one cycle may be divided into sections having a predetermined width and each switching state may be assigned, or timing information for switching the switching state to a different switching state may be added to the switching state information.

As shown in FIG. 15, the power conversion unit 12 is operating in the switching state SW1 at the present time t (k), and the previous switching state is SW7. The switching state SW1 at the present time t (k) also continues from the previous cycle (k-1), and the switching states SW1 and SW7 are the switching states SW (k-1) within the previous cycle (k-1). ..

The pattern determination unit 14 includes switching states SW (k-1) in at least one previous cycle (k-1) including the switching state SW1 at the present time t (k), such as SW1 and SW7, and a current value idq (k). ) And the current command value idqref (k), the switching pattern SWP (k) (: SW3, SW4, SW2, SW2, SW6, SW7) for one cycle of the current cycle (k) is performed by the pattern generation function PGF. ) Is generated. The switching pattern SWP (k) is given to the power conversion unit 12 as a command of the switching state for one cycle, and the switching elements Q1 to Q6 are on / off controlled.
Then, the pattern determination unit 14 generates a switching pattern SWP (k + 1) for the next cycle (k + 1) at the time point t (k + 1).

In this case, although the pattern determination unit 14 receives and acquires the actual switching state SW (k-1) in the previous cycle (k-1) from the power conversion unit 12, the pattern determination unit 14 has shown. The switching state SW (k-1) in the previous cycle (k-1) may be acquired from the switching pattern SWP (k-1) generated in the previous cycle.

Since one cycle T of the control cycle is usually much longer than the duration of one switching state, the switching pattern SWP is composed of a combination of a plurality of switching states. However, if the control cycle can be shortened to the same level as the duration of the switching state, one switching state may be set to one cycle.

Further, the switching state SW (k-1) in the previous cycle can be applied only in the current switching state, but it is desirable that there are a plurality of switching states. When the control cycle is short, the switching state within the previous cycle is not limited to the immediately preceding cycle (k-1), and the switching state of, for example, two previous cycles (k-2) may be adopted together.

Next, the functions and operations of each part of the power conversion device 1 will be described with reference to FIG.
The power conversion unit 12 converts the DC power supplied from the DC power supply 2 into AC power based on the switching pattern SWP (k) determined by the pattern determination unit 14, and outputs the DC power to the rotary mechanical device 3. The method for determining the switching pattern SWP (k) will be described later.

The current detection unit 13 detects a three-phase alternating current between the power conversion unit 12 and the rotary mechanical device 3, and outputs this as a current detection value iuvw (k) to the uvw / dq converter 11.
Here, any current detection unit such as a CT (current transformer) detector, a shunt resistor, or the like may be used as the current detection unit 13. Of the three-phase currents, the currents of two phases may be detected and the current of the remaining one phase may be calculated. Further, a one-shunt current detection method that restores the three-phase alternating current value by one current detection unit may be used.

The uvw / dq converter 11 converts the current value iuvw (k) detected by the current detection unit 13 into the current value idq (k) on the biaxial dq coordinates and outputs the current value idq (k) to the pattern determination unit 14. At this time, the phase generated in the power converter 1 can be used as the phase information of the magnetic pole position of the rotating mechanical device 3 required for the uvw / dq converter 11. When a phase and speed detector such as an encoder is installed in the rotary mechanical device 3, the detected phase may be used.

In the first embodiment, since the current command value is an example of idqref (k) which is the current command value on the dq coordinate, the current detection value iuvw (k) is set to the current value idq (k) on the dq coordinate. Is being converted to. If the current command value is the command value iuref (k), ivref (k), iwref (k) of the three-phase alternating current, the uvw / dq converter 11 does not perform coordinate conversion of the current detection value iuvw (k). It may be output to the pattern determination unit 14 as it is.

If the current command values are the two-phase alternating currents iαref (k) and iβref (k), the current detection value iuvw (k) is set to αβ by using the uvw / αβ converter instead of the uvw / dq converter 11. The current values iα (k) and iβ (k) on the coordinates may be converted and output to the pattern determination unit 14.

Next, an operation example of the power conversion device 1 of the first embodiment will be described with reference to a flowchart showing a processing procedure of FIG. The processing procedure described below is an example of the learning method and the electric motor control method of the present application. Therefore, each process may be changed as much as possible, and the process can be omitted, replaced, and added as appropriate according to the embodiment.

First, in step S1, it is determined whether to execute machine learning A or motor control B. When executing machine learning A (step S1: Yes), machine learning A is performed to create a trained model. When the machine learning A is not executed (step S1: No), the motor control B is executed using the trained model in which the machine learning A is performed. In this case, since the processing contents are different between the processing when the machine learning A is performed and the processing when the electric motor control B is performed, the processing procedure when the machine learning A is performed will be described first.

Machine learning A is executed according to the configuration of the machine learning device 10 of FIG.
As shown in FIG. 4, the machine learning device 10 includes an input data acquisition unit 10a, a label acquisition unit 10b, a learning unit 10c, and a pattern generation function storage unit 10d.

In carrying out machine learning A, the machine learning device 10 performs learning with teacher data based on teacher data prepared in advance. Learning with teacher data will be described later.
Here, the control method acquired as the teacher data is a so-called pulse width modulation (PWM) method in which the three-phase voltage command value is normalized and the switching pattern SWP (k) is determined by the triangular wave carrier comparison modulation method. In comparison, it is a control method that reduces the switching loss of the power conversion unit 12, and is, for example, model predictive control (Model Predictive Control), selective harmonic elimination, and low-order harmonic elimination (Low-order Harmonic). Elimination), optimal pulse pattern (Optimized Pulse Patterns), and other control methods.

In step S2, the input data acquisition unit 10a of the machine learning device 10 has the current command value idqref (k), the dq coordinate current value idq (k), and the switching state SW (k-) of the previous cycle from the teacher data prepared in advance. 1) is acquired as input data and output to the learning unit 10c. The switching state SW (k-1) of the previous cycle used as input data is data obtained when, for example, model prediction control is performed in advance, and is the switching element Q1 of the power conversion unit 12 of FIG. This is data for controlling on / off of Q6.

In step S3, the label acquisition unit 10b of the machine learning device 10 acquires the switching pattern SWP (k) as a label from the teacher data prepared in advance and outputs it to the learning unit 10c.

In step S4, the learning unit 10c of the machine learning device 10 includes a set of data (hereinafter, a teacher data set) including input data input from the input data acquisition unit 10a and a label input from the label acquisition unit 10b. Acquire as) and execute learning with teacher data.

The learning unit 10c builds a learned model by performing learning with teacher data based on the teacher data set input as described above.

The machine learning A for the electric motor control B in the first embodiment is learning with teacher data by a neural network configured by combining perceptrons. Specifically, a teacher data set consisting of input data indicating the motor state and a label corresponding to the motor state is given to the neural network, and weighting is performed for each perceptron so that the output of the neural network is the same as the label. Repeat learning while changing.

In the learning process, the weighting value is adjusted so as to reduce the output error of each perceptron by repeating the process of backpropagation (also called backpropagation, error backpropagation method).

In this way, the characteristics of the teacher data set are learned, and a trained model for estimating the result from the input is inductively acquired. That is, in the learning with teacher data, as described above, the error between the label and the output data is eliminated while adjusting the weighting value.

In this way, the learning with teacher data carried out by the learning unit 10c controls the switching elements Q1 to Q6 of the power conversion unit 12 so that the switching loss is smaller than that of the pulse width modulation (PWM) method as the learning result. A trained model for determining the switching pattern SWP (k) to be used is obtained. Then, the trained model constructed by the learning unit 10c is output to the pattern generation function storage unit 10d in the next stage.

The neural network used by the learning unit 10c for learning may have three layers, but the number of layers may be further increased. Learning may be performed by so-called deep learning (also called deep learning).

In step S5, the pattern generation function storage unit 10d of the machine learning device 10 stores the learned model obtained by learning with teacher data in the learning unit 10c as the pattern generation function PGF. The pattern generation function PGF may be updated by periodically executing the processes of steps S2 to S5.

The pattern generation function PGF stored in the pattern generation function storage unit 10d is output to the pattern determination unit 14 when the motor control B described later is executed. Then, the pattern determination unit 14 switches the power conversion unit 12 based on the pattern generation function PGF, the dq coordinate current value idq (k), the current command value idqref (k), and the switching state SW (k-1) of the previous cycle. The pattern SWP (k) is determined.

The machine learning device 10 for realizing the above-mentioned processing is realized by, for example, the hardware configuration shown in FIG. That is, the machine learning device 10 is composed of a processor 30 and a storage device 31 included in the processor 30.

The storage device 31 includes a volatile storage device 311 such as a RAM (Random Access Memory) and a non-volatile auxiliary storage device 312 such as an HDD (Hard Disk Drive) and an SSD (Solid State Drive). As the non-volatile auxiliary storage device 312, a flash memory or the like may be used instead of the HDD.

The processor 30 executes various learning programs input from the storage device 31.
Since the storage device 31 includes the volatile storage device 311 and the auxiliary storage device 312, various learning programs are input to the processor 30 from the auxiliary storage device 312 via the volatile storage device 311.

The processor 30 may output data such as the learning result of the learning program to the volatile storage device 311 of the storage device 31, or stores these data in the auxiliary storage device 312 via the volatile storage device 311. May be good.

The learning program is a program including instructions for causing the processor 30 of the machine learning device 10 to execute the learning process with teacher data and to generate the learning result data as a result of the machine learning A. The teacher data is subjected to machine learning A by the machine learning device 10 so as to acquire the switching pattern SWP (k) of the power conversion unit 12 for reducing the switching loss of the power conversion unit 12 as compared with the pulse width modulation (PWM) method. It is data to carry out.

The machine learning device 10 can be realized by a PC (Personal Computer), a server device, or the like. However, since the machine learning device 10 has a large amount of calculation associated with machine learning A, for example, a GPU (Graphics Processing Units) is mounted on a PC, and a GPU called GPGPU (General-Purpose computing on Graphics Processing Units) is used. May be used for the arithmetic processing associated with the machine learning A so that the processing can be performed at high speed.

Regarding the specific hardware configuration of the machine learning device 10, it is possible to omit, replace, and add components as appropriate according to various embodiments. For example, the machine learning device 10 may include a plurality of processors. Further, the processor 30 may be composed of a CPU (Central Processing Unit), an FPGA (Field-Programmable Gate Array), or the like.

Next, the processing content of the electric motor control B performed using the trained model after the machine learning A is performed as described above will be described by returning to the flowchart shown in FIG.

First, in step S1, it is determined whether to execute machine learning A or motor control B. Since the processing when the machine learning A is performed has been described above, the processing content when the motor control B is performed (step S1: No) will be described here.

In step S6, the pattern determination unit 14 acquires the pattern generation function PGF, which is a learned model stored in the pattern generation function storage unit 10d of the machine learning device 10.

Next, in step S7, the pattern determination unit 14 uses the current command value idqref (k), the dq coordinate current value idq (k), and the switching state SW (k-1) of the previous cycle of the power conversion unit 12 as input data. get.

In step S8, the pattern determination unit 14 uses the input data (current command value idqref (k), dq coordinate current value idq (k), switching state SW (k-1) of the previous cycle of the power conversion unit 12), and the machine. The switching pattern SWP (k) is generated based on the pattern generation function PGF obtained from the learner 10. Then, the generated switching pattern SWP (k) is output to the power conversion unit 12.

In step S9, the power conversion unit 12 supplies AC power to the rotary machine device 3 based on the switching pattern SWP (k) output from the pattern determination unit 14, and the rotary machine device 3 supplies the current command value idqref ( It is driven so as to make the switching loss of the power conversion unit 12 smaller than that of the pulse width modulation (PWM) method while following the dq coordinate current value idq (k) on the dq coordinate with respect to k).

As described above, the power conversion device 1 of the first embodiment has a current detection unit 13 that detects the current flowing through the rotary machine device 3 and a machine learner 10 that outputs a pattern generation function for determining a switching pattern. Switching based on the current command value idqref (k), the dq coordinate current value idq (k), the switching state SW (k-1) of the previous cycle of the power conversion unit 12, and the pattern generation function from the machine learner 10. The machine learner 10 includes a pattern determination unit 14 that determines a pattern and a power conversion unit 12 that controls switching elements Q1 to Q6 according to the switching pattern and outputs AC power to the rotating mechanical device 3. Learning with teacher data is performed so that the switching loss of unit 12 is smaller than that of the pulse width modulation (PWM) method, a pattern generation function is output, and the pattern determination unit 14 determines the switching pattern of the power conversion unit 12 accordingly. decide.

Therefore, the power conversion device 1 of the first embodiment makes the dq coordinate current value idq (k) follow the current command value idqref (k), and the power conversion unit 12 is more than the pulse width modulation (PWM) method. The rotary mechanical device 3 can be driven so as to reduce the switching loss of the above.

In the above description of the first embodiment, the teacher data prepared in advance is used for creating the trained model by the learning with the teacher data in the machine learning device 10, but the electric motor control B is not limited to this. While doing so, teacher data may be measured and learning with teacher data may be performed. Further, the machine learning device 10 may not be included in the power conversion device 1, and only the pattern generation function may be acquired from the machine learning device 10 by the pattern determination unit 14 of the power conversion device 1.

Embodiment 2.
In the power conversion device of the second embodiment, as the performance of the trained model created in the machine learner, in addition to making the switching loss of the power conversion unit 12 smaller than that of the pulse width modulation (PWM) method, the rotating machine At least one of the drive sound of the device, the mechanical vibration of the rotating machine device, the current harmonic of the rotating machine device, and the follow-up time of the current detection value to the current command value is set for the pulse width modulation (PWM) method. It is something that makes it possible to acquire something that makes it smaller.

Hereinafter, regarding the configuration and operation of the power conversion device and the machine learning device according to the second embodiment, FIG. 6 which is a block diagram showing the configuration of the power conversion device, and the operation of the power conversion device and the machine learning device are shown in a flowchart. This will be described with reference to FIG. In the power conversion device of the second embodiment, the same or corresponding parts as those of the first embodiment are designated by the same reference numerals.

Since the basic functions and configurations of the power conversion device of the second embodiment are the same as those of the first embodiment (FIG. 1) described above, redundant description will be omitted below, and the first and second embodiments will be implemented. The differences between the second and second forms will be described in detail.

As shown in FIG. 6, the entire system of the second embodiment is composed of a power conversion device 1, a DC power supply 2, and a rotary mechanical device 3.

The power conversion device 1 includes a machine learning device 10A, a power conversion unit 12 which is a main circuit, a current detection unit 13, a pattern determination unit 14A, a speed detection unit 15, a position detection unit 16, and a state observation unit 17.

In the first embodiment, the current command value idqref (k), the dq coordinate current value idq (k), the switching state SW (k-1) of the previous cycle, and the pattern generation function are input to the pattern determination unit 14. Was there.

On the other hand, in the second embodiment, the pattern determination unit 14A has a preset control target, a current command value idqref (k), a switching state SW (k-1) of the previous cycle, a pattern generation function, and the like. The state quantity output by the state observation unit 17 to be described in detail later is input.

Here, the control targets include, for example, reduction of switching loss of the power conversion unit 12, reduction of driving noise of the rotary mechanical device 3, reduction of mechanical vibration of the rotary mechanical device 3, and current harmonics of the rotary mechanical device 3. This is a target value for reducing the current detection value and the follow-up time of the current detection value to the current command value. In addition, this control target is, for example, as a table associated with a number, the control target No. In No. 1, the switching loss is reduced, and the control target No. In 2, the switching loss and the mechanical vibration may be reduced in advance.

Further, in the second embodiment, the pattern determination unit 14A outputs a type selection command of the trained model to the machine learning device 10A. This trained model type selection command is a command to select the performance of the trained model that matches the control target and read the pattern generation function from the machine learner 10A, and is used in the subsequent processing of the motor control B. explain. In the drawings of this embodiment (FIGS. 6 and 7), and also in the drawings of other embodiments, the trained model type selection command is represented as TSC for simplification of notation. ..

Next, the functions of the speed detection unit 15, the position detection unit 16, and the state observation unit 17 of the second embodiment, which are different from the first embodiment, will be described.

The speed detection unit 15 detects the mechanical speed information ωrm (k) of the rotary mechanical device 3 and outputs it to the state observation unit 17. As the speed information of the rotary mechanical device 3, the electrical speed information ωre (k) may be detected.

The position detection unit 16 detects the mechanical phase information θrm (k) of the rotary mechanical device 3 and outputs it to the state observation unit 17. As the phase information of the rotating mechanical device 3, the electrical phase information θre (k) may be detected.

The state observing unit 17 observes the driving state of the rotating mechanical device 3 based on the current, speed, and phase of the rotating mechanical device 3 detected by the current detecting unit 13, the speed detecting unit 15, and the position detecting unit 16, respectively. The state quantity is output.

That is, the state observation unit 17 has a current command value idqref (k), a current detection value iuvw (k) acquired from the current detection unit 13, a speed detection value ωrm (k) acquired from the speed detection unit 15, and a position detection unit. The state quantity of the rotating mechanical device 3 is observed based on the position detection value θrm (k) obtained from 16. Here, the state quantities observed by the state observing unit 17 include, for example, the dq coordinate current value idq (k), the rotating machine parameter of the rotating machine device 3, the magnetic flux of the rotating machine device 3, the output torque, and the rotating machine device 3. It includes at least one of the harmonics of the flowing current and the rise time of the dq coordinate current value idq (k) with respect to the current command value idqref (k) to the rotating mechanical device 3.

Note that the rotary machine parameters of the rotary machine device 3 are, for example, values such as resistance, inductance, and moment of inertia of the rotary machine device 3. Each parameter of the rotary mechanical device 3 may be calculated by the state observation unit 17 or may be input to the state observation unit 17.

Next, the processing procedure of the machine learning device 10A and the pattern determination unit 14A, which are the features of the second embodiment, will be described with reference to the flowchart shown in FIG. The processing procedure described below is an example of the learning method and the electric motor control method of the present application. Therefore, each process may be changed as much as possible, and the process can be omitted, replaced, and added as appropriate according to the embodiment.

In the second embodiment as well, as in the first embodiment, the machine learning A performs learning with teacher data based on the teacher data prepared in advance, but the method of creating the teacher data is different.

That is, in the first embodiment, the machine learning device 10 acquires data for executing the motor control B as teacher data so that the switching loss of the power conversion unit 12 is smaller than that of the pulse width modulation (PWM) method. Was there. On the other hand, in the second embodiment, the machine learning device 10A not only reduces the switching loss according to the control target, but also reduces the driving sound of the rotating machine device 3 and the rotating machine device 3 Data is acquired when the motor control B is performed so as to reduce at least one of the mechanical vibration, the current harmonic of the rotating mechanical device 3, and the follow-up time of the current detected value to the current command value.

Therefore, in the first embodiment, only one pattern generation function was created in search of a trained model for reducing the switching loss, but in the second embodiment, the performance is determined according to the control target. Create and save multiple pattern generation functions with different values. The method of saving the pattern generation function created by learning with teacher data for that purpose and the method of acquiring the trained model for executing the motor control B will be described next.

In steps S2 to S4, a trained model is created by learning with teacher data according to the same procedure as in the first embodiment.

In step S10, the trained model created by learning with teacher data is saved as a pattern generation function. After that, the process returns to step S1, the teacher data prepared in advance is changed, and learning with teacher data in steps S2 to S4 is performed again to reduce the switching loss and reduce the driving sound of the rotating mechanical device 3. Learned to have at least one of the following performances: reduction of mechanical vibration of rotary machinery 3, reduction of current harmonics of rotary machinery 3, and reduction of follow-up time of current detection value to current command value. Create a model and save it as a separate pattern generation function.

In learning with teacher data, a plurality of trained models may be individually created for each teacher data prepared in advance, and all pattern generation functions corresponding to the trained models may be saved. Alternatively, select only the necessary teacher data from all the teacher data prepared in advance, create multiple trained models based on the selected teacher data, and save only the pattern generation function corresponding to the trained model. You may try to do it.

By performing the above-mentioned learning with teacher data, a pattern generation function for each teacher data prepared in advance can be acquired and saved. In that case, according to the control target, as a pattern generation function, in addition to making the switching loss of the power conversion unit 12 smaller than that of the pulse width modulation (PWM) method, the drive sound of the rotary machine device 3 and the rotary machine device 3 Switching pattern SWP of the power conversion unit 12 so as to realize the performance of reducing at least one of the mechanical vibration, the current harmonic of the rotating mechanical device 3, and the follow-up time of the current detected value to the current command value. You can get a function that determines k).

Next, the electric motor control B performed by using the trained model after the machine learning A is performed as described above will be described with reference to the flowchart showing the processing procedure of FIG. 7. The processing procedure described below is an example of the electric motor control method of the present application. Therefore, each process described below may be changed as much as possible, and the processes can be omitted, replaced, and added as appropriate according to the embodiment.

First, in step S1, it is determined whether to execute machine learning A or motor control B. Regarding the processing when the machine learning A is performed, the difference from the first embodiment in the second embodiment has already been described. Therefore, here, the embodiment when the processing of the electric motor control B is performed (step S1: No). The difference from 1 will be described.

In step S11, the pattern determination unit 14A acquires the control target. As described above, the control targets are, for example, reduction of switching loss of the power conversion unit 12, reduction of driving noise of the rotary mechanical device 3, reduction of mechanical vibration of the rotary mechanical device 3, and current harmonics of the rotary mechanical device 3. , Reduction of current detection value tracking time to current command value, etc.

In step S12, the pattern determination unit 14A generates a trained model type selection command TSC according to the acquired control target, and outputs this to the machine learner 10A. The trained model type selection command TSC in this case is a command for selecting the trained model stored as a table associated with the numbers in the pattern generation function storage unit 10d. For example, the trained model No. In No. 1, a trained model that reduces switching loss, a trained model No. 1. In No. 2, it is a signal for reading a trained model that matches the control target, such as a trained model that reduces switching loss and mechanical vibration.

Then, the machine learner 10A outputs a pattern generation function conforming to the type selection command TSC of the trained model from the trained models saved in step S10, so that the pattern determination unit 14A acquires this pattern generation function. To do.

Subsequently, in step S7, the pattern determination unit 14A uses the current command value idqref (k) used in the learning with teacher data, the state quantity from the state observation unit 17, and the switching state SW of the previous cycle from the power conversion unit 12 ( k-1) is acquired as input data. In this case, as the value to be acquired as input data, all the input data used for each training with teacher data performed to acquire a plurality of trained models may be input, or a specific trained model may be input. It is also possible to input only the input data used for the learning with the teacher data performed in order to acquire.
After that, in steps S8 to S9, the rotary mechanical device 3 is driven in the same procedure as in the first embodiment.

By performing the above-mentioned electric motor control, the power conversion device 1 of the second embodiment reads a pattern generation function, which is a trained model conforming to the control target, from the machine learning device 10A, and reads the switching pattern SWP of the power conversion unit 12. (K) is determined. Therefore, not only the switching loss of the power conversion unit 12 is made smaller than that of the pulse width modulation (PWM) method, but also the driving sound of the rotary machine device 3, the mechanical vibration of the rotary machine device 3, and the rotary machine device 3 It is possible to drive the rotating mechanical device 3 so as to reduce at least one of the current harmonics and the follow-up time of the current detected value to the current command value.

In the second embodiment, a method of creating and using a plurality of trained models has been described on the basis of making the switching loss smaller than that of the pulse width modulation (PWM) method, but the present invention is not limited to this. , A trained model may be created based on the performance related to other modulation methods.

Embodiment 3.
The power conversion device of the third embodiment performs reinforcement learning on a pattern generation function that has been trained with teacher data used in the pattern determination unit.

Hereinafter, regarding the configuration and operation of the power conversion device and the machine learning device according to the third embodiment, FIG. 8 is a block diagram showing the configuration of the power conversion device, and FIG. 8 is a flowchart showing the operation of the power conversion device and the machine learning device. This will be described based on 9. In the power conversion device of the third embodiment, the same or corresponding parts as those of the first and second embodiments are designated by the same reference numerals.

Since the basic functions and configurations of the power conversion device of the third embodiment are the same as those of the first embodiment (FIG. 1) described above, duplicated description will be omitted below, and the first embodiment and the first embodiment will be omitted here. The differences in Form 3 of the above will be described in detail below.

As shown in FIG. 8, the entire system of the third embodiment is composed of a power conversion device 1, a DC power supply 2, and a rotary mechanical device 3.

The power conversion device 1 includes a machine learning device 10B, a power conversion unit 12, a current detection unit 13, a state observation unit 17B, and a pattern determination unit 14B, which are main circuits.

In the third embodiment, reinforcement learning of the pattern generation function is performed while executing motor control based on the pattern generation function which is a trained model in which learning with teacher data is performed. Therefore, the machine learning device 10B is provided with a reward calculation unit 10g and a function update unit 10h.

Here, first, the concept of reinforcement learning will be explained. Reinforcement learning is learning agents to maximize value in a given environment. That is, in the third embodiment, the state of the rotating mechanical device 3 (given) so as to make the switching loss of the power conversion unit 12 smaller (maximize the value) than the trained model created in the first embodiment. This means that a trained model (agent) that appropriately selects the switching pattern SWP (k) of the power conversion unit 12 is generated according to the environment).

FIG. 9 is a flowchart for explaining an example of a processing procedure for performing reinforcement learning of the learned pattern generation function that has been trained with the teacher data described in FIG. The processing procedure described below is an example of reinforcement learning of the present application. Therefore, each process of these procedures may be changed as much as possible, and the processes can be omitted, replaced, and added as appropriate according to the embodiment.

First, in step S14, the pattern determination unit 14B sets the current command value idqref (k), the dq coordinate current value idq (k), the switching state SW (k-1) of the previous period, and the switching state SW (k-1) of the previous period as the initial state of the rotary mechanical device 3. Acquire the pattern generation function acquired by learning by learning with teacher data. It should be noted that these values acquired as the initial state may all start from the state of 0, or may start using the value in the middle of control.

In step S15, the pattern determination unit 14B determines the switching pattern SWP (k) based on the initial state of the rotary machine device 3 acquired in step S14 and the current pattern generation function obtained by the machine learning device 10B.

Next, in step S16, the power conversion unit 12 drives the rotary mechanical device 3 based on the switching pattern SWP (k) output by the pattern determination unit 14B. Then, the machine learning device 10B has a current command value idqref (k), a dq coordinate current value idq (k) given by the state observation unit 17B, a previous switching pattern SWP (k-1), and a current switching pattern SWP ( The switching transition number SWcount indicating the deviation of the number of times of on / off of the switching elements Q1 to Q6 of the power conversion unit 12 with k) is acquired.

In step S17, the reward calculation unit 10g calculates the current deviation between the current command value idqref (k) and the dq coordinate current value idq (k), and determines whether or not the current deviation is within the specified value. When it is determined that the current deviation is within the specified value (step S17: Yes), the process proceeds to step S18, the preset reward (change amount Δ1) is increased, and it is determined that the current deviation exceeds the specified value. In (step S17: No), the process proceeds to step S19 to reduce the preset reward (change amount Δ1).

In step S20, the reward calculation unit 10g determines whether or not the switching transition number SWcount obtained from the state observation unit 17B is within the specified value. When the switching transition count SWcount is within the specified value (step S20: Yes), the process proceeds to step S21 to increase the preset reward (change amount Δ2), and it is determined that the switching transition count SWcount exceeds the specified value. In (step S20: No), the process proceeds to step S22 to reduce the preset reward (change amount Δ2).

In step S23, the function update unit 10h sets the current deviations of each weighting coefficient and bias of the neural network constituting the pattern generation function based on the rewards (change amounts Δ1, Δ2) obtained by the reward calculation unit 10g. Update the value function to adjust to reduce switching loss while keeping it within the specified range. Then, the pattern generation function is updated based on the updated value function.

Here, the update of the pattern generation function is to adjust each weighting coefficient and bias of the neural network constituting the pattern generation function.
After that, the process returns to step S15, the switching pattern SWP (k) is determined based on the updated pattern generation function, and the same process is repeated.

By performing the above-mentioned reinforcement learning, the power conversion device 1 of the third embodiment has a switching loss smaller than that of the pattern generation function related to the pulse width modulation (PWM) method created in the first embodiment. The pattern generation function is updated. Therefore, the rotary mechanical device 3 can be driven with the switching loss of the power conversion unit 12 smaller than that of the pattern generation function learned by the learning with the teacher data of the first embodiment.

In the third embodiment, the method of reinforcement learning of the trained model trained by the learning with the teacher data of the first embodiment has been described, but the trained model of the second embodiment is to be reinforcement-learned. You may.

That is, in addition to making the switching loss of the power conversion unit 12 smaller than that of the pulse width modulation (PWM) method, the driving sound of the rotary machine device 3, the mechanical vibration of the rotary machine device 3, and the current harmonic of the rotary machine device 3 , And a pattern generation function that can reduce at least one of the follow-up time of the current detection value to the current command value is strengthened and learned, and a pattern generation function with further improved performance is created. May be good.
Further, the pattern generation function obtained by reinforcement learning may be used as the trained model in the machine learning devices 10 and 10A of the first embodiment or the second embodiment.

As described above, the power conversion device 1 of the third embodiment causes switching loss by reinforcement learning of the trained model that has been trained with teacher data in the first embodiment or the second embodiment. In addition, at least one of the driving sound of the rotating mechanical device 3 related to the modulation method, the mechanical vibration of the rotating mechanical device 3, the current harmonic of the rotating mechanical device 3, and the follow-up time of the current detected value to the current command value. The rotary mechanical device 3 can be driven so as to further reduce the size.

Embodiment 4.
The power conversion device of the fourth embodiment has a configuration in which the switching state of the power conversion unit 12 provided in the power conversion device 1 of the first embodiment is calculated based on the voltage command value. By calculating the switching state of the power conversion unit 12 based on the voltage command value, the time constant of the dq coordinate current value idq (k) with respect to the current command value idqref (k) can be set as designed, and further from the speed command of the electric motor. It can also be applied to a control method that calculates the voltage command value and does not go through the current command value.

Hereinafter, the configuration of the power conversion device and the machine learning device of the fourth embodiment will be described with reference to FIG. 10 which is a block diagram showing the configuration of the power conversion device and FIG. 11 which is a block diagram showing the configuration of the machine learning device. ..
The hardware configuration diagram for realizing the power conversion device is shown in FIG. 2, the operation flowchart of the power conversion device and the machine learning device is shown in FIG. 3, and the hardware configuration diagram for realizing the machine learning device is shown in FIG. In common with. Therefore, in the following, the description overlapping with the first embodiment will be omitted, and the points different from the first embodiment will be described in detail. Further, in the power conversion device of the fourth embodiment, the same or corresponding parts as those of the first embodiment are designated by the same reference numerals.

As shown in FIG. 10, the entire system of the fourth embodiment is composed of a power conversion device 1, a DC power supply 2, and a rotary mechanical device 3.

The power conversion device 1 includes a machine learning device 610, a uvw / dq converter 11, a main circuit power conversion unit 12, a current detection unit 13, a pattern determination unit 614, and a PI (Proportional International) current controller 618. Compared with the first embodiment, the power conversion device 1 is further provided with the PI current controller 618.

In FIG. 10, in order to calculate the voltage command value vdqref (k), the PI current controller 618 calculates from the current command value idqref (k) and the dq coordinate current value idq (k). However, instead of the PI current controller 618, a P (Proportional) current controller, an I (Integral) current controller, a PID (Proportional Integral Differential) current controller, and an IP (Integral-Proportional) current controller are used. Alternatively, the voltage command value may be calculated from the speed command value by a control method such as V / f control. Further, in FIG. 10, although the voltage command value vdqref (k) in the dq coordinate is used as the voltage command value, the voltage command value in the αβ coordinate and the voltage command value in the uvw coordinate may be used.

Next, the configuration of the machine learning device 610 used in the fourth embodiment will be described.
As shown in FIG. 11, the machine learning device 610 includes an input data acquisition unit 610a, a label acquisition unit 10b, a learning unit 10c, and a pattern generation function storage unit 10d. Since the data handled by the input data acquisition unit 610a is changed in the fourth embodiment as compared with the first embodiment, the changed contents will be mainly described.

The machine learning device 610 performs learning with teacher data based on the teacher data prepared in advance. Since the method of learning with teacher data is the same as that of the first embodiment, the description thereof is omitted here.

The input data included in the teacher data of the machine learning device 610 is the voltage command value vdqref (k) and the switching state SW (k-1) of the previous cycle of the power conversion unit 12.

As shown in FIG. 11, by performing learning with teacher data in the learning unit 10c based on the teacher data prepared in advance, switching from the voltage command value vdqref (k) and the switching state SW (k-1) of the previous cycle. A trained model that determines the pattern SWP (k) can be created.

The operations of machine learning and electric motor control in the fourth embodiment are the same as those in the flowchart of FIG. 3, and the difference from the first embodiment is that the voltage command value vdqref (k) and the switching state of the previous cycle are different in step S2. The SW (k-1) is acquired as input data, and the voltage command value vdqref (k) and the switching state SW (k-1) of the previous cycle are acquired as input data in step S7.

In the fourth embodiment, the switching pattern SWP (k) is calculated based on the voltage command value vdqref (k), the switching state SW (k-1) of the previous cycle, and the pattern generation function which is a learned model. The PI current controller 618 can design the time constant of the dq coordinate current value idq (k) of the electric motor with respect to the current command value idqref (k). Further, the configuration can be applied to a control method that does not use a current command value for calculating a voltage command value from a speed command of an electric motor.

As described above, the power conversion device 1 of the fourth embodiment has a current detection unit 13 that detects the current flowing through the rotary machine device 3 and a machine learner 610 that outputs a pattern generation function for determining a switching pattern. The PI current controller 618 that calculates the voltage command value vdqref (k) from the current command value idqref (k) and the dq coordinate current value idq (k), the voltage command value vdqref (k), and the previous time of the power conversion unit 12. The pattern determination unit 614, which determines the switching pattern based on the periodic switching state SW (k-1) and the pattern generation function from the machine learner 610, and the switching elements Q1 to Q6 are controlled and rotated according to the switching pattern. The mechanical device 3 is provided with a power conversion unit 12 that outputs AC power, and the machine learner 610 performs learning with teacher data so that the switching loss of the power conversion unit 12 is smaller than that of the pulse width modulation (PWM) method. The pattern generation function is output, and the pattern determination unit 614 determines the switching state of the power conversion unit 12 accordingly.

Therefore, the power conversion device 1 of the fourth embodiment has the same effect as that of the first embodiment, and the dq coordinates of the rotating mechanical device 3 with respect to the current command value idqref (k) as compared with the first embodiment. The PI current controller 618 can design the time constant of the current value idq (k). Further, the configuration can be applied to a control method that does not use a current command value for calculating a voltage command value from a speed command of an electric motor.

In the above description of the fourth embodiment, the teacher data prepared in advance is used for creating the trained model by the learning with the teacher data in the machine learning device 610, but the present invention is not limited to this, and the electric motor control B is used. While doing so, the teacher data may be measured to perform learning with teacher data, or the machine learning device 610 is not included in the power conversion device 1 and only the pattern generation function is included in the pattern determination unit 614 of the power conversion device 1. May be obtained from the machine learning device 610.

Embodiment 5.
The power conversion device of the fifth embodiment has a configuration in which the switching pattern of the power conversion unit 12 provided in the power conversion device 1 of the second embodiment is calculated based on the voltage command value. By calculating the switching pattern of the power conversion unit 12 based on the voltage command value, it can be applied to a control method that calculates the voltage command value from the speed command of the electric motor without using the current command value. Further, even in such a control method, the switching loss of the power conversion unit 12 is reduced, the driving sound of the rotating mechanical device 3 is reduced, the mechanical vibration of the rotating mechanical device 3 is reduced, and the current harmonic of the rotating mechanical device 3 is reduced. The effect of reducing the current detection value tracking time to the current command value can be obtained.

Hereinafter, the configuration of the power conversion device according to the fifth embodiment will be described with reference to FIG. 12, which is a block showing the configuration of the power conversion device. The configuration of the machine learning device is shown in FIG. 11, and the operation flowchart of the power conversion device and the machine learning device is shown in FIG. 7, which is common to the second embodiment or the fourth embodiment. Therefore, in the following, the duplicated description will be omitted, and the points different from those of the second embodiment or the fourth embodiment will be described in detail. In the power conversion device of the fifth embodiment, the same or corresponding parts as those of the second embodiment or the fourth embodiment are designated by the same reference numerals.

As shown in FIG. 12, the entire system of the fifth embodiment is composed of a power conversion device 1, a DC power supply 2, and a rotary mechanical device 3.

The power conversion device 1 of the fifth embodiment includes a machine learning device 710, a main circuit power conversion unit 12, a current detection unit 13, a pattern determination unit 714, a speed detection unit 15, a position detection unit 16, and a state observation unit 17. , And a V / f controller 719. Compared with the second embodiment, the configuration further includes a V / f controller 719.

In FIG. 12, the Vf voltage command value vfref (k) is calculated from the speed command value wref (k) of the electric motor by the V / f controller 719. However, instead of V / f controller 719, from the current command value idqref (k) and dq coordinate current value idq (k), PI current controller 618, P current controller, I current controller, PID current controller, I The voltage command value vdqref (k) may be calculated by the −P current controller, or the voltage command value may be changed to a voltage command value in αβ coordinates or a voltage command value in uvw coordinates.

The machine learning device 710 used in the fifth embodiment has the same configuration as that of FIG. 11 of the fourth embodiment, but controls the input data and the label data included in the teacher data in the same manner as in the second embodiment. The learning department individually creates a plurality of trained models for each teacher data prepared in advance according to the goal.

The machine learning and electric motor control in the fifth embodiment are the same as those in FIG. 7 of the operation flowchart of the power converter and the machine learning device. The difference from the second embodiment is that at least the Vf voltage command value vfr (k) and the switching state SW (k-1) of the previous cycle are acquired as input data in step S2, and in step S7, in step S12. At least the Vf voltage command value vref (k) and the switching state SW (k-1) of the previous cycle are acquired as input data corresponding to the acquired trained model.

In the fifth embodiment, the Vf voltage command value vfr (k) and the pattern generation function matching the control target are read from the machine learning device 710 to determine the switching pattern SWP (k) of the power conversion unit 12. Therefore, even in a control method such as the V / f controller 719 that does not use the current command value for calculating the voltage command value from the speed command of the electric motor, the switching loss of the power conversion unit 12 is reduced as compared with the pulse width modulation (PWM) method. In addition to making it smaller, of the drive sound of the rotary machine device 3, the mechanical vibration of the rotary machine device 3, the current harmonic of the rotary machine device 3, and the follow-up time of the current detection value to the current command value. The effect of driving the rotary mechanical device 3 so as to reduce at least one of them can be obtained.

In the description of the fifth embodiment, the switching pattern SWP (k) is calculated based on the Vf voltage command value Vfref (k) calculated by the V / f controller 719. However, the current command value is described. The voltage command value vdqref (k) is calculated by the PI current controller 618 from the idqref (k) and the dq coordinate current value idq (k), and the switching pattern SWP (k) is calculated from the voltage command value vdqref (k). It may be.

Embodiment 6.
The power conversion device of the sixth embodiment has a configuration in which the switching pattern of the power conversion unit 12 provided in the power conversion device 1 of the third embodiment is calculated based on the voltage command value. By calculating the switching pattern of the power conversion unit 12 based on the voltage command value, it is used in the pattern determination unit while keeping the time constant of the dq coordinate current value idq (k) with respect to the current command value idqref (k) as designed. Reinforcement learning can be performed on the pattern generation function that has been trained with teacher data. Further, it can be applied to a control method that does not use a current command value for calculating a voltage command value from a speed command of an electric motor.

Hereinafter, the configuration of the power conversion device according to the sixth embodiment will be described with reference to FIG. 13, which is a block showing the configuration of the power conversion device. The operation flowchart of the power conversion device and the machine learning device is shown in FIG. 9, which is common to the third embodiment. Therefore, in the following, the duplicated description will be omitted, and the differences from the third embodiment will be described in detail. In the power conversion device of the sixth embodiment, the same or corresponding parts as those of the third embodiment are designated by the same reference numerals.

As shown in FIG. 13, the entire system of the sixth embodiment is composed of a power conversion device 1, a DC power supply 2, and a rotary mechanical device 3.

The power conversion device 1 includes a machine learning device 810, a power conversion unit 12, a current detection unit 13, a state observation unit 17, a PI current controller 618, and a pattern determination unit 814, which are main circuits. As compared with the third embodiment, the PI current controller 618 is further provided.

In FIG. 13, in order to calculate the voltage command value vdqref (k), the PI current controller 618 calculates from the current command value idqref (k) and the dq coordinate current value idq (k). However, instead of the PI current controller 618, a P current controller, an I current controller, a PID current controller, and an IP current controller may be used, and the speed command is performed by a control method such as V / f control. The voltage command value may be calculated from the value. In the present embodiment, the voltage command value is set to the dq coordinate, but it may be changed to the voltage command value of the αβ coordinate or the voltage command value of the uvw coordinate.

The reinforcement learning and the electric motor control in the sixth embodiment are almost the same as those in FIG. 9 of the operation flowchart of the power converter and the machine learning device, but the difference from the third embodiment is the pattern determination unit 814 in step S15. Is the voltage command value vdqref (k) calculated by the PI current controller 618 of the current command value idqref (k) and the dq coordinate current value idq (k) in the initial state of the rotary machine device 3 acquired in step S14, and the machine. The switching pattern SWP (k) is determined based on the current pattern generation function obtained by the learner 810.

This embodiment 6 updates the trained model pattern generation function so that the switching loss is further smaller than the pattern generation function related to the pulse width modulation (PWM) method created in the fourth embodiment. The rotating mechanical device 3 can be driven with the switching loss of the power conversion unit 12 smaller than that of the learned model obtained by the learning with the teacher data of the fourth embodiment.

In the sixth embodiment, the method of reinforcement learning of the trained model obtained by the learning with the teacher data of the fourth embodiment has been described, but the trained model of the fifth embodiment is reinforcement-learned. May be good.

That is, in addition to making the switching loss of the power conversion unit 12 smaller than that of the pulse width modulation (PWM) method, the driving sound of the rotary machine device 3, the mechanical vibration of the rotary machine device 3, and the current harmonic of the rotary machine device 3 , And a pattern generation function that can reduce at least one of the follow-up time of the current detection value to the current command value is strengthened and learned, and a pattern generation function with further improved performance is created. May be good.
Further, the pattern generation function obtained by reinforcement learning may be used as the trained model in the machine learning device 610 of the fourth embodiment or the machine learning device 710 of the fifth embodiment.

As described above, the power conversion device 1 of the sixth embodiment causes switching loss by reinforcement learning of the trained model that has been trained with teacher data in the fourth embodiment or the fifth embodiment. In addition, at least one of the driving sound of the rotating mechanical device 6 related to the modulation method, the mechanical vibration of the rotating mechanical device 6, the current harmonic of the rotating mechanical device 6, and the follow-up time of the current detected value to the current command value. The rotary mechanical device 3 can be driven so as to further reduce the size. Further, as compared with the third embodiment, the time constant of the dq coordinate current value idq (k) with respect to the current command value idqref (k) is designed by calculating the switching pattern of the power conversion unit 12 based on the voltage command value. Reinforcement learning can be performed on the pattern generation function that has been trained with the teacher data used in the pattern determination unit 814 while keeping the same. Further, it can be applied to a control method that does not use a current command value for calculating a voltage command value from a speed command of an electric motor.

Although the present application describes various exemplary embodiments and examples, the various features, embodiments, and functions described in one or more embodiments are applications of a particular embodiment. It is not limited to, but can be applied to embodiments alone or in various combinations.
Therefore, innumerable variations not illustrated are envisioned within the scope of the techniques disclosed in the present application. For example, it is assumed that at least one component is modified, added or omitted, and further, at least one component is extracted and combined with the components of other embodiments.

1 Power converter, 2 DC power supply, 3 Rotating machine device, 10, 10A, 10B, 610, 710, 810 Machine learning device, 10a, 610a Input data acquisition unit, 10b Label acquisition unit, 10c Learning unit, 10d Pattern generation function Storage unit, 10g reward calculation unit, 10h function update unit, 11 uvw / dq converter, 12 power conversion unit, 12a switching element, 13 current detection unit, 14, 14A, 14B, 614,714,814 pattern determination unit, 20 , 30 processor, 21,31 storage device, 15 speed detection unit, 16 position detection unit, 17,17B state observation unit, 311 volatile storage device, 312 auxiliary storage device, 618 PI current controller, 719 V / f controller ..

Claims (15)

1. A power conversion unit that is equipped with a plurality of switching elements and converts DC power into AC power and supplies it to the rotary machine device, a current detection unit that detects the current flowing through the rotary machine device, and each set control cycle. Based on the pattern generation function given by the machine learner, the switching states for one cycle in the plurality of switching elements are determined, and a switching pattern consisting of a combination of the switching states for one cycle is generated to generate the power conversion. Equipped with a pattern determination unit that outputs and controls the unit
   The machine learning device executes machine learning based on either the current command value, the current detection value detected by the current detection unit, or the voltage command value included in the teacher data, and the power conversion unit of the power conversion unit. It generates the pattern generation function that determines the switching pattern.
   The pattern determination unit inputs both the pattern generation function given by the machine learning device and either the current command value and the current detection value or the voltage command value to execute arithmetic processing. A power conversion device that determines the switching pattern of the power conversion unit.
2. The power conversion device according to claim 1, wherein the pattern determination unit acquires the pattern generation function for reducing the switching loss of the power conversion unit from the machine learning device as compared with the case of the pulse width modulation method.
3. From the machine learner, the pattern determining unit determines the driving sound of the rotating machine device, the mechanical vibration of the rotating machine device, the current harmonics included in the current flowing through the rotating machine device, and the current to the current command value. The power conversion device according to claim 1 or 2, wherein the pattern generation function for making at least one of the tracking times of the detected values smaller than that in the case of the pulse width modulation method is acquired.
4. When determining the switching pattern of the power conversion unit, the pattern determining unit determines the magnetic flux response value, the speed response value, the position response value, the magnetic flux command value, the speed command value, and the position of the rotating mechanical device. Any one of claims 1 to 3, which acquires at least one of a command value, a switching state in the previous control cycle of the power conversion unit, and a rotary machine parameter of the rotary machine device as a state quantity. The power converter according to the section.
5. The power conversion device according to any one of claims 1 to 4, further comprising the machine learning device.
6. A pattern generation function that determines a switching pattern consisting of a combination of switching states for one set control cycle of multiple switching elements that make up the power conversion unit that converts DC power to AC power and supplies it to the rotating machinery. , Executes machine learning based on teacher data and outputs it.
   An input data acquisition unit that acquires either the current command value included in the teacher data, the current detection value detected by the current detection unit that detects the current flowing through the rotating mechanical device, or the voltage command value as input data. ,
   A label acquisition unit that acquires the switching pattern of the power conversion unit included in the teacher data as a label, and
   The switching of the power conversion unit based on either one of the current command value and the current detection value or the voltage command value obtained by the input data acquisition unit and a switching pattern obtained by the label acquisition unit. A machine learning device equipped with a learning unit that generates a trained model that determines the switching pattern of an element.
7. The input data acquisition unit acquires the switching state of the power conversion unit in the previous cycle as the input data included in the teacher data, and at the same time,
   The sixth aspect of claim 6, wherein the label acquisition unit acquires the switching pattern of the power conversion unit of the current cycle corresponding to the switching state of the power conversion unit of the previous cycle as the label included in the teacher data. Machine learning device.
8. The input data acquisition unit, as the input data included in the teacher data, includes a magnetic flux response value, a speed response value, and a position response value of the rotary machine device, a magnetic flux command value, a speed command value, and a position command of the rotary machine device. The machine learning device according to claim 6 or 7, wherein the value and at least one of the rotary machine parameters of the rotary machine device are acquired.
9. Claim 6 or claim 7 that the label acquisition unit acquires the switching pattern for reducing the switching loss of the power conversion unit as the label included in the teacher data as compared with the case of the pulse width modulation method. The machine learning device described in.
10. The label acquisition unit, as the label included in the teacher data, is at least one of a current harmonic included in the current flowing through the rotating machine device and a follow-up time of the current detection value to the current command value. The machine learning device according to any one of claims 6, 7, or 9, which acquires a switching pattern in which one is smaller than that in the case of the pulse width modulation method.
11. The learning unit updates the pattern generation function based on the reward calculation unit that calculates the reward based on the current command value, the current detection value, or the switching state, and the reward input from the reward calculation unit. The machine learning device according to any one of claims 6 to 10, further comprising a function update unit.
12. In the switching pattern of the power conversion unit, the reward calculation unit increases the reward when the number of switching times of the power conversion unit is smaller than the current state, and the current state when the difference between the current detection value and the current command value exceeds the specified value. The machine learning device according to claim 11, which reduces the reward more than.
13. The machine learning device according to claim 12, wherein the reward calculation unit increases the reward when the mechanical vibration of the rotating mechanical device is made smaller than the current state in the switching pattern in the power conversion unit.
14. The machine learning device according to claim 12 or 13, wherein the reward calculation unit increases the reward when the driving sound of the rotating mechanical device is made smaller than the current state in the switching pattern in the power conversion unit.
15. A trained model for determining a switching pattern of the switching element constituting the power conversion unit by performing machine learning using the machine learning device according to any one of claims 6 to 14. How to generate a trained model.

Images