Inference device and inference method

Abstract

An inference device (10) that infers control data for controlling driving electric power supplied to an electric motor (61), comprising: a data acquisition unit (111) that acquires condition data indicating a condition of the electric motor (61); and an inference unit (113) that uses a learned model (20) for inferring the control data on the basis of the condition data to infer the control data on the basis of the condition data acquired by the data acquisition unit (111).

Description

Inference device and inference method

[0001] The present disclosure relates to an inference device and an inference method for inferring control data for controlling drive power supplied to an electric motor.

[0002] BACKGROUND ART Electric motors that are driven by supplied electric power are known. For example, Japanese Patent Application Laid-Open Publication No. 2013-167256 (Patent Document 1) discloses a motor that drives a compressor.

[0003] JP 2013-167256 A

[0004] Electric motors such as the motor disclosed in JP 2013-167256 A (Patent Document 1) are generally configured to be driven using drive power supplied based on control data from a control device. The energy efficiency of the electric motor is determined depending on energy losses such as mechanical loss, iron loss, and copper loss. For this reason, for example, the control device is configured to determine control data that minimizes energy loss using a predetermined calculation formula such as a voltage equation.

[0005] However, because the multiple components that make up an electric motor have individual variations, the performance of an electric motor made up of multiple components varies from motor to motor. As a result, with conventional methods, it has been difficult to calculate optimal control data according to the state of each individual electric motor due to error factors that are not reflected in predetermined calculation formulas such as voltage equations.

[0006] The present disclosure has been made to solve the above-mentioned problems, and aims to provide a technique for inferring optimal control data according to the state of each electric motor.

[0007] An inference device according to the present disclosure infers control data for controlling drive power supplied to an electric motor, and includes a data acquisition unit that acquires status data indicating the status of the electric motor, and an inference unit that infers the control data based on the status data acquired by the data acquisition unit using a trained model for inferring the control data based on the status data.

[0008] An inference method according to the present disclosure is a method for inferring control data for controlling drive power supplied to an electric motor by a computer. The inference method includes, as processing executed by the computer, a step of acquiring status data indicating a status of the electric motor, and a step of inferring the control data based on the status data acquired in the acquiring step, using a trained model for inferring the control data based on the status data.

[0009] According to the present disclosure, control data can be inferred based on status data indicating the status of an electric motor using a trained model, thereby making it possible to infer optimal control data according to the status of each individual electric motor.

[0010] FIG. 1 is a diagram illustrating a configuration of an electric motor according to an embodiment. FIG. 2 is a diagram illustrating a cross section of an electric motor according to an embodiment. FIG. 3 is a diagram illustrating a configuration of an electric motor unit and an inference device according to an embodiment. FIG. 4 is a diagram illustrating an overview of reinforcement learning. FIG. 5 is a diagram illustrating an overview of reinforcement learning. FIG. 6 is a diagram illustrating B1 (action), B2 (state), C (output), and D (reward criterion) in reinforcement learning according to an embodiment. FIG. 7 is a diagram illustrating a configuration of a learning device in a learning phase. FIG. 8 is a flowchart relating to processing executed by a learning device (control unit) in the learning phase. FIG. 9 is a diagram illustrating a configuration of an inference device in an utilization phase. FIG. 10 is a flowchart relating to processing executed by an inference device (control unit) in the utilization phase.

[0011] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. While multiple embodiments will be described below, it is anticipated from the beginning that the configurations described in each embodiment will be appropriately combined. Note that identical or corresponding parts in the drawings will be designated by the same reference numerals, and their description will not be repeated.

[0012] [Configuration of the Electric Motor] An electric motor 61 according to an embodiment will be described with reference to Figures 1 and 2. The electric motor 61 drives various actuators such as a compressor mounted in an air conditioner, for example.

[0013] FIG. 1 is a diagram showing the configuration of an electric motor 61 according to an embodiment. In FIG. 1, when the electric motor 61 is viewed from the side, the direction transverse to the electric motor 61 is defined as the X-axis direction, the direction longitudinal to the electric motor 61 that is perpendicular to the X-axis is defined as the Y-axis direction, and the direction perpendicular to the X-axis and Y-axis is defined as the Z-axis direction. FIG. 1 shows a longitudinal cross section of the electric motor 61 taken along the X-Y plane. FIG. 2 is a diagram showing a transverse cross section of the electric motor 61. FIG. 2 shows a transverse cross section of the electric motor 61 taken along the X-Z plane at line A-A' shown in FIG. 1.

[0014] As shown in FIGS. 1 and 2, the electric motor 61 includes a housing 60, a stator 611, a rotor 612, a shaft 613, and a glass terminal 67 for supplying power to the electric motor 61.

[0015] The housing 60 accommodates components such as a stator 611 and a rotor 612. The stator 611 is formed of an iron core or a coil, and has a circular or nearly circular cross section. The stator 611 is formed by fastening electromagnetic steel plates having a predetermined thickness stacked in the Y-axis direction using caulking or the like. The stator 611 is formed in an annular shape centered in the axial direction, and a central hole 619 having a circular cross section is formed in the center of the stator 611 in order to accommodate the rotor 612. The rotor 612, which rotates around the Y-axis direction, is disposed in the central hole 619 of the stator 611. The rotor 612 is rotatable in the central hole 619 formed in the stator 611 in a direction along the X-Z plane, around the Y-axis direction.

[0016] Furthermore, a plurality of slots 614 are formed in the stator 611 along the circumferential direction. A winding 615 is attached to each of the plurality of slots 614. Driving power for driving the electric motor 61 is supplied to the winding 615 via glass terminals 67. The winding 615 is attached to the stator 611 by a well-known winding method such as a distributed winding method or a concentrated winding method, and the method of attaching the winding 615 is not particularly limited.

[0017] The rotor 612 has a circular or nearly circular cross section. The rotor 612 is formed by laminating electromagnetic steel sheets of a predetermined thickness in the Y-axis direction and fastening them together using caulking, rivets, or the like. The rotor 612 is disposed inside the stator 611 so as to fit into the central hole 619 without contacting the stator 611. A shaft hole 616 having a circular cross section is formed in the center of the rotor 612 along the Y-axis direction, through which the shaft 613 passes. The rotor 612 also has a plurality of air holes 617 formed surrounding the shaft hole 616. A plurality of permanent magnets 618 are provided outside the plurality of air holes 617. In the electric motor 61 shown in FIG. 2 , six permanent magnets 618 are disposed at equal intervals of 60 degrees around the circumferential direction of the rotor 612.

[0018] The electric motor 61 is not limited to an interior permanent magnet (IPM) motor in which the permanent magnet 618 is embedded inside the rotor 612, but may be a surface permanent magnet (SPM) motor in which the permanent magnet 618 is attached to the outer surface of the rotor 612, or a motor having another configuration.

[0019] [Motor Operation] In the electric motor 61 configured as described above, when a current flows through the windings 615 of the stator 611 due to the driving power supplied from the glass terminals 67, a rotating magnetic field is generated in the stator 611. The permanent magnets 618 are attracted to the rotating magnetic field generated in the stator 611, causing the rotor 612 to rotate. As the rotor 612 rotates, the shaft 613 inserted in the rotor 612 rotates. The number of rotations per unit time, i.e., the rotational speed, and the output torque of the rotor 612 change depending on the driving power (driving current) supplied to the electric motor 61.

[0020] As a result, the electric motor 61 can transmit the rotational force of the shaft 613 to an actuator such as a compression mechanism connected to the shaft 613, thereby driving the actuator.

[0021] [Configuration of the Electric Motor Unit] An electric motor unit 600 according to an embodiment will be described with reference to Fig. 3. Fig. 3 is a diagram showing the configuration of the electric motor unit 600 and the inference device 10 according to an embodiment. As shown in Fig. 3, the electric motor unit 600 includes an electric motor 61, a drive circuit 601, a power supply circuit 602, and a control device 100.

[0022] The drive circuit 601 supplies drive power to the glass terminal 67 of the electric motor 61 using power source power supplied from the power source circuit 602 based on control data (control command values) from the control device 100. The power source circuit 602 transforms power supplied from a battery (not shown) and supplies the power source power to the drive circuit 601.

[0023] The control device 100 is a computing entity that controls the drive power supplied from the drive circuit 601 to the electric motor 61 by executing various programs. The control device 100 is composed of a computer such as a processor. The processor may be, for example, a microcontroller, a central processing unit (CPU), or a microprocessing unit (MPU). The processor has the function of executing various processes by executing programs, but some or all of these functions may be implemented using dedicated hardware circuits such as an application-specific integrated circuit (ASIC), a graphics processing unit (GPU), or a field-programmable gate array (FPGA). The term "processor" is not limited to a processor in the narrow sense that executes processes using a stored program, such as a CPU or MPU, but may also include hardwired circuits such as an ASIC, GPU, or FPGA. Therefore, the term "processor" can also be interpreted as a processing circuit whose processes are predefined by computer-readable code and / or hardwired circuits. The processor may be composed of a single chip or multiple chips. Furthermore, the processor and associated processing circuitry may be comprised of multiple computers interconnected by wire or wirelessly, such as via a local area network or a wireless network. The processor and associated processing circuitry may also be comprised of a cloud computer that performs remote calculations based on input data and outputs the results of the calculations to other devices at remote locations.

[0024] Furthermore, the control device 100 may include a storage unit for storing program code, work memory, etc., when the processor executes various programs. The storage unit may be one or more non-transitory computer-readable media. Examples of the storage unit include volatile memory such as dynamic random access memory (DRAM) and static random access memory (SRAM), or non-volatile memory such as read-only memory (ROM) and flash memory. The storage unit may also be one or more computer-readable storage media. Examples of the storage unit include storage devices such as hard disk drives (HDDs) and solid-state drives (SSDs).

[0025] [Individual Variations of Electric Motors] The performance of the electric motor 61 can be affected by individual variations of the electric motor 61. Specifically, the individual variations of the electric motor 61 can be affected by individual variations of the components that make up the electric motor 61. For example, there are individual variations in the shape of the housing 60, the inner diameter, outer diameter, and width dimensions of the stator 611, the inner diameter and outer diameter dimensions of the rotor 612, and the shape and inclination of the shaft 613. Furthermore, the dimensions of the gaps that occur when the housing 60, stator 611, rotor 612, and shaft 613 are combined also vary due to individual variations of the components or the state of assembly.

[0026] In particular, the gap between the stator 611 and the rotor 612 is easily affected by the misalignment between the stator 611 and the rotor 612, the inner diameter of the stator 611, the outer diameter of the rotor 612, the shape and inclination of the shaft 613, and the degree to which the stator 611 and the rotor 612 are fixed. For example, if the outer diameter of the rotor 612 is smaller than the standard, the gap between the stator 611 and the rotor 612 will be large, and if the outer diameter of the rotor 612 is larger than the standard, the gap between the stator 611 and the rotor 612 will be small. If the inner diameter of the stator 611 is smaller than the standard, the gap between the stator 611 and the rotor 612 will be small, and if the inner diameter of the stator 611 is larger than the standard, the gap between the stator 611 and the rotor 612 will be large.

[0027] Furthermore, a major factor affecting the individual variations of the electric motor 61 is the amount of magnetic flux in the rotor 612 that is linked to the windings 615. The amount of magnetic flux in the rotor 612 that is linked to the windings 615 generates an induced voltage based on the law of electromagnetic induction. The magnitude of the induced voltage is proportional to the magnitude of the magnetic flux in the rotor 612 that is linked to the windings 615. In other words, the amount of magnetic flux in the rotor 612 that is linked to the windings 615 corresponds to the induced voltage. The amount of magnetic flux in the rotor 612 that is linked to the windings 615 varies mainly depending on the magnetic flux density of the permanent magnets 618 inserted in the rotor 612, the dimensions of the permanent magnets 618, the outer diameter of the rotor 612, and the inner diameter of the stator 611, and variations in these factors also tend to cause fluctuations in the driving power supplied to the electric motor 61. For example, when the amount of magnetic flux in the rotor 612 decreases, the current flowing through the windings 615 of the motor 61 also increases, causing copper loss and increasing the driving power supplied to the motor 61. This causes the input power of the motor 61 to fluctuate.

[0028] [Control of drive power by control device] Here, an example of control of drive power by the control device 100 will be described. The energy efficiency of the electric motor 61 is determined depending on energy losses such as mechanical loss, iron loss, and copper loss. For this reason, the control device 100 is generally configured to determine control data that minimizes energy loss using a predetermined calculation formula (theoretical formula) such as the voltage equation shown in the following formula (1):

[0029]

[0030] In equation (1), T is the output torque of the electric motor 61, p is the number of pole pairs of the permanent magnet 618, Φ ａ is the induced voltage constant, I ａ is the value of the drive current supplied to the motor 61, β is the phase of the drive current supplied to the motor 61, L ｑ is the q-axis inductance, and L ｄ As shown in equation (1), the output torque T of the electric motor 61 is expressed by the value I of the driving current supplied to the electric motor 61. ａ and phase β. Note that the value of the drive current I ａ is an index value indicating the magnitude of the drive current supplied to the motor 61. The phase β of the drive current is a value indicating the phase relationship of the current waveform with respect to the rotation angle of the rotor 612.

[0031] When the output torque T is determined, the control device 100 uses the formula (1) to determine the value I of the drive current for calculating the output torque T. ａ and phase β, where the value of the drive current I ａ When is minimum, copper loss in the winding 615, which is proportional to the square of the product of the winding resistance and the drive current, is also minimum. ａ In this way, the control device 100 is configured to minimize the copper loss by minimizing the drive current value I that is minimum according to a specific output torque T. ａ and phase β are determined, and the determined drive current value I ａ and phase β to the drive circuit 601, thereby controlling the drive power of the electric motor 61. Such a control method of the electric motor 61 by the control device 100 is also called "maximum torque / current control." The control device 100 can improve the energy efficiency of the electric motor 61 to some extent by minimizing copper loss.

[0032] However, in addition to copper loss, the electric motor 61 also generates mechanical loss and iron loss. Therefore, even if copper loss is minimized, the control device 100 may not be able to truly maximize energy efficiency. In particular, iron loss is a function of the output torque T, the number of pole pairs p, and the induced voltage constant Φ used in equation (1). ａ , the value of the driving current I ａ , the phase β of the driving current, and the q-axis inductance L ｑ , and the d-axis inductance L ｄ Iron loss may vary nonlinearly with factors such as magnetic saturation, stress distribution due to crimping, shrink fitting, punching of the electromagnetic steel sheet, and modulation factor. Furthermore, magnetic saturation, stress distribution due to crimping, shrink fitting, punching of the electromagnetic steel sheet, and modulation factor are all significantly affected by individual variations in each electric motor 61, such as dimensional variations in the components that make up the electric motor 61. Therefore, it is difficult for the control device 100 to theoretically predict iron loss, as is the case with copper loss, and it is difficult to truly maximize energy efficiency while taking iron loss into consideration.

[0033] Furthermore, even for copper loss, which is relatively easy to predict based on formula (1), there are error factors that are not reflected in the theoretical formula, such as pulsation in the operating load of the electric motor 61 that is unexpected by the control device 100, deviations in the observed values ​​observed by the control device 100, and deviations in the calculated values ​​used for various controls that accompany deviations in the observed values. For this reason, even when control is performed based on formula (1), the control device 100 may not be able to minimize copper loss, and it is difficult to calculate optimal control data that truly maximizes energy efficiency in accordance with the state of each electric motor 61.

[0034] Furthermore, the motor 61 will not operate unless the sum of the induced voltage caused by the amount of magnetic flux of the rotor 612 interlinked with the windings 615 and the voltage consumption due to the drive power supplied to the motor 61 is lower than the bus voltage of the drive power supplied from the drive circuit 601 to the motor 61. Specifically, the control device 100 needs to determine the control data so as to satisfy a predetermined calculation formula (theoretical formula) shown in the following formula (2).

[0035]

[0036] In equation (2), R is the winding resistance of the winding 615, I ａ is the value of the drive current supplied to the motor 61, β is the phase of the drive current supplied to the motor 61, ω is the rotation speed of the rotor 612, L ｑ is the q-axis inductance, Φ ａ is the induced voltage constant, L ｄ is the d-axis inductance, and V ｏｍ and ωΦ respectively indicate the bus voltage. ａ corresponds to the induced voltage.

[0037] Because the induced voltage is generated in proportion to the rotation speed of the rotor 612, the electric motor 61 incorporating the permanent magnet 618 is likely to become unable to drive, particularly in the high rotation speed range. Generally, the control device 100 controls the drive power so that the winding field magnetic flux is generated in a direction that weakens the magnetic flux of the permanent magnet 618, thereby enabling the electric motor 61 to be driven even in the high rotation speed range. This control method of the electric motor 61 by the control device 100 is also called "field-weakening control." The control device 100 can achieve field-weakening control by adjusting the phase β of the drive current.

[0038] For example, in equation (2), when the phase β of the drive current is gradually increased, the control device 100 ｄ I ａ Since the value of "sinβ" can be increased, ａ +RI ａ cosβ-ωL ｄ I ａ By performing such field weakening control, the control device 100 can reduce the induced voltage ωΦ caused by the magnetic flux of the magnet. ａ As a result, the electric motor 61 can be driven even in a high rotation speed range.

[0039] However, when the control device 100 controls the electric motor 61 with field-weakening control, the phase β of the drive current is greatly varied in accordance with the operating load, and therefore it is more difficult to calculate control data that truly maximizes energy efficiency than when the electric motor 61 is not controlled with field-weakening control. If the control device 100 does not control the electric motor 61 with field-weakening control, the control device 100 calculates the value I of the drive current that satisfies the output torque T based on the formula (1). ａ and phase β, and from among them, the value of the drive current I ａ and the phase β. However, in reality, the control device 100 needs to control the electric motor 61 by field-weakening control so as to satisfy not only the formula (1) but also the formula (2), and therefore it is even more difficult to calculate optimal control data that truly maximizes energy efficiency.

[0040] Thus, by simply calculating the control data based on the theoretical formula (1), the control device 100 cannot calculate optimal control data that truly minimizes energy loss while taking into account each energy loss, such as copper loss, mechanical loss, and iron loss. As a result, there is a risk that the energy efficiency of the electric motor 61 will decrease relatively. Furthermore, the control device 100 needs to calculate optimal control data even when field-weakening control is performed to satisfy the constraints shown in formula (2).

[0041] Furthermore, the control device 100 does not include a sensor that detects the rotation angle of the rotor 612 in the electric motor 61, and is configured to estimate the rotation angle of the rotor 612 by sensorless control based on a predetermined calculation formula (theoretical formula) not shown. The electric motor 61 rotates the rotor 612 at a rotation angle calculated by the control device 100 using the predetermined calculation formula. Examples of the predetermined calculation formula (theoretical formula) for estimating the rotation angle of the rotor 612 include an induced voltage utilization method that estimates the rotation angle based on an observed value of an induced voltage, and a saliency utilization method that estimates the rotation angle based on data on the magnetic saliency of the electric motor 61.

[0042] When the rotation angle of the rotor 612 is estimated by sensorless control based on a predetermined calculation formula (theoretical formula), errors are likely to occur in the estimation of the rotation angle, making it difficult for the control device 100 to accurately calculate the rotation angle. For example, in the case of the induced voltage utilization method, it is known that the accuracy of the estimation of the rotation angle deteriorates in the low rotation speed range where induced voltage is hardly generated. Also, in the case of the saliency utilization method, it is known that the accuracy of the estimation of the rotation angle deteriorates in the medium rotation speed range or higher because it becomes difficult to observe saliency.

[0043] Therefore, when the control device 100 estimates the rotation angle of the rotor 612 based on a calculation formula such as an induced voltage method or a saliency method, it is difficult to determine control data that minimizes energy loss due to the influence of estimation errors. For example, the current phase β represents the phase relationship of the current waveform with respect to the rotation angle of the rotor 612. Therefore, the control device 100 calculates the value I of the drive current based on formulas (1) and (2). ａ When determining the phase β, if an estimation error occurs in the rotation angle of the rotor 612, the electric motor 61 will be energized at a phase different from the originally intended current waveform, and there is a risk that the electric motor 61 will not be able to be controlled appropriately.

[0044] As described above, it is difficult for the control device 100 to infer optimal control data that truly maximizes energy efficiency, considering factors such as variations among individual electric motors 61, the need to satisfy the constraints shown in formula (2), and the possibility of estimation errors in the rotation angle of the rotor 612. Therefore, the present disclosure provides an inference device 10 that uses AI (artificial intelligence) to infer optimal control data that truly maximizes the energy efficiency of the electric motor unit 600, while considering factors such as variations among individual electric motors 61, the need to satisfy the constraints shown in formula (2), and the possibility of estimation errors in the rotation angle of the rotor 612.

[0045] [Inference Device] An inference device 10 according to an embodiment will be described with reference to Fig. 3. As shown in Fig. 3, the inference device 10 is communicably connected to the control device 100 of the electric motor unit 600, and together with the electric motor unit 600, constitutes an inference system 1000. The inference device 10 includes a control unit 11, a storage unit 12, and an input / output unit 13 as its main functional components.

[0046] The control unit 11 is a computing entity that executes various programs to infer control data for controlling the drive power. The control unit 11 is composed of a computer such as a processor. The processor may be, for example, a microcontroller, a CPU, or an MPU. The processor has the function of executing various processes by executing programs, but some or all of these functions may be implemented using dedicated hardware circuits such as an ASIC, a GPU, or an FPGA. The term "processor" is not limited to a processor in the narrow sense that executes processes using stored programs, such as a CPU or an MPU, but may also include hardwired circuits such as an ASIC, a GPU, or an FPGA. Therefore, the term "processor" can also be interpreted as a processing circuitry in which processes are predefined by computer-readable code and / or hardwired circuits. The processor may be composed of a single chip or multiple chips. Furthermore, the processor and related processing circuits may be composed of multiple computers interconnected by wire or wirelessly via a local area network or a wireless network. The processor and associated processing circuitry may be configured as a cloud computer that performs computations remotely based on input data and outputs the results of the computations to other devices at remote locations.

[0047] The memory unit 12 provides a storage area for storing program code, work memory, etc. when the control unit 11 executes various programs. The memory unit 12 may be one or more non-transitory computer readable media. Examples of the memory unit 12 include volatile memories such as DRAM and SRAM, and non-volatile memories such as ROM and flash memory. The memory unit 12 may also be one or more computer readable storage media. Examples of the memory unit 12 include storage devices such as HDDs and SSDs.

[0048] The input / output unit 13 is an output interface that outputs the control data inferred by the control unit 11 to the control device 100 of the electric motor unit 600. The input / output unit 13 is also an input interface that acquires, from the control device 100 of the electric motor unit 600, status data that indicates the status of the electric motor 61, which is necessary for the control unit 11 to infer the control data.

[0049] The control unit 11 includes a data acquisition unit 111 , a model generation unit 112 , an inference unit 113 , and a presentation unit 114 .

[0050] The data acquisition unit 111 acquires status data indicating the status of the electric motor 61 from the control device 100 via the input / output unit 13. The model generation unit 112 generates a trained model 20 (described later) for inferring the control data based on the status data, using learning data 30 (described later) including the control data and status data acquired from the electric motor 61 driven based on the control data. The inference unit 113 uses the trained model 20 to infer the control data based on the status data.

[0051] The presentation unit 114 presents to the user information relating to the driving of the electric motor 61 based on the control data inferred by the inference unit 113. For example, the presentation unit 114 outputs presentation data to the display device 500 via the input / output unit 13 to present that the electric motor 61 is being controlled based on the control data inferred by the inference device 10 or to present the effect of controlling the electric motor 61 based on the control data inferred by the inference device 10. The display device 500 may display the effect obtained by using the inference device 10, for example, by displaying an estimated value of energy loss that could be reduced by using the inference device 10, as the effect of controlling the electric motor 61, based on the presentation data.

[0052] In this way, the inference device 10 can encourage users to use the inference device 10 by informing the user that it is controlling the electric motor 61 based on the control data inferred by the inference device 10 or by informing the user of the effect of controlling the electric motor 61 based on the control data inferred by the inference device 10. The process by which the inference device 10 determines the control data for the electric motor 61 tends to be a black box, but by notifying the user as described above, the user can determine the effect of the inference device 10 based on the operating results of the electric motor 61, and it is possible to ensure that the inference device 10 is being operated appropriately.

[0053] The display device 500 is configured to be able to communicate with the inference device 10 via a network. The display device 500 is an information terminal used by a user of an air conditioner equipped with an electric motor 61. The display device 500 may be implemented as a general-purpose computer, or as a dedicated computer for controlling the air conditioner equipped with the electric motor 61. For example, the display device 500 may be an information terminal that executes predetermined information processing, such as a desktop personal computer (PC), a laptop PC, a smartphone, a smartwatch, a wearable device, a tablet PC, or a remote controller installed in a building. The display device 500 may be mounted on the same device as the inference device 10 or the control device 100.

[0054] 4 to 8, an application example of the inference device 10 in the learning phase will be described. As described above, the inference device 10 is configured to infer control data for controlling the drive power supplied to the electric motor 61 based on status data indicating the status of the electric motor 61. To achieve this type of inference, the inference device 10 performs reinforcement learning using AI.

[0055] 4 and 5 are diagrams for explaining an overview of reinforcement learning. As shown in FIG. 4, in reinforcement learning, an agent (acting subject) in a certain environment observes the current state (environmental parameters) and determines the action to be taken based on a policy. A policy is a rule for determining an action, and optimizing the policy optimizes the selection of an action. The environment dynamically changes depending on the agent's actions, and the agent is given a reward based on a reward criterion in response to the environmental changes. The agent repeats such actions and learns the action that will obtain the most reward through a series of actions. In FIG. 4, the current state is represented by B2 (state), the agent's action is represented by B1 (action), and the reward criterion is represented by D (reward criterion).

[0056] Q-learning and TD-learning are well-known as representative reinforcement learning methods. For example, in the case of Q-learning, the general update formula for the action value function Q(s, a) is expressed as the following formula (3):

[0057]

[0058] In formula (3), s ｔ is the state of the environment at time t, and a ｔ indicates the action at time t. ｔ Therefore, the state is s ｔ＋１ It changes to r ｔ＋１ indicates the reward that the agent can receive depending on the change in state, γ indicates the discount rate, and α indicates the learning coefficient. Note that γ is in the range of 0<γ≦1, and α is in the range of 0<α≦1. B1 (action) is action a ｔ and B2 (state) is state s ｔ and the state s at time t ｔ Best action in a ｔ is learned.

[0059] In equation (3), if the action value function Q of the action a with the highest Q value at time t+1 is larger than the action value function Q of the action a executed at time t, the action value function Q is increased; conversely, the action value function Q is decreased. In other words, the action value function Q(s, a) is updated so that the action value function Q of the action a at time t approaches the best action value at time t+1. This allows the best action value in a certain environment to be propagated sequentially to the action value in the previous environment.

[0060] As shown in Figure 5, in the learning phase, the inference device 10 executes a learning program 40 to generate (update) a learned model 20 based on learning data 30 including B1 (action) and B2 (state) and D (reward criterion).

[0061] In the utilization phase, the inference device 10 uses the trained model 20 to obtain C (output) based on B2 (state).

[0062] FIG. 6 is a diagram for explaining B1 (action), B2 (state), C (output), and D (reward criterion) in reinforcement learning according to the embodiment.

[0063] As shown in FIG. 6, in the inference device 10, control data for controlling the drive power supplied from the drive circuit 601 to the electric motor 61 is used as B1 (action). The control data is the value of the drive current in the drive power (I ａ ) and the phase of the drive current (β in equations (1) and (2)).

[0064] In the inference device 10, state data indicating the current operating state of the electric motor 61 is used as B2 (state). The state data includes both the rotation speed of the rotor 612 in the electric motor 61 when the rotor 612 is actually rotating (hereinafter also referred to as "actual rotation speed"), and a predetermined target rotation speed of the rotor 612 (hereinafter also referred to as "target rotation speed"). Furthermore, in addition to the actual rotation speed and target rotation speed of the electric motor 61, the state data also includes the bus voltage in the driving power (V in Equation (2)). ｏｍ ) may be included.

[0065] The status data may include any data that indicates the status of the electric motor 61. For example, the status data may include the output torque of the electric motor 61, the rotation speed or rotation angle of the rotor 612, motor constants such as an induced voltage constant or a torque constant required for drive control of the electric motor 61, the moment of inertia, various control gains, data related to the status of a compressor, a fan, an air conditioner, or the like that is driven by the electric motor 61, and the like.

[0066] In the inference device 10, control data for controlling the drive power supplied to the electric motor 61 is used as C (output). The control data is, like B1 (action), the value of the drive current in the drive power (I ａ ) and the phase of the drive current (β in equations (1) and (2)).

[0067] In the inference device 10, as D (reward standard), driving performance data related to the driving performance of the electric motor 61 and power efficiency data related to the power efficiency of the electric motor 61 are used. The driving performance data includes the absolute value of the difference between the actual rotation speed and the target rotation speed of the electric motor 61. The power efficiency data includes at least one of the value of the drive current in the drive power, the value of the drive power, and the value of the power supply power supplied from the power supply circuit 602 to the drive circuit 601.

[0068] 7 is a diagram showing the configuration of the learning device 110 in the learning phase. The learning device 110 is realized by the control unit 11 of the inference device 10. The learning device 110 can exchange data with each of the learning program storage unit 121 and the trained model storage unit 122. The learning program storage unit 121 and the trained model storage unit 122 are realized by the storage unit 12 of the inference device 10.

[0069] 7 , the learning device 110 includes a data acquisition unit 111 and a model generation unit 112. The learning device 110 executes a learning program 40 stored in a learning program storage unit 121 to generate a trained model 20 based on training data 30 including B1 (action) and B2 (state) and D (reward criterion).

[0070] The data acquisition unit 111 acquires learning data 30 including B1 (action) and B2 (state). Specifically, the data acquisition unit 111 acquires the control data shown in Fig. 6 as B1 (action). The data acquisition unit 111 acquires the state data shown in Fig. 6 as B2 (state).

[0071] The model generation unit 112 generates a trained model 20 that infers C (output) from B2 (state) using training data 30 including B1 (action) and B2 (state) acquired by the data acquisition unit 111. The model generation unit 112 stores the generated trained model 20 in the trained model storage unit 122.

[0072] The model generation unit 112 includes a reward calculation unit 1121 and a function update unit 1122. The reward calculation unit 1121 calculates a reward based on D (reward standard).

[0073] Specifically, the remuneration calculation unit 1121 calculates the remuneration based on the operating performance data of the electric motor 61 shown in FIG. 6 as D (remuneration standard). For example, the remuneration calculation unit 1121 recognizes the current actual rotation speed and the target rotation speed of the electric motor 61 based on the status data B2 (status). Furthermore, the remuneration calculation unit 1121 can also recognize the bus voltage in the driving power based on the status data B2 (status). Because the value and phase of the driving current are affected by the bus voltage, including the bus voltage in the status data allows the remuneration calculation unit 1121 to recognize the operating status of the electric motor 61.

[0074] In this state, the reward calculation unit 1121 calculates the difference between the actual rotation speed and the target rotation speed of the electric motor 61 when at least one of the value and phase of the drive current is controlled in accordance with the control data of B1 (action). The reward calculation unit 1121 calculates the reward based on the absolute value of the calculated difference.

[0075] Here, at least one of the value and phase of the drive current is controlled according to the control data of B1 (action), and therefore the absolute value of the difference between the actual rotation speed and the target rotation speed of the electric motor 61 varies depending on the control data. For example, if the value of the current supplied to the electric motor 61 becomes excessively small, the electric motor 61 cannot generate sufficient torque, and the actual rotation speed of the electric motor 61 becomes smaller than the target rotation speed. The smaller the actual rotation speed of the electric motor 61 becomes than the target rotation speed, the larger the absolute value of the difference between the actual rotation speed and the target rotation speed. Furthermore, if the value of the current supplied to the electric motor 61 becomes excessively large, the torque of the electric motor 61 becomes too large, and the actual rotation speed of the electric motor 61 becomes larger than the target rotation speed. The larger the actual rotation speed of the electric motor 61 becomes than the target rotation speed, the larger the absolute value of the difference between the actual rotation speed and the target rotation speed. Therefore, when the reward calculation unit 1121 controls at least one of the value and phase of the drive current in accordance with the control data of B1 (action), the smaller the absolute value of the difference between the actual rotation speed and the target rotation speed of the electric motor 61, the more appropriate the control data of B1 (action) is, and the larger the absolute value of the difference between the actual rotation speed and the target rotation speed of the electric motor 61, the less appropriate the control data of B1 (action) is, and the smaller the reward is.

[0076] As a result, the learning device 110 can train the estimation model so that it can infer appropriate control data that can realize operation of the electric motor 61 that satisfies the desired operating load depending on the operating state of the electric motor 61, and generate a learned model 20.

[0077] Furthermore, the reward calculation unit 1121 may calculate the reward based on the power efficiency data of the electric motor 61 shown in FIG. 6 in addition to the driving performance data as D (reward standard). Specifically, the reward calculation unit 1121 acquires at least one of the drive current value, drive power value, and source power value when at least one of the drive current value and phase is controlled in accordance with the control data of B1 (action) in a state corresponding to the state data of B2 (state). The learning device 110 can acquire the drive current value, drive power value, and source power value from the control device 100 of the electric motor unit 600.

[0078] The reward calculation unit 1121 calculates a reward based on the value of the drive current. Here, the copper loss generated in the electric motor 61 is determined in proportion to the square of the value of the drive current supplied to the electric motor 61. Therefore, when at least one of the value and phase of the drive current is controlled in accordance with the control data of B1 (action), the reward calculation unit 1121 increases the reward as the value of the drive current becomes smaller, since it is determined that the energy loss due to copper loss is more successfully reduced, and decreases the reward as the value of the drive current becomes larger, since it is determined that the energy loss due to copper loss is not as successfully reduced.

[0079] As a result, the learning device 110 can train the estimation model so as to minimize the value of the drive current, and generate the learned model 20. Therefore, the learning device 110 can realize operation of the electric motors 61 in a manner that minimizes energy loss due to copper loss, depending on the operating state of each electric motor 61.

[0080] The reward calculation unit 1121 also calculates a reward based on the value of the drive power. Here, the drive power supplied from the drive circuit 601 to the electric motor 61 is determined by the sum of the workload determined by the operating load of the electric motor 61 and the total energy loss, such as copper loss, mechanical loss, and iron loss, generated in the electric motor 61. When the electric motor 61 can operate with a constant operating load, i.e., with a constant workload determined by the operating load, minimizing the drive power supplied from the drive circuit 601 to the electric motor 61 is equivalent to minimizing the total energy loss generated in the electric motor 61. Therefore, when at least one of the value and phase of the drive current is controlled in accordance with the control data of B1 (action), the reward calculation unit 1121 increases the reward as the value of the drive power decreases, since it is determined that the total energy loss generated in the electric motor 61 is more successfully reduced. Conversely, the reward calculation unit 1121 decreases the value of the drive power as the value of the drive power decreases, since it is determined that the total energy loss generated in the electric motor 61 is not as successfully reduced.

[0081] This allows the learning device 110 to train the estimation model so as to minimize the value of the driving power, and generate the learned model 20. Therefore, the learning device 110 can realize operation of the electric motor 61 in a manner that minimizes not only copper loss but also all energy losses occurring in the electric motor 61, such as mechanical loss and iron loss, depending on the operating state of each electric motor 61.

[0082] The reward calculation unit 1121 also calculates a reward based on the value of the power supply power. Here, the power supply power supplied from the power supply circuit 602 to the drive circuit 601 is determined by the sum of the drive power supplied from the drive circuit 601 to the electric motor 61 and the total energy loss, such as copper loss, iron loss, and switching loss, generated in the drive circuit 601. When the electric motor 61 can operate with a constant operating load, i.e., with a constant workload determined by the operating load, minimizing the power supply power supplied from the power supply circuit 602 to the drive circuit 601 is equivalent to minimizing the total energy loss generated in the drive circuit 601 and the electric motor 61. Therefore, when at least one of the value and phase of the drive current is controlled in accordance with the control data for B1 (action), the reward calculation unit 1121 increases the reward as the value of the power supply power decreases, since it is determined that the total energy loss generated in the drive circuit 601 and the electric motor 61 is more successfully reduced. Conversely, the reward calculation unit 1121 decreases the value of the power supply power as the value of the power supply power increases, since it is determined that the total energy loss generated in the drive circuit 601 and the electric motor 61 is less successfully reduced.

[0083] This allows the learning device 110 to train the estimation model so as to minimize the value of the power supply power, and generate the learned model 20. Therefore, the learning device 110 can realize operation of the electric motor 61 in a way that minimizes total energy loss, such as copper loss, mechanical loss, and iron loss, not only in the electric motor 61 but also in the entire electrical system in the electric motor unit 600 including the drive circuit 601, depending on the operating state of the individual electric motor 61.

[0084] The function update unit 1122 updates the function for inferring control data based on the reward calculated by the reward calculation unit 1121. For example, in the case of Q-learning, the function update unit 1122 updates the action value function Q(S ｔ , a ｔ ) is used as a function to infer the control data.

[0085] The learned model storage unit 122 stores the action value function Q(S ｔ , a ｔ ) is stored as the trained model 20.

[0086] 8 is a flowchart showing the processing executed by the learning device 110 (control unit 11) in the learning phase. Note that FIG. 8 shows the processing executed by the inference device 10 corresponding to the learning device 110. Also, in FIG. 8, "S" is used as an abbreviation for "STEP."

[0087] 8, the learning device 110 acquires the control data shown in Fig. 6 as B1 (action) and acquires the state data shown in Fig. 6 as B2 (state) by the data acquisition unit 111 (S11). Note that the data acquisition unit 111 is not limited to acquiring B1 (action) and B2 (state) at the same time, and may acquire B1 (action) and B2 (state) at different times.

[0088] The learning device 110 calculates a reward based on the driving performance data of the electric motor 61, which is D (reward standard), using the reward calculation unit 1121 (S12). Note that the learning device 110 may also calculate a reward based on the power efficiency data of the electric motor 61, which is D (reward standard), in addition to the driving performance data of the electric motor 61, using the reward calculation unit 1121.

[0089] If the reward calculation unit 1121 of the learning device 110 determines that the reward should be increased based on the reward criterion, it increases the reward (S13), and if it determines that the reward should be decreased based on the reward criterion, it decreases the reward (S14).

[0090] Specifically, the reward calculation unit 1121 controls the drive power of the electric motor 61 based on the control data of B1 (action) in the state indicated by the state data of B2 (state), and calculates the reward based on the operating performance data or power efficiency data of the electric motor 61 acquired from the control device 100 at that time. When the absolute value of the difference between the actual rotation speed and the target rotation speed of the electric motor 61 is a first value, the reward calculation unit 1121 increases the reward more than when the absolute value is a second value greater than the first value. When the value of the drive current is the first value, the reward calculation unit 1121 increases the reward more than when the value of the drive current is a second value greater than the first value. When the value of the drive power is the first value, the reward calculation unit 1121 increases the reward more than when the value of the drive power is a second value greater than the first value. When the value of the source power is the first value, the reward calculation unit 1121 increases the reward more than when the value of the source power is a second value greater than the first value.

[0091] The learning device 110 uses the function update unit 1122 to update the action value function Q(s) of the formula (3) stored in the trained model storage unit 122 based on the reward calculated by the reward calculation unit 1121. ｔ , a ｔ ) is updated (S15).

[0092] The learning device 110 repeatedly executes the above-described processes from S11 to S15, and generates the action-value function Q(s ｔ , a ｔ ) as the trained model 20. After that, the training device 110 ends this process.

[0093] As described above, the learning device 110 according to the embodiment uses learning data 30 including control data for controlling the drive power supplied to the electric motor 61 and status data indicating the status of the electric motor 61 to generate a trained model 20 for inferring control data from the status data. This allows the learning device 110 to generate a trained model 20 that can infer optimal control data that truly maximizes the energy efficiency of the electric motor unit 600.

[0094] [Utilization Phase] An application example of the inference device 10 in the utilization phase will be described with reference to Figures 9 and 10. Figure 9 is a diagram showing the configuration of the inference device 10 in the utilization phase. The inference device 10 is capable of exchanging data with the trained model storage unit 122.

[0095] 9, the inference device 10 includes a data acquisition unit 111 and an inference unit 113. The inference device 10 uses the trained model 20 to obtain C (output) based on B2 (state).

[0096] The data acquisition unit 111 acquires the status data shown in FIG. 6 as B2 (status) from the control device 100 of the electric motor unit 600.

[0097] The inference unit 113 reads the learned model 20 from the learned model storage unit 122, and uses the learned model 20 to infer the control data shown in Figure 6 as C (output) based on the state data of B2 (state). The inference unit 113 outputs the inferred control data to the control device 100 of the electric motor unit 600.

[0098] 10 is a flowchart showing the processing executed by the inference device 10 (control unit 11) in the utilization phase. In FIG. 10, "S" is used as an abbreviation for "STEP."

[0099] 10, the inference device 10 acquires the status data shown in FIG. 6 as B2 (status) by the data acquisition unit 111 (S21). The status data acquired at this time includes the rotation speed and target rotation speed of the electric motor 61 at the moment when the inference device 10 acquires the status data. Furthermore, the status data acquired at this time may include the bus voltage of the drive power supplied from the drive circuit 601 to the electric motor 61.

[0100] The inference device 10 inputs the state data to the trained model 20 by the inference unit 113 (S22), infers control data, and outputs it to the control device 100 of the electric motor unit 600 (S23). After that, the inference device 10 ends this process.

[0101] As described above, the inference device 10 according to the embodiment uses the learned model 20 to infer control data for controlling the driving power supplied to the electric motor 61 based on status data indicating the status of the electric motor 61.

[0102] This allows the inference device 10 to infer optimal control data that satisfies the desired operating load based on the individual variations and operating conditions of each electric motor 61, and that minimizes all energy losses, such as copper loss, mechanical loss, and iron loss, in the entire electrical system of the electric motor unit 600, including the drive circuit 601, thereby truly maximizing energy efficiency.

[0103] The control device 100 controls the driving of the electric motor 61 based on the control data inferred by the inference device 10, and thereby calculates the control data (the value of the driving current I ａ and phase β), it is possible to realize operation of the electric motor 61 that truly maximizes energy efficiency.

[0104] Even when performing field-weakening control that satisfies the constraints shown in equation (2), the control device 100 can realize operation of the electric motor 61 that truly maximizes energy efficiency by controlling the drive of the electric motor 61 based on the control data inferred by the inference device 10.

[0105] Even when the control device 100 estimates the rotation angle of the rotor 612 using sensorless control based on a predetermined calculation formula (theoretical formula), it can control the driving of the electric motor 61 based on the control data inferred by the inference device 10, thereby realizing operation of the electric motor 61 in a way that truly maximizes energy efficiency.

[0106] [Modification] A modification of the above-described embodiment will be described below. Note that only the differences from the above-described embodiment will be described below.

[0107] The inference device 10 may generate the trained model 20 using the model generation unit 112 before shipping the trained model 20, and then retrain the trained model 20 using the training data 30 obtained after shipping the trained model 20 to generate a new trained model 20.

[0108] Specifically, the reinforcement learning shown in FIG. 8, which was performed before the trained model 20 was shipped, was used to calculate the action value function Q(s ｔ , a ｔ After the trained model 20 is shipped, the action value function Q(s) of the formula (3) is calculated again by the reinforcement learning shown in FIG. 8 in the actual use environment (user use environment) of the electric motor 61. ｔ , a ｔ ) may be generated.

[0109] In this way, the inference device 10 may generate a new trained model 20 that can more accurately infer appropriate control data by performing reinforcement learning (transfer learning) in a post-shipment environment using the trained model 20, which is the outcome of reinforcement learning performed in a pre-shipment environment. This allows the inference device 10 to shorten the learning time in the actual usage environment for generating the trained model 20.

[0110] The control device 100 may control the driving power of the electric motor 61 by using both the control data calculated using the theoretical formulas (1) and (2) and the control data inferred by the inference device 10.

[0111] Specifically, the control device 100 drives the electric motor 61 based on control data calculated using the theoretical equations (1) and (2) during a predetermined period. Meanwhile, the inference device 10 generates a trained model 20 using state data acquired from the electric motor 61, which is driven based on the control data calculated using the theoretical equations (1) and (2) during the predetermined period. After the predetermined period has elapsed, the control device 100 drives the electric motor 61 based on the control data inferred by the trained model 20. The predetermined period may be a period during which the inference device 10 can accumulate training data 30 for generating a trained model 20 capable of inferring appropriate control data.

[0112] In this way, the control device 100 calculates the control data using the theoretical formulas of Equation (1) and Equation (2) until sufficient training data 30 for generating the trained model 20 is accumulated, and after sufficient training data 30 for generating the trained model 20 is accumulated, the control device 100 generates the trained model 20 using the training data 30 and infers the control data using the generated trained model 20. This allows the control device 100 to control the electric motor 61 using appropriate control data.

[0113] The control device 100 may compare the control data calculated using the theoretical formulas (1) and (2) with the control data inferred by the trained model 20, and if the difference between the two exceeds a threshold, drive the electric motor 61 based on the control data calculated using the theoretical formulas (1) and (2).

[0114] Here, if the inference device 10 acquires state data in a situation different from the situation at the time of learning, there is a risk that the inference device 10 will infer inappropriate control data. If the control device 100 controls the electric motor 61 based on the inappropriate control data, there is a risk that the electric motor 61 will enter an unexpected operating state. Therefore, if the control data inferred by the trained model 20 significantly deviates from the control data calculated using the theoretical formulas (1) and (2), the control device 100 drives the electric motor 61 by prioritizing the control data calculated using the theoretical formulas (1) and (2), thereby preventing the electric motor 61 from entering an unexpected operating state.

[0115] The inference device 10 may generate the trained model 20 using a predetermined amount or more of training data by the model generation unit 112. Specifically, after the inference device 10 has accumulated a predetermined amount or more of training data 30 for generating the trained model 20 capable of inferring appropriate control data, the inference device 10 generates the trained model 20 using the predetermined amount or more of training data 30. This allows the inference device 10 to infer appropriate control data using the trained model 20 that has been sufficiently trained by reinforcement learning.

[0116] The model generation unit 112 of the inference device 10 may include fewer types of data as state data in the learning data 30 than the types of data obtainable from the electric motor unit 600. The state data obtainable by the inference device 10 from the control device 100 of the electric motor unit 600 includes the actual and target rotation speeds of the electric motor 61, the bus voltage in the drive power, the output torque of the electric motor 61, the rotation speed or rotation angle of the rotor 612, motor constants such as an induced voltage constant or a torque constant required for drive control of the electric motor 61, the moment of inertia, and various control gains. The inference device 10 generates the trained model 20 using the actual and target rotation speeds of the electric motor 61 and the bus voltage in the drive power, among these data. Note that the inference device 10 does not necessarily need to use the bus voltage in the drive power to infer the control data.

[0117] Furthermore, when selecting some of the status data from the status data obtainable from the electric motor unit 600, the inference device 10 may perform dimensionality reduction on the status data obtainable from the electric motor unit 600, thereby selecting fewer types of data than the types of data obtainable from the electric motor unit 600 as status data to include in the training data 30. When performing dimensionality reduction, the inference device 10 may use a known dimensionality reduction algorithm such as principal component analysis. It is preferable that the inference device 10 preferentially selects status data that is particularly closely related to the control data from the status data obtainable from the electric motor unit 600.

[0118] In this way, the inference device 10 generates the trained model 20 using only some of the state data rather than using all of the state data that can be obtained from the motor unit 600, thereby reducing the computational load required for learning and shortening the calculation time.

[0119] [Summary] The inference device 10 of the present disclosure includes a data acquisition unit 111 that acquires status data indicating the status of the electric motor 61, and an inference unit 113 that infers control data based on the status data acquired by the data acquisition unit 111 using a trained model 20 for inferring control data based on the status data.

[0120] With this configuration, the inference device 10 can use the learned model 20 to infer control data based on status data indicating the status of the electric motor 61, thereby being able to infer optimal control data according to the status of each individual electric motor 61.

[0121] The inference device 10 further includes a model generation unit 112 that generates a trained model 20 using training data 30 including control data and state data.

[0122] With this configuration, the inference device 10 can use the learning data 30 to generate a learned model 20 that can infer optimal control data depending on the state of each electric motor 61.

[0123] The model generation unit 112 includes a reward calculation unit 1121 that calculates a reward based on a reward standard, and a function update unit 1122 that updates a function for inferring control data based on the reward.

[0124] With this configuration, the inference device 10 can generate a learned model 20 that can infer optimal control data according to the state of each electric motor 61 through reinforcement learning using rewards.

[0125] The remuneration calculation unit 1121 calculates the remuneration based on the driving performance data relating to the driving performance of the electric motor 61 and the power efficiency data relating to the power efficiency of the electric motor 61 as the remuneration standard.

[0126] With this configuration, the inference device 10 can generate a learned model 20 that can infer optimal control data according to the state of each electric motor 61 by performing reinforcement learning using rewards calculated based on the operating performance data and power efficiency data of the electric motor 61.

[0127] The driving performance data used as the remuneration standard includes the absolute value of the difference between the actual rotation speed and the target rotation speed of the electric motor 61 .

[0128] With this configuration, the inference device 10 performs reinforcement learning using a reward calculated based on the absolute value of the difference between the actual rotation speed and the target rotation speed of the electric motor 61, and can infer control data that can realize operation of the electric motor 61 that satisfies the desired operating load depending on the state of each electric motor 61.

[0129] The driving power is supplied from the driving circuit 601 to the electric motor 61. The driving circuit 601 supplies the driving power to the electric motor 61 using the power supply power supplied from the power supply circuit 602. The power efficiency data includes at least one of the value of the driving current in the driving power, the value of the driving power, and the value of the power supply power.

[0130] With this configuration, the inference device 10 performs reinforcement learning using a reward calculated based on at least one of the value of the drive current in the drive power, the value of the drive power, and the value of the power supply power, and can infer control data that can realize operation of the electric motor 61 in a way that minimizes total energy loss generated in the electric motor 61, such as copper loss, mechanical loss, and iron loss, depending on the state of each electric motor 61.

[0131] The control data includes at least one of the value of the drive current in the drive power and the phase of the drive current.

[0132] With this configuration, the inference device 10 can use the learning data 30 to infer the optimum value and phase of the drive current depending on the state of each electric motor 61 .

[0133] The status data indicating the status of the electric motor 61 includes the actual rotation speed and the target rotation speed of the electric motor 61 .

[0134] With this configuration, the inference device 10 can infer optimal control data according to the state of each electric motor 61 based on the actual rotation speed and target rotation speed of the electric motor 61 .

[0135] The status data indicating the status of the electric motor 61 includes the bus voltage of the drive power supplied from the drive circuit 601 to the electric motor 61 .

[0136] With this configuration, the inference device 10 can infer optimal control data according to the state of each electric motor 61 based on the bus voltage of the driving power.

[0137] The electric motor 61 includes a stator 611 , a rotor 612 , and a permanent magnet 618 provided on at least one of the stator 611 and the rotor 612 .

[0138] With this configuration, even when performing field-weakening control on the electric motor 61, the control device 100 can control the driving of the electric motor 61 based on the control data inferred by the inference device 10, thereby realizing operation of the electric motor 61 in a way that truly maximizes energy efficiency.

[0139] The electric motor 61 rotates the rotor 612 at a rotation angle calculated based on a predetermined calculation formula (theoretical formula).

[0140] With this configuration, even when the control device 100 estimates the rotation angle of the rotor 612 using sensorless control based on a predetermined calculation formula (theoretical formula), it can control the drive of the electric motor 61 based on the control data inferred by the inference device 10, thereby realizing operation of the electric motor 61 that truly maximizes energy efficiency.

[0141] The model generation unit 112 generates the trained model 20 before shipping the trained model 20, and then retrains the trained model 20 using the training data 30 obtained after shipping the trained model 20 to generate a new trained model 20.

[0142] With this configuration, the inference device 10 can shorten the learning time in an actual usage environment to generate the trained model 20.

[0143] The electric motor 61 is driven for a predetermined period based on control data calculated using the theoretical equations (1) and (2). The model generation unit 112 generates the trained model 20 using state data acquired from the electric motor 61, which is driven for a predetermined period based on the control data calculated using the theoretical equations (1) and (2). After the predetermined period has elapsed, the electric motor 61 is driven based on the control data inferred by the trained model 20.

[0144] With this configuration, the control device 100 controls the electric motor 61 based on control data calculated using the theoretical formulas (1) and (2) until sufficient learning data 30 for generating the trained model 20 has been accumulated, and after sufficient learning data 30 for generating the trained model 20 has been accumulated, the control device 100 can control the electric motor 61 based on control data inferred by the trained model 20, and therefore can control the electric motor 61 using appropriate control data.

[0145] When the difference between the control data calculated using the theoretical formulas (1) and (2) and the control data inferred by the trained model 20 exceeds a threshold, the electric motor 61 is driven based on the control data calculated using the theoretical formulas (1) and (2).

[0146] With this configuration, when the control data inferred by the learned model 20 significantly deviates from the control data calculated using the theoretical formulas (1) and (2), the control device 100 drives the electric motor 61 by prioritizing the control data calculated using the theoretical formulas (1) and (2), thereby preventing the electric motor 61 from entering an unexpected operating state.

[0147] The model generation unit 112 generates a trained model 20 using a predetermined amount or more of training data 30.

[0148] With this configuration, the inference device 10 can infer appropriate control data using the learned model 20 that has been sufficiently trained by reinforcement learning.

[0149] The model generating unit 112 includes, as state data in the learning data 30, fewer types of data than the types of data that can be acquired regarding the state of the electric motor 61.

[0150] According to this configuration, the inference device 10 generates the trained model 20 using only some of the state data rather than using all of the state data that can be obtained from the control device 100, thereby reducing the computational load required for learning and shortening the computation time.

[0151] The model generation unit 112 performs dimensionality reduction on the data that can be obtained regarding the state of the electric motor 61, and includes fewer types of data as state data in the learning data 30 than the types of data that can be obtained regarding the state of the electric motor 61.

[0152] According to this configuration, the inference device 10 generates the trained model 20 using only some of the state data rather than using all of the state data that can be obtained from the motor unit 600, thereby reducing the computational load required for learning and shortening the calculation time.

[0153] The inference device 10 further includes a presentation unit 114 that presents information relating to the driving of the electric motor 61 based on the control data inferred by the inference unit 113 .

[0154] With this configuration, the inference device 10 can encourage the user to use the inference device 10 .

[0155] The inference method disclosed herein includes, as processing executed by the inference device 10, a step (S21) of acquiring status data indicating the status of the electric motor 61, and steps (S22, S23) of inferring control data based on the status data acquired by the acquiring step (S21) using a trained model 20 for inferring control data based on the status data.

[0156] With this configuration, the inference device 10 can use the learned model 20 to infer control data based on status data indicating the status of the electric motor 61, thereby being able to infer optimal control data according to the status of each individual electric motor 61.

[0157] The embodiments disclosed herein should be considered to be illustrative in all respects and not restrictive. The scope of the present disclosure is defined by the claims, not by the description of the above embodiments, and is intended to include all modifications within the meaning and scope of the claims.

[0158] 10 inference device, 11 control unit, 12 memory unit, 13 input / output unit, 20 trained model, 30 training data, 40 training program, 60 housing, 61 electric motor, 67 glass terminal, 100 control device, 110 learning device, 111 data acquisition unit, 112 model generation unit, 113 inference unit, 114 presentation unit, 121 training program storage unit, 122 trained model storage unit, 500 display device, 600 electric motor unit, 601 drive circuit, 602 power supply circuit, 611 stator, 612 rotor, 613 shaft, 614 slot, 615 winding, 616 shaft hole portion, 617 air hole portion, 618 permanent magnet, 619 central hole portion, 1000 inference system, 1121 reward calculation unit, 1122 function update unit.

Claims

1. An inference device that infers control data for controlling a driving power supplied to an electric motor, A data acquisition unit that acquires status data indicating a status of the electric motor; an inference unit that infers the control data based on the state data acquired by the data acquisition unit, using a trained model for inferring the control data based on the state data; a model generation unit that generates the trained model using learning data including the control data and the state data; The model generation unit a remuneration calculation unit that calculates remuneration based on the remuneration standard; a function update unit that updates a function for inferring the control data based on the reward; The reward calculation unit calculates the reward based on driving performance data related to the driving performance of the electric motor and power efficiency data related to the power efficiency of the electric motor as the reward standard.

2. The inference device according to claim 1 , wherein the driving performance data includes an absolute value of a difference between a rotation speed of the electric motor and a target rotation speed.

3. The driving power is supplied to the electric motor from a driving circuit, the drive circuit supplies the drive power to the electric motor using power supply power supplied from a power supply circuit; 3. The inference device according to claim 1, wherein the power efficiency data includes at least one of a value of a drive current at the drive power, a value of the drive power, and a value of the power supply power.

4. 3. The inference device according to claim 1, wherein the control data includes at least one of a value of a drive current in the drive power and a phase of the drive current.

5. 3. The inference device according to claim 1, wherein the state data includes a rotation speed and a target rotation speed of the electric motor.

6. The inference device according to claim 1 or 2, wherein the state data includes a bus voltage in the driving power.

7. The inference device according to claim 1 or 2, wherein the electric motor comprises a stator, a rotor, and a permanent magnet provided on at least one of the stator and the rotor.

8. The inference device according to claim 7 , wherein the electric motor rotates the rotor at a rotation angle calculated using a predetermined formula.

9. 2. The inference device according to claim 1, wherein the model generation unit generates the trained model before shipment of the trained model, and then retrains the trained model using the learning data acquired after shipment of the trained model to generate a new trained model.

10. The electric motor is driven based on the control data calculated using a predetermined formula for a predetermined period of time, The model generation unit generates the trained model by using the state data acquired from the electric motor that is driven based on the control data calculated using the predetermined formula during the predetermined period; The inference device according to claim 1 , wherein the electric motor is driven based on the control data inferred by the trained model after the predetermined period has elapsed.

11. 2. The inference device according to claim 1, wherein the electric motor is driven based on the control data calculated using a predetermined formula when a difference between the control data calculated using the predetermined formula and the control data inferred by the trained model exceeds a threshold value.

12. The inference device according to claim 1 , wherein the model generation unit generates the trained model using a predetermined amount or more of the learning data.

13. The inference device according to claim 1 , wherein the model generation unit includes in the learning data as the state data fewer types of data than types of data that can be acquired regarding the state of the electric motor.

14. The inference device according to claim 13, wherein the model generation unit performs dimensionality reduction on data obtainable regarding the state of the electric motor, so that fewer types of data than the types of data obtainable from the electric motor are included in the learning data as the state data.

15. 3. The inference device according to claim 1, further comprising a presentation unit that presents information regarding the driving of the electric motor based on the control data inferred by the inference unit.

16. 1. A method for inferring control data for controlling a driving power supplied to an electric motor by a computer, comprising: The process executed by the computer is acquiring status data indicative of a status of the electric motor; Inferring the control data based on the state data acquired by the acquiring step, using a trained model for inferring the control data based on the state data; generating the trained model using learning data including the control data and the state data; The step of generating the trained model includes: calculating a reward based on the reward criterion; updating a function for inferring the control data based on the reward; An inference method, wherein the step of calculating the reward calculates the reward based on operational performance data related to operational performance of the electric motor and power efficiency data related to power efficiency of the electric motor as the reward criteria.