Power converter

Abstract

Apply machine learning to sensorless control of an electric motor performed by a power conversion device to achieve stable and high-precision control characteristics. 【Solution means】The power conversion device 1 includes a phase calculation unit 11 that calculates a phase estimation value θ with respect to the phase of the motor M, ^ a vector control calculation unit 8 that calculates a d-axis voltage command value v dc and a q-axis current command value v qc for the motor M based on the d-axis current detection value i dc_IM ** and a q-axis current detection value i qc_IM ** and a machine learning device 10. The machine learning device 10 has a learned machine learning model, and inputs an input value including at least one of the d-axis voltage command value v dc_IM ** , the q-axis voltage command value v qc_IM ** , the d-axis current detection value i dc and the q-axis current detection value i qc into the machine learning model, and outputs a slip frequency estimation value ω s ^^ and a second rotational frequency estimation value ω r ^^ which are output values related to the phase of the motor M. The phase calculation unit 11 calculates the phase estimation value θ ^ based on these output values.

Description

[Technical Field]

[0001] This invention relates to a power conversion device. [Background technology]

[0002] As background technology for the present invention, for example, Patent Document 1 is known. Patent Document 1 describes a machine learning device that performs correction of electrical parameter setpoints and gain correction of speed control and current control in a vector control system of a permanent magnet synchronous motor, thereby achieving highly accurate speed control and current control. [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2022-77267 [Overview of the project] [Problems that the invention aims to solve]

[0004] In recent years, in order to reduce costs and improve reliability, a sensorless control method for electric motors has been proposed that estimates the magnetic pole position based on the electric motor's current and voltage, without using position sensors such as resolvers to detect the electric motor's magnetic pole position, and performs vector control of the electric motor based on the estimation result. However, the machine learning device described in Patent Document 1 is based on the premise that a position sensor is mounted on the electric motor, and therefore cannot be applied to power conversion devices that perform sensorless control of electric motors.

[0005] In view of the above problems, the present invention aims to achieve stable and highly accurate control characteristics by applying machine learning to the sensorless control of an electric motor performed by a power conversion device. [Means for solving the problem]

[0006] The power conversion device according to the present invention includes a phase calculation unit that calculates a phase estimation value with respect to the phase of an electric motor, a coordinate conversion unit that converts the current value of the alternating current flowing through the electric motor into a current value in a rotating coordinate system based on the phase estimation value calculated by the phase calculation unit, a vector control calculation unit that calculates a voltage command value for the electric motor based on the current value in the rotating coordinate system, a power converter that generates the alternating current based on the voltage command value, and a machine learning device that has a learned machine learning model and inputs an input value including at least one of the voltage command value and the current value in the rotating coordinate system into the machine learning model to output an output value related to the phase. The phase calculation unit calculates the phase estimation value based on the output value.

Advantages of the Invention

[0007] According to the present invention, machine learning can be applied to the sensorless control of an electric motor performed by a power conversion device, and stable and highly accurate control characteristics can be realized.

Brief Description of the Drawings

[0008] [Figure 1] Configuration diagram of the power conversion device according to the first embodiment of the present invention. [Figure 2] Functional block diagram of the machine learning model used in the first embodiment of the present invention. [Figure 3] Block diagram showing a configuration example of each neuron in the machine learning model. [Figure 4] Diagram showing an example of a learning data acquisition system used for acquiring learning data of the machine learning model. [Figure 5] Diagram showing an example of learning data. [Figure 6] Waveform diagram in which the values of the time-series data of the input signal and the time-series data of the teacher signal in the learning data are graphed. [Figure 7] Diagram showing the state of learning of the machine learning model. [Figure 8] Explanation diagram of the improvement effect of the control characteristics according to the present invention. [Figure 9]A diagram illustrating the configuration of a power conversion device according to a second embodiment of the present invention. [Figure 10] A functional block diagram of a machine learning model used in a second embodiment of the present invention. [Figure 11] A diagram illustrating the configuration of a power conversion device according to a third embodiment of the present invention. [Figure 12] A functional block diagram of a machine learning model used in a third embodiment of the present invention. [Figure 13] A diagram showing the configuration of a power conversion device according to a fourth embodiment of the present invention. [Figure 14] A diagram showing the configuration of a power conversion device according to a fifth embodiment of the present invention. [Figure 15] A diagram showing the configuration of a power conversion device according to the sixth embodiment of the present invention. [Modes for carrying out the invention]

[0009] Embodiments of the present invention will be described in detail below with reference to the drawings. Common components in each drawing are given the same reference numerals. Furthermore, the embodiments described below are not limited to those shown.

[0010] <First Embodiment> Figure 1 is a diagram showing the configuration of a power converter according to the first embodiment of the present invention. The power converter 1 shown in Figure 1 is connected to a motor M, which is an induction motor, and controls the drive of the motor M by controlling the AC power supplied to the motor M. Current flows through the motor M in accordance with the AC power supplied from the power converter 1, consisting of a magnetic flux axis (d axis) component and a torque axis (q axis) component perpendicular to the magnetic flux axis. This generates torque in the motor M, and the motor M is driven.

[0011] The power conversion device 1 includes a power converter 2 connected to a DC power supply 3, a coordinate conversion unit 5, a speed control calculation unit 6, a d-axis current flux setting unit 7, a vector control calculation unit 8, a frequency calculation unit 9, a machine learning device 10, a phase calculation unit 11, and a coordinate conversion unit 12. For each component except the power converter 2, it is realized by, for example, the software of a microcomputer mounted on the power conversion device 1 or the like.

[0012] The power converter 2 converts the DC power supplied from the DC power supply 3 into AC power based on the three-phase AC voltage command values v u * , v v * , v w * input from the coordinate conversion unit 12, thereby generating three-phase AC currents i u , i v , i w and outputting them to the motor M. Thereby, drive control of the motor M is performed using variable voltage and variable frequency AC power corresponding to the voltage command values v u * , v v * , v w * . Note that the power converter 2 can be configured by combining a plurality of semiconductor elements such as, for example, IGBT (Insulated Gate Bipolar Transistor) and MOSFET (Metal Oxide Semiconductor Field Effect Transistor).

[0013] An electric current detector 4 is provided on the AC electric wire connecting the power converter 2 and the motor M. The electric current detector 4 detects the three-phase AC currents i u , i v , i w flowing through the motor M, and outputs AC current detection values i uc , i vc , i wc corresponding to these detection results. Note that the electric current detector 4 may be mounted inside the power converter 2 or the motor M. Also, the three-phase AC currents i u , i v , iw Any two of these, for example, the U phase and the W phase, are alternating current i u ,i w The current detectors 4 detect each of these currents, and the remaining one phase (V phase) AC current is determined in the power converter 1 under the AC condition (i u +i v +i w From =0), i v =-(i u +i w You can also calculate it as follows:

[0014] The coordinate transformation unit 5 calculates the phase estimate value θ of the motor M calculated by the phase calculation unit 11. ^ Based on this, the 3-phase AC current detection value i input from the current detector 4 is used. uc ,i vc ,i wc The d-axis current detection value i represents the current value in the rotating coordinate system. dc and q-axis current detection value i qc Convert to.

[0015] The speed control calculation unit 6 receives the rotation frequency command value ω from the higher-level device to the power converter 1. r * and the second rotation frequency estimate ω output from the machine learning device 10 r ^^ Based on the difference, the torque command value τ for motor M is calculated. * The q-axis current command value i is obtained by calculating this and dividing it by a predetermined torque coefficient. q * Perform the calculation and output the result.

[0016] The d-axis current flux setting unit 7 sets the preset d-axis current command value i d * and the d-axis secondary frequency command value f 2d * Outputs.

[0017] The vector control calculation unit 8 calculates the current command values ​​i for the d axis and q axis. d * ,i q * And the current detection value i in the rotating coordinate system dc ,iqc The primary frequency command value ω1 calculated by the phase calculation unit 11 * And the d-axis secondary frequency command value f 2d * Based on this, the voltage command values ​​for the d-axis and q-axis of the induction motor M are v dc_IM ** ,v qc_IM ** The system calculates and outputs the results of these calculations. At this time, the vector control calculation unit 8 uses, for example, the electrical circuit parameters of the motor M to calculate the d-axis voltage command value v dc_IM ** and q-axis voltage command value v qc_IM ** It is possible to perform the following calculations.

[0018] The frequency calculation unit 9 calculates the q-axis voltage command value v qc_IM ** And the d-axis secondary frequency command value f 2d * And the q-axis current command value i q * And the detected current value i on the q axis of the rotating coordinate system qc Using this, the slip frequency command value ω for motor M s * And the first rotational frequency estimate ω represents the estimated rotational frequency of motor M. r ^ The unit performs calculations on and outputs the results of these calculations. At this time, the frequency calculation unit 9 uses, for example, the electrical circuit parameters of the motor M to determine the slip frequency command value ω s * and the first rotation frequency estimate ω r ^ It is possible to perform the following calculations.

[0019] The machine learning device 10 is a device that performs AI calculations using a trained machine learning model. It inputs predetermined input values ​​corresponding to the drive state of the motor M into the machine learning model and outputs output values ​​related to the phase of the motor M. In this embodiment, for example, the d-axis voltage command value v is used as an input value to the machine learning model. dc_IM ** and q-axis voltage command value v qc_IM ** And the current detection value i in the rotating coordinate systemdc , i qc and the slip frequency command value ω obtained by the frequency calculation unit 9 s * and the first rotational frequency estimated value ω r ^ are used. In the machine learning device 10, a machine learning model that has been pre-learned from the relationship between these values and the slip frequency and rotational frequency of the motor M is stored. By using this machine learning model, the machine learning device 10 can obtain, as output values related to the phase of the motor M, the slip frequency estimated value ω s * of the motor M with respect to the slip frequency command value ω s ^^ and the second rotational frequency estimated value ω r ^ obtained by correcting the first rotational frequency estimated value ω r ^^ and output them.

[0020] The phase calculation unit 11 calculates a phase estimated value θ with respect to the phase of the motor M ^ . In the present embodiment, the phase calculation unit 11 calculates, based on the slip frequency estimated value ω s ^^ and the second rotational frequency estimated value ω r ^^ obtained by the machine learning device 10, the phase estimated value θ ^ and the command value ω1 with respect to the fundamental frequency of the motor M * and outputs them by calculation. <00​​​​​​​​​​​​​​​​​​​​​​v * ,v w * As mentioned above, when this is input to the power converter 2, the three-phase alternating current i u ,i v ,i w A signal is generated, and the drive control of motor M is performed.

[0022] Next, the processing details of the power converter 1 of this embodiment will be described. First, the basic operation of the speed sensorless vector control method using the machine learning device 10, which is a feature of the power converter 1 of this embodiment, will be described below.

[0023] In the speed control calculation unit 6, the rotation frequency command value ω r * The second rotation frequency estimate ω r ^^ To ensure that it follows, proportional control and integral control are used to generate the torque command τ according to equation (1) below. * and q-axis current command value i q * Perform the calculation.

[0024]

number

[0025] The definitions of each constant in equation (1) are as follows: K sp : Proportional gain for speed control K si Integral gain of speed control P m :Polar logarithm Φ 2d d-axis secondary magnetic flux M: Mutual inductance L2: Secondary inductance

[0026] Furthermore, in equation (1), the asterisk (*) above each constant indicates that it is a value predetermined according to the electrical circuit parameters of the motor M. This point is also true for each constant in the following equations.

[0027] In the vector control calculation unit 8, first, the setting value R1 of the primary winding resistance, which is an electrical circuit parameter of the motor M, is determined. * The set values ​​L for the primary and secondary leakage inductances in the windings of motor M. s * And the setting value Φ for the secondary magnetic flux of the d axis 2d * And the current detection values ​​i of the d-axis and q-axis of the rotating coordinate system dc ,i qc And the current command values ​​i for the d axis and q axis d * ,i q * And the primary frequency command value ω1 * Using the above, the voltage reference values ​​v for the d axis and q axis are calculated according to equation (2) below. dc_IM * ,v qc_IM * Perform the calculation.

[0028]

number

[0029] Note that in equation (2), T acr This represents the response time constant of the current control.

[0030] Next, the current detection values ​​i for the d axis and q axis dc ,i qc However, the current command values ​​i for the d axis and q axis d * ,i q * To track these changes, proportional and integral control are used to calculate the voltage correction values ​​Δv for the d-axis and q-axis according to equation (3) below. dc_IM Δv qc_IM Perform the calculation.

[0031]

number

[0032] The definitions of each constant in equation (3) are as follows: K pd_IM : Proportional gain of current control on the d axis K id_IM : Integral gain of current control on the d axis K pq_IM : Proportional gain of current control on the q axis K iq_IM Integral gain of current control on the q axis

[0033] Based on the above calculation, the voltage reference values ​​v for the d axis and q axis are obtained. dc_IM * ,v qc_IM * And the voltage correction value Δv for the d axis and q axis. dc_IM Δv qc_IM Once these values ​​are obtained, they are added together according to equation (4) below to obtain the voltage command values ​​v for the d axis and q axis. dc_IM ** , v qc_IM ** Perform the calculation.

[0034]

number

[0035] In the frequency calculation unit 9, the q-axis voltage command value v qc_IM ** And the d-axis secondary frequency command value f 2d * And the set value M of mutual inductance * Using and the following equation (5), the slip frequency command value ω s * Perform the calculation.

[0036]

number

[0037] Furthermore, the q-axis voltage command value v qc_IM ** And the q-axis current command value i q * Using the electrical parameters of motor M, the first rotational frequency estimate ω is calculated according to equation (6) below.r ^ Perform the calculation.

[0038]

number

[0039] The definitions of each constant in equation (6) are as follows: T acr : Response time constant of current control R2 ’ : Primary side equivalent value of secondary winding resistance T obs : Time constant of the disturbance observer

[0040] The phase calculation unit 11 calculates the slip frequency estimate ω, which is the output of the machine learning device 10. s ^^ and the estimated second rotation frequency ω r ^^ Using this, the primary frequency command value ω1 is calculated according to equation (7) below. * The following is calculated. Also, according to equation (8), the phase estimate θ ^ Perform the calculation.

[0041]

number

number

[0042] The basic operation of the speed sensorless vector control method in the power converter 1 of this embodiment is as described above. Next, the details of the machine learning device 10, which is a feature of this embodiment, will be described.

[0043] Figure 2 is a functional block diagram of a machine learning model used in the first embodiment of the present invention. The machine learning device 10 of this embodiment has a pre-trained machine learning model 100. The machine learning model 100 is a deep learning model for AI computation using, for example, a neural network, and calculates and outputs output values ​​for a predetermined combination of input values ​​using AI computation. Specifically, as shown in Figure 2, for example, the machine learning model 100 is composed of an input layer 101, a hidden layer 102, and an output layer 103.

[0044] Multiple types of input values ​​for AI calculation are input to the input layer 101. In this embodiment, for example, the six types of input values ​​shown in Figure 2 (d-axis voltage command value v dc_IM ** q-axis voltage command value v qc_IM ** d-axis current detection value i dc q-axis current detection value i qc , slip frequency command value ω s * and the first rotation frequency estimate ω r ^ The combination of the above is input to the input layer 101. Note that it is not necessary to input all six of these input values ​​to the input layer 101. Alternatively, other variables, such as the active power value of the rotating coordinate system (i qc ×v qc_IM ** ) and reactive power value (i dc ×v dc_IM ** An input value further including at least one of the above may be input to the input layer 101.

[0045] The hidden layer 102 is composed of multiple neurons 104 arranged hierarchically, and performs a predetermined weighting operation on each input value input to the input layer 101, according to the hierarchy. By repeating this process, forward and backward propagation characteristics are applied to each input value within the hidden layer 102. The result of the operation at the final layer of the hidden layer 102 is input to the output layer 103.

[0046] The output layer 103 generates a predetermined output value using the calculation result from the final layer of the hidden layer 102. In this embodiment, for example, there are two types of output values ​​shown in Figure 2 (estimated slip frequency ω s ^^ and the estimated second rotation frequency ω r ^^ The combination of ) is generated by the output layer 103. The output values ​​generated by the output layer 103 are output from the machine learning model 100 as the output values ​​of the machine learning device 10.

[0047] Figure 3 is a block diagram showing an example configuration of each neuron 104 in the machine learning model 100 of Figure 2. Neuron 104 receives output values ​​X1~X from the previous layer. j For each of these, the pre-trained weight values ​​W1~W j Using the weight information 1041 consisting of the above, the bias information 1042 representing the bias value b, and the activation function 1043 represented by the function f, the calculated value y at the neuron 104 is obtained by performing a calculation represented by, for example, the following equation (9).

[0048]

number

[0049] The calculated value y of each neuron 104 obtained by equation (9) is input to the next layer. In the hidden layer 102, such calculations are performed at each neuron 104.

[0050] Note that the machine learning model 100 described in Figures 2 and 3 is just one example, and other machine learning models may be used. As long as it can be used in the AI ​​calculations performed by the machine learning device 10 to obtain a desired output value for any input value, it is not limited to the neural network structure shown in Figure 2, but machine learning models with various structures can be used in the machine learning device 10.

[0051] Next, we will explain the training method for the machine learning model used in the machine learning device 10. As mentioned above, the machine learning device 10 utilizes a pre-trained machine learning model. To create this model, for example, the machine learning model is trained using a test motor having characteristics equivalent to the motor M in Figure 1, before the power conversion device 1 is put into operation.

[0052] Figure 4 shows an example of a training data acquisition system used to acquire training data for a machine learning model. In training the machine learning model used in the machine learning apparatus 10 of this embodiment, for example, training data acquired by the training data acquisition system 13 shown in Figure 4, which is configured by arranging the test device 13a and the load device 13b on a common base 13c, is used.

[0053] The test device 13a comprises a vector control controller 13a1, a test motor 13a2 having characteristics equivalent to the motor M in Figure 1 and driven and controlled by the vector control controller 13a1, and an encoder 13a3 attached to the test motor 13a2 for detecting the phase θ of the test motor 13a2. The vector control controller 13a1 receives a rotation frequency command value ω from an external source. r * The phase θ of the test motor 13a2 detected by the encoder 13a3, and the detected value i of the U-phase current among the three-phase AC currents flowing through the test motor 13a2. uc and the detected value i of the W-phase current wc Based on this, a well-known vector control calculation is performed to drive and control the test motor 13a2.

[0054] The load device 13b comprises a load controller 13b1, a load motor 13b2 which is mechanically connected to the output shaft of the test motor 13a2 via a coupling 13d and applies load torque to the test motor 13a2, and an encoder 13a3 attached to the load motor 13b2 which detects the drive state (e.g., torque) of the load motor 13b2. The load controller 13b1 receives a torque command value τ from an external source. *Based on the drive state of the load motor 13b2 detected by the encoder 13b3, the load motor 13b2 is controlled to drive.

[0055] In the learning data acquisition system 13, the test device 13a is operated in speed control mode, and the load device 13b is operated in torque control mode. The rotational frequency command value ω is then given to the test device 13a. r * The torque command value τ applied to the load device 13b at each rotational frequency is changed in increments of a few percent from stop to base frequency. * By changing this, the test motor 13a2 is driven stably. From the internal signals of the vector control controller 13a1 in this state, time-series data of input signals corresponding to each input value used in the machine learning model 100 in Figure 2, and time-series data of training signals corresponding to each output value output from the machine learning model 100 are acquired as training data.

[0056] Figure 5 shows an example of training data acquired by the training data acquisition system shown in Figure 4. The training data 14 shown in Figure 5 consists of time-series data 141 of the input signal and time-series data 142 of the teacher signal. In the time-series data 141 of the input signal, the six types of input values ​​(d-axis voltage command value v) shown in Figure 2 are used as combinations of input signals for the training data 14. dc_IM ** q-axis voltage command value v qc_IM ** d-axis current detection value i dc q-axis current detection value i qc , slip frequency command value ω s * and the first rotation frequency estimate ω r ^ The following shows an example of time-series data of the training signal 142. In the training signal time-series data 142, the combination of training signals used in the training data 14 is the two types of output values ​​(slip frequency estimate ω) shown in Figure 2. s ^^ and the estimated second rotation frequency ω r ^^ ) corresponding slip frequency detection value ω s_detand rotation frequency detection value ω r_det An example of time-series data is shown.

[0057] The above rotational frequency detection value ω r_det The phase θ detected by encoder 13a3 is obtained by the following equation (10). Also, the slip frequency detected value ω s_det The primary frequency command value ω1 * and rotation frequency detection value ω r_det Using this, it can be calculated by the following equation (11).

[0058]

number

number

[0059] Figure 6 is a waveform diagram graphing the time-series data 141 of the input signal and the time-series data 142 of the teacher signal in the training data 14 shown in Figure 5. In Figure 6, graph 1311 shows the d-axis voltage command value v dc_IM ** Graph 1312 shows the q-axis voltage command value v qc_IM ** Graph 1313 shows the d-axis current detection value i dc Graph 1314 shows the q-axis current detection value i qc Graph 1315 shows the slip frequency command value ω s * Graph 1316 shows the first rotation frequency estimate ω r ^ The values ​​of the time-series data 141 acquired for each input signal are shown. Graph 1321 shows the slip frequency detection value ω s_det Graph 1322 shows the rotational frequency detected value ω r_det The values ​​of the time-series data 143 obtained for each training signal are shown.

[0060] Figure 7 shows the training process of a machine learning model. In training a machine learning model, first, combinations of multiple types of training data 14 acquired under various conditions are input to the deep learning device 15 as input data. The deep learning device 15 is, for example, a personal computer, a supercomputer, or a workstation, and is equipped with software to perform deep learning. Alternatively, a microprocessor with deep learning capabilities may be used as the deep learning device 15.

[0061] The deep learning device 15 detects the slip frequency ω represented by the training signal of the input data set. s_det and rotation frequency detection value ω r_det Based on the time-series data, the slip frequency estimate ω is output from the machine learning model 100. s ^^ Slip frequency detected value ω s_det The error with, and the estimated second rotation frequency ω r ^^ and rotation frequency detection value ω r_det The machine learning model 100 is trained to minimize the error between each of the following. Here, for the time series data of each input signal in the input data set, the weight values ​​W1 to W1 of the weight information 1041 explained in Figure 3 are obtained by forward and backward propagation in the hidden layer 102 of the machine learning model 100. j The bias value b of the bias information 1042 is updated accordingly. This allows a trained machine learning model 100 to be obtained using the time-series data of each input signal and output signal in the input data set.

[0062] The trained machine learning model 100 obtained by the deep learning device 15 is sent to the machine learning device 10 in the power converter 1 shown in Figure 1, and used in the AI ​​calculations performed by the machine learning device 10. As a result, the machine learning device 10 uses the trained machine learning model 100 to determine the d-axis voltage command value v dc_IM ** q-axis voltage command value v qc_IM ** d-axis current detection value i dc q-axis current detection value i qc , slip frequency command value ω s* and the first rotation frequency estimate ω r ^ From each input value, the slip frequency estimate ω s ^^ and the estimated second rotation frequency ω r ^^ It becomes possible to obtain each output value.

[0063] Figure 8 is an explanatory diagram illustrating the improvement effect of the control characteristics according to the present invention. Figure 8(a) shows an example of control characteristics when the motor M is driven and controlled without applying the present invention, i.e., without using the machine learning device 10. On the other hand, Figure 8(b) shows an example of control characteristics when the motor M is driven and controlled using the machine learning device 10, i.e., when applying the present invention. In these figures, the upper graph shows the slip frequency ω of the motor M. s The graphs in the lower section show the time evolution of each element, and each graph represents the rotational frequency ω of the motor M. r and the primary frequency command value ω1 * These graphs show the time evolution of each variable. Note that in these graphs, the values ​​of each variable are normalized by dividing them by 2π.

[0064] If the present invention is not applied, the primary frequency command value ω1 shown in Figure 8(a) * In the aforementioned equation (11), ω s_det =ω s , ω r_det =ω r By replacing each of these, the slip frequency ω s and rotation frequency ω r It is calculated as the sum of these values.

[0065] If this invention is not applied, in principle of velocity estimation, ω1 * =0Hz corresponds to a singularity, and it is known that the speed estimation operation of the motor M breaks down at this singularity. Therefore, conventionally, as shown in Figure 8(a), the primary frequency command value ω1 * The lower limit is a predetermined value greater than 0, for example, ω1 * Set to =0.3Hz, slip frequency ω s When reducing the primary frequency command value ω1* The rotation frequency command value ω should not fall below this lower limit. r * It is common practice to automatically adjust in the increasing direction. As a result, if the present invention is not applied, the slip frequency ω s When the rotation frequency ω decreases r It can be seen that the voltage rises, making it impossible to control the rotation state of motor M to a constant level.

[0066] On the other hand, when the present invention is applied, the machine learning device 10 determines the slip frequency ω s The estimated slip frequency ω is an estimated value of s ^^ And, rotation frequency ω r The estimated second rotation frequency ω is an estimated value of r ^^ Therefore, the primary frequency command value ω1 can be determined. * This is outside the constraints of the singularity mentioned above, and as shown in Figure 8(b), ω1 * Even if the frequency becomes 0Hz, the motor M's speed estimation operation will not fail, and stable drive control of the motor M will be possible.

[0067] According to the first embodiment of the present invention described above, the following effects are achieved.

[0068] (1) Power converter 1 has a phase estimate value θ for the phase of motor M. ^ A phase calculation unit 11 that performs calculations, and the phase estimate value θ calculated by the phase calculation unit 11. ^ Based on this, the alternating current i flows through the motor M. u ,i v ,i w AC current detection value i representing the current value uc ,i vc ,i wc The d-axis current detection value i is the current value in the rotating coordinate system. dc and q-axis current detection value i qc A coordinate transformation unit 5 converts to a d-axis current detection value i dc and q-axis current detection value i qc Based on this, the d-axis voltage command value v for motor M dc_IM** and q-axis voltage command value v qc_IM ** A vector control calculation unit 8 that calculates the d-axis voltage command value v dc_IM ** and q-axis voltage command value v qc_IM ** The corresponding voltage command value v u * ,v v * ,v w * Based on the alternating current i u ,i v ,i w The power converter 2 generates a power converter v, and the machine learning device 10 comprises a power converter 2 that generates a power converter v. In this power converter 1, the machine learning device 10 has a trained machine learning model 100 and the d-axis voltage command value v dc_IM ** q-axis voltage command value v qc_IM ** d-axis current detection value i dc and q-axis current detection value i qc Input values ​​including at least one of the above are input to the machine learning model 100 to obtain the slip frequency estimate ω, which is an output value related to the phase of the motor M. s ^^ and the estimated second rotation frequency ω r ^^ The phase calculation unit 11 outputs the following values: Based on these output values, it calculates the phase estimate θ according to equations (7) and (8). ^ This is calculated. In this way, machine learning can be applied to the sensorless control of the motor M performed by the power converter 1, thereby achieving stable and highly accurate control characteristics.

[0069] (2) In this embodiment, the motor M is an induction motor. In this case, the power converter 1 controls the q-axis voltage command value v qc_IM ** and q-axis current detection value i qc Using this, according to equation (6), the slip frequency command value ω represents the command value of the slip frequency of motor M. s * And the first rotational frequency estimate ω represents the estimated rotational frequency of motor M. r ^It includes a frequency calculation unit 9 that calculates and . The input value of the machine learning model 100 is the d-axis voltage command value v dc_IM ** q-axis voltage command value v qc_IM ** d-axis current detection value i dc and q-axis current detection value i qc At least one of the above, and slip frequency command value ω s * and the first rotation frequency estimate ω r ^ and include. The machine learning device 10 uses these input values ​​to generate a slip frequency estimate ω, which represents an estimate of the slip frequency. s ^^ And the first rotation frequency estimate ω r ^ Corrected second rotation frequency estimate ω r ^^ The output value includes and . In this way, machine learning can be applied to the speed sensorless control of motor M, which is an induction motor, to achieve stable and highly accurate control characteristics.

[0070] (3) The machine learning model 100 used in the machine learning device 10 is the slip frequency detection value ω when the test motor 13a2, which is an induction motor having characteristics equivalent to motor M, is in a stable driving state. s_det and rotation frequency detection value ω r_det This signal is used as the training signal and has been trained by the deep learning device 15. In this way, the slip frequency estimate ω using the machine learning model 100 in the machine learning device 10 is s ^^ and the estimated second rotation frequency ω r ^^ This enables accurate and highly precise AI calculations.

[0071] <Second Embodiment> Figure 9 is a configuration diagram of a power conversion device according to a second embodiment of the present invention. In the first embodiment described above, an example was given in which AI calculations using a machine learning model were applied to speed estimation of motor M, which is an induction motor, in the speed sensorless vector control. In contrast, this embodiment describes an example in which AI calculations using a machine learning model were applied to phase error estimation of motor Ma, which is a permanent magnet synchronous motor, in the position sensorless vector control.

[0072] The power converter 1a shown in Figure 9 is connected to a motor Ma, which is a permanent magnet synchronous motor, and controls the drive of motor Ma by controlling the AC power supplied to motor Ma. The power converter 1a includes a power converter 2, a coordinate transformation unit 5 and a coordinate transformation unit 12, similar to those described in Figure 1 in the first embodiment, as well as a speed control calculation unit 6a, a d-axis current setting unit 7a, a vector control calculation unit 8a, a phase error calculation unit 9a, a machine learning device 10a and a phase calculation unit 11a. However, the coordinate transformation unit 5 uses the phase estimate value θ of motor M in the first embodiment. ^ Instead, the phase estimate value θ of the motor Ma is calculated by the phase calculation unit 11a. a ^ Using this, the 3-phase AC current detection value i input from the current detector 4 is used. uc ,i vc ,i wc The d-axis current detection value i represents the current value in the rotating coordinate system. dc and q-axis current detection value i qc It is converted to the following. In addition, the coordinate transformation unit 12 converts the phase estimate value θ of the motor M in the first embodiment. ^ Instead, the phase estimate value θ of the motor Ma is calculated by the phase calculation unit 11a. a ^ The d-axis voltage command value v output from the vector control calculation unit 8a is used with this. dc_PM ** and q-axis voltage command value v qc_PM ** The voltage command value v for 3-phase AC u * ,v v * ,v w *This converts the signal to [the specified value]. Note that each component except for power converter 2 is implemented using, for example, the software of a microcontroller mounted on power converter 1a.

[0073] In the following, the power converter 2, coordinate transformation unit 5, and coordinate transformation unit 12, which are common components of the power converter 1a in the first embodiment, will not be described, and the other components will be described separately.

[0074] The speed control calculation unit 6a receives the rotation frequency command value ω from the higher-level device to the power converter 1a. r * and the rotation frequency estimate ω output from the phase calculation unit 11a a ^ Based on the difference, the torque command value τ for motor Ma is calculated. a * The calculation is performed, and the q-axis current command value i is obtained from the result of that calculation. q * The speed control calculation unit 6a calculates and outputs the rotation frequency command value ω. r * The rotation frequency estimate ω a ^ To ensure that it follows, proportional control and integral control are used to generate the torque command τ according to equation (12) below. a * and q-axis current command value i q * Perform the calculation.

[0075]

number

[0076] Of the constants in equation (12), proportional gain K sp , integral gain K si and polar logarithm P m The definitions of each constant are the same as those in equation (1) described in the first embodiment. The definitions of the other constants are as follows: K e : Induced voltage coefficient L dd-axis inductance L q q-axis inductance

[0077] The d-axis current setting unit 7a sets the pre-set d-axis current command value i d * Outputs.

[0078] The vector control calculation unit 8a calculates the current command values ​​i for the d axis and q axis. d * ,i q * And the current detection value i in the rotating coordinate system dc ,i qc And the rotation frequency estimate ω calculated by the phase calculation unit 11a a ^ Based on this, the voltage command values ​​for the d-axis and q-axis of motor Ma, which is a permanent magnet synchronous motor, are v dc_PM ** ,v qc_PM ** The system calculates the following and outputs the results. At this time, the vector control calculation unit 8a uses, for example, the electrical circuit parameters of the motor Ma to calculate the d-axis voltage command value v dc_PM ** and q-axis voltage command value v qc_PM ** It is possible to perform the following calculations.

[0079] Specifically, the vector control calculation unit 8a first determines the set value R of the winding resistance, which is an electrical circuit parameter of the motor Ma. * And the set values ​​L for the d-axis and q-axis inductances in the windings of motor Ma. d * ,L q * And the set value K of the induced voltage coefficient e * And the current command values ​​i for the d-axis and q-axis of the rotating coordinate system d * ,i q * And, the estimated rotation frequency ω a ^ Using the above, the voltage reference values ​​v for the d axis and q axis are calculated according to equation (13) below. dc_PM * ,vqc_PM * Perform the calculation.

[0080]

number

[0081] Note that in equation (13), T acr This, like equation (2) above, represents the response time constant of the current control.

[0082] Next, the current detection values ​​i for the d axis and q axis dc ,i qc However, the current command values ​​i for the d axis and q axis d * ,i q * To track these changes, proportional and integral control are used to calculate the voltage correction values ​​Δv for the d-axis and q-axis according to equation (14) below. dc_PM Δv qc_PM Perform the calculation.

[0083]

number

[0084] The definitions of each constant in equation (14) are as follows: K pd_PM : Proportional gain of current control on the d axis K id_PM : Integral gain of current control on the d axis K pq_PM : Proportional gain of current control on the q axis K iq_PM Integral gain of current control on the q axis

[0085] Based on the above calculation, the voltage reference values ​​v for the d axis and q axis are obtained. dc_PM * ,v qc_PM * And the voltage correction value Δv for the d axis and q axis. dc_PM Δv qc_PMOnce these values ​​are obtained, they are added together according to equation (15) below to obtain the voltage command values ​​v for the d axis and q axis. dc_PM ** ,v qc_PM ** Perform the calculation.

[0086]

number

[0087] The phase error calculation unit 9a calculates the voltage command values ​​v for the d axis and q axis. dc_PM ** ,v qc_PM ** and current detection value i dc ,i qc And, the estimated rotation frequency ω a ^ Using this, the first phase error estimate Δθ represents the estimated error between the control phase used to control motor Ma and the actual phase of motor Ma. a ^ The calculation is performed and output. At this time, the phase error calculation unit 9a uses, for example, the electrical circuit parameters of the motor Ma to calculate the first phase error estimate Δθ. a ^ It is possible to perform the following calculations.

[0088] Specifically, the phase error calculation unit 9a calculates the voltage command values ​​v for the d axis and q axis. dc_PM ** ,v qc_PM ** and current detection value i dc ,i qc And the set values ​​R for the winding resistance and q-axis inductance of motor Ma. * ,L q * Using the above, the first phase error estimate Δθ is calculated according to the following equation (16), which is an extended induced voltage equation. a ^ Perform the calculation.

[0089]

number

[0090] The machine learning device 10a, similar to the machine learning device 10 described in the first embodiment, is a device that performs AI calculations using a trained machine learning model. It inputs predetermined input values ​​corresponding to the driving state of motor Ma into the machine learning model and outputs output values ​​related to the phase of motor Ma. In this embodiment, for example, the d-axis voltage command value v is used as an input value to the machine learning model. dc_PM ** and q-axis voltage command value v qc_PM ** And the current detection value i in the rotating coordinate system dc ,i qc And the first phase error estimate Δθ obtained by the phase error calculation unit 9a a ^ And the rotational frequency estimated value ω obtained by the phase calculation unit 11a. a ^ The following are used. The machine learning device 10a stores a pre-trained machine learning model based on the relationship between these values ​​and the phase error of the motor Ma. Using this machine learning model, the machine learning device 10a obtains the first phase error estimate Δθ as an output value related to the phase of the motor Ma. a ^ Corrected second phase error estimate Δθ a ^^ It can output.

[0091] The phase calculation unit 11a calculates the phase estimate θ for the phase of the motor Ma. a ^ The following is calculated. In this embodiment, the phase calculation unit 11a calculates the second phase error estimate Δθ obtained by the machine learning device 10a. a ^^ Based on this, the phase estimate θ a ^ And the estimated rotational frequency ω of motor Ma a ^ The program calculates and outputs the results.

[0092] Specifically, the phase calculation unit 11a calculates the second phase error estimate Δθ, which is the output of the machine learning device 10a. a ^^ The predetermined phase error command value Δθ c* (For example, Δθ) c * To track (=0), proportional and integral control are used to estimate the rotation frequency ω according to equation (17) below. a ^ The following is calculated. Also, according to equation (18), the phase estimate θ a ^ Perform the calculation.

[0093]

number

number

[0094] The basic operation of the position sensorless vector control method in the power converter 1a of this embodiment is as described above. Next, the details of the machine learning device 10a, which is a feature of this embodiment, will be described.

[0095] Figure 10 is a functional block diagram of a machine learning model used in a second embodiment of the present invention. The machine learning device 10a of this embodiment has a pre-trained machine learning model 100a. The machine learning model 100a is a deep learning model for AI computation using, for example, a neural network, similar to the machine learning model 100 described in the first embodiment, and it calculates and outputs output values ​​for a predetermined combination of input values ​​using AI computation. Specifically, as shown in Figure 10, for example, the machine learning model 100a is configured by combining an input layer 101a, a hidden layer 102a, and an output layer 103a.

[0096] Multiple types of input values ​​for AI calculation are input to the input layer 101a. In this embodiment, for example, the six types of input values ​​shown in Figure 10 (d-axis voltage command value v dc_PM ** q-axis voltage command value v qc_PM ** d-axis current detection value i dc q-axis current detection value i qc , First phase error estimate Δθ a ^and rotation frequency estimate ω a ^ The combination of the above is input to the input layer 101a. Note that it is not necessary to input all six of these input values ​​to the input layer 101a. Alternatively, other variables, such as the active power value of the rotating coordinate system (i qc ×v qc_PM ** ) and reactive power value (i dc ×v dc_PM ** An input value further including at least one of the above may be input to the input layer 101a.

[0097] The hidden layer 102a is composed of multiple neurons 104a arranged hierarchically, and performs a predetermined weighting operation on each input value input to the input layer 101a, according to the hierarchy. By repeating this process, forward and backward propagation characteristics are multiplied to each input value within the hidden layer 102a. The result of the operation at the final layer of the hidden layer 102a is input to the output layer 103a. The configuration of each neuron 104a in the hidden layer 102a is the same as the configuration of each neuron 104 described in Figure 3 in the first embodiment.

[0098] The output layer 103a generates a predetermined output value from the calculation result of the final layer of the hidden layer 102a. In this embodiment, for example, one type of output value (second phase error estimate Δθ) is shown in Figure 10. a ^^ The output value generated by the output layer 103a is output from the machine learning model 100a as the output value of the machine learning device 10a.

[0099] Note that the machine learning model 100a described in Figure 10 is just one example, and other machine learning models may be used. As long as the machine learning device 10a can obtain a desired output value for any input value by being used in the AI ​​calculations it performs, it is not limited to the neural network structure shown in Figure 10, but various machine learning models with different structures can be used in the machine learning device 10a.

[0100] Next, the learning method for the machine learning model used in the machine learning device 10a will be described. Similar to the first embodiment, the machine learning device 10a utilizes a pre-trained machine learning model 100a. To create this model, the machine learning model 100a is trained before the power conversion device 1a is put into operation, using, for example, a test motor having characteristics equivalent to the motor Ma shown in Figure 9.

[0101] In this embodiment, as in the first embodiment, the training data acquired when the test motor 13a2 is stably driven in the training data acquisition system 13 shown in Figure 4 can be used to train the machine learning model 100a. However, the test motor 13a2 in the test device 13a has characteristics equivalent to the motor Ma shown in Figure 9. The training signal is the second phase error estimate Δθ, which is the output value of the machine learning model 100a. a ^^ The detected phase error corresponding to Δθ a This is used.

[0102] Phase error detection value Δθ a This is the phase estimate θ a ^ and the phase θ of encoder 13a3 a The detected value is used to obtain the following equation (19).

[0103]

number

[0104] The machine learning model 100a is trained using the training data acquired as described above. The trained machine learning model 100a is sent to the machine learning device 10a in the power converter 1a shown in Figure 9, and used in the AI ​​calculations performed by the machine learning device 10a. As a result, the machine learning device 10a uses the trained machine learning model 100a to determine the d-axis voltage command value v dc_PM ** q-axis voltage command value v qc_PM ** d-axis current detection value i dc q-axis current detection value iqc , First phase error estimate Δθ a ^ and rotation frequency estimate ω a ^ From each input value, the output value is the second phase error estimate Δθ a ^^ It becomes possible to calculate this.

[0105] According to the second embodiment of the present invention described above, the following effects are achieved.

[0106] (1) The power converter 1a has a phase estimate value θ for the phase of the motor Ma. a ^ A phase calculation unit 11a that performs calculations, and the phase estimate value θ calculated by the phase calculation unit 11a. a ^ Based on this, the alternating current i flowing through motor Ma u ,i v ,i w AC current detection value i representing the current value uc ,i vc ,i wc The d-axis current detection value i is the current value in the rotating coordinate system. dc and q-axis current detection value i qc A coordinate transformation unit 5 converts to a d-axis current detection value i dc and q-axis current detection value i qc Based on this, the d-axis voltage command value v for motor Ma dc_PM ** and q-axis voltage command value v qc_PM ** A vector control calculation unit 8a that calculates the d-axis voltage command value v dc_PM ** and q-axis voltage command value v qc_PM ** The corresponding voltage command value v u * ,v v * ,v w * Based on the alternating current i u ,i v ,i wThe system comprises a power converter 2 that generates a d-axis voltage command value v, and a machine learning device 10a. In this power converter 1a, the machine learning device 10a has a trained machine learning model 100a and dc_PM ** q-axis voltage command value v qc_PM ** d-axis current detection value i dc and q-axis current detection value i qc Input values ​​including at least one of the above are input to the machine learning model 100a to obtain a second phase error estimate Δθ, which is an output value related to the phase of the motor Ma. a ^^ The phase calculation unit 11a outputs the following based on this output value, according to equations (17) and (18): a ^ This is calculated. In this way, as in the first embodiment, machine learning can be applied to the sensorless control of motor Ma performed by the power converter 1a to achieve stable and highly accurate control characteristics.

[0107] (2) In this embodiment, the motor Ma is a permanent magnet synchronous motor. In this case, the power converter 1a receives the voltage command values ​​v for the d axis and q axis. dc_PM ** ,v qc_PM ** and current detection value i dc ,i qc Using this, according to equation (16), the first phase error estimate Δθ represents the estimated error between the control phase used to control motor Ma and the actual phase of motor Ma. a ^ It includes a phase error calculation unit 9a that calculates the following. The input value of the machine learning model 100a is the d-axis voltage command value v dc_IM ** q-axis voltage command value v qc_IM ** d-axis current detection value i dc and q-axis current detection value i qc At least one of the above, and the first phase error estimate Δθ a ^ The machine learning device 10a calculates the first phase error estimate Δθ for these input values. a ^Corrected second phase error estimate Δθ a ^^ The output value includes [the specified value]. In this way, machine learning can be applied to the position sensorless control of the synchronous motor Ma, thereby achieving stable and highly accurate control characteristics.

[0108] (3) The machine learning model 100a used in the machine learning device 10a is the phase error detection value Δθ when the test motor 13a2, which is a synchronous motor having characteristics equivalent to motor Ma, is in a stable driving state. a This has been used as the training signal for training. Therefore, the second phase error estimate Δθ using the machine learning model 100a in the machine learning device 10a is obtained. a ^^ This enables accurate and highly precise AI calculations.

[0109] <Third Embodiment> Figure 11 is a configuration diagram of a power conversion device according to a third embodiment of the present invention. In the second embodiment described above, an example was given in which AI calculations using a machine learning model were applied to the phase error estimation of motor Ma, which is a permanent magnet synchronous motor, using position sensorless vector control. In contrast, this embodiment describes an example in which AI calculations using a machine learning model were applied to the phase error estimation of motor Mb, which is a synchronous reluctance motor, using position sensorless vector control.

[0110] The power converter 1b shown in Figure 11 is connected to a motor Mb, which is a synchronous reluctance motor, and controls the drive of the motor Mb by controlling the AC power supplied to the motor Mb. The power converter 1b includes a power converter 2, a coordinate transformation unit 5 and a coordinate transformation unit 12, similar to those described in Figure 1 in the first embodiment, as well as a speed control calculation unit 6b, a vector control calculation unit 8b, a phase error calculation unit 9b, a machine learning device 10b and a phase calculation unit 11b. However, the coordinate transformation unit 5 uses the phase estimate value θ of the motor M in the first embodiment. ^ Instead, the phase estimate value θ of the motor Mb is calculated by the phase calculation unit 11b. b ^Using this, the 3-phase AC current detection value i input from the current detector 4 is used. uc ,i vc ,i wc The d-axis current detection value i represents the current value in the rotating coordinate system. dc and q-axis current detection value i qc It is converted to the following. In addition, the coordinate transformation unit 12 converts the phase estimate value θ of the motor M in the first embodiment. ^ Instead, the phase estimate value θ of the motor Mb is calculated by the phase calculation unit 11b. b ^ The d-axis voltage command value v output from the vector control calculation unit 8b is used with this. dc_SYC ** and q-axis voltage command value v qc_SYC ** The voltage command value v for 3-phase AC u * ,v v * ,v w * This converts the signal to [the specified value]. Note that each component except for power converter 2 is implemented using, for example, the software of the microcontroller mounted on power converter 1b.

[0111] In the following, the power converter 2, coordinate transformation unit 5, and coordinate transformation unit 12, which are common components of the power converter 1b in the first embodiment, will not be described, and the other components will be described separately.

[0112] The speed control calculation unit 6b receives the rotation frequency command value ω from the higher-level device to the power converter 1b. r * and the rotation frequency estimate ω output from the phase calculation unit 11b b ^ Based on the difference, the torque command value τ for motor Mb is calculated. b * The current command values ​​i for the d axis and q axis are calculated from the calculation result. d * ,i q * The speed control calculation unit 6b calculates and outputs the rotation frequency command value ω. r * The rotation frequency estimate ωb ^ To ensure that it follows, proportional control and integral control are used to generate the torque command τ according to equation (20) below. * and d-axis current command value i d * Perform the calculation.

[0113]

number

[0114] Furthermore, according to the following equation (21), the q-axis current command value i q * Perform the calculation.

[0115]

number

[0116] In equation (21), T d This is the first-order lag time constant.

[0117] The vector control calculation unit 8b calculates the current command values ​​i for the d axis and q axis. d * ,i q * And the current detection value i in the rotating coordinate system dc ,i qc And the rotation frequency estimate ω calculated by the phase calculation unit 11b b ^ Based on this, the voltage command values ​​for the d-axis and q-axis of the motor Mb, which is a synchronous reluctance motor, are v dc_SYC ** ,v qc_SYC ** The vector control calculation unit 8b calculates the d-axis voltage command value v, for example, using the electrical circuit parameters of the motor Mb, similar to the vector control calculation unit 8a described in the second embodiment. dc_SYC ** and q-axis voltage command value v qc_SYC ** It is possible to perform the following calculations.

[0118] Specifically, the vector control calculation unit 8b first determines the set value R of the winding resistance, which is an electrical circuit parameter of the motor Mb. * And the set values ​​L for the d-axis and q-axis inductances in the motor Mb windings. d * ,L q * And the current command values ​​i for the d-axis and q-axis of the rotating coordinate system d * ,i q * And, the estimated rotation frequency ω b ^ Using the above, the voltage reference values ​​v for the d axis and q axis are calculated according to equation (22) below. dc_SYC * ,v qc_SYC * Perform the calculation.

[0119]

number

[0120] Note that in equation (22), T acr This, like equations (2) and (13) mentioned above, represents the response time constant of the current control.

[0121] Next, the current detection values ​​i for the d axis and q axis dc ,i qc However, the current command values ​​i for the d axis and q axis d * ,i q * To track these changes, proportional and integral control are used to calculate the voltage correction values ​​Δv for the d-axis and q-axis according to equation (23) below. dc_SYC Δv qc_SYC Perform the calculation.

[0122]

number

[0123] The definitions of each constant in equation (23) are as follows: K pd_SYC : Proportional gain of current control on the d axis K id_SYC : Integral gain of current control on the d axis K pq_SYC : Proportional gain of current control on the q axis K iq_SYC Integral gain of current control on the q axis

[0124] Based on the above calculation, the voltage reference values ​​v for the d axis and q axis are obtained. dc_SYC * ,v qc_SYC * And the voltage correction value Δv for the d axis and q axis. dc_SYC Δv qc_SYC Once these values ​​are obtained, they are added together according to equation (24) below to obtain the voltage command values ​​v for the d axis and q axis. dc_SYC ** ,v qc_SYC ** Perform the calculation.

[0125]

number

[0126] The phase error calculation unit 9b calculates the voltage command values ​​v for the d axis and q axis. dc_SYC ** ,v qc_SYC ** and current detection value i dc ,i qc And, the estimated rotation frequency ω b ^ Using these, the first phase error estimate Δθ represents the estimated error between the control phase used to control the motor Mb and the actual phase of the motor Mb. b ^ The calculation and output are performed. At this time, the phase error calculation unit 9b uses, for example, the electrical circuit parameters of the motor Mb to calculate the first phase error estimate Δθ. b ^ It is possible to perform the following calculations.

[0127] Specifically, the phase error calculation unit 9b calculates the voltage command values ​​v for the d axis and q axis. dc_SYC ** ,v qc_SYC ** and current detection value idc ,i qc And the set values ​​R for the winding resistance of motor Mb and the d-axis inductance. * ,L d * Using the above, the first phase error estimate Δθ is calculated according to the following equation (25), which is an extended induced voltage equation. b ^ Perform the calculation.

[0128]

number

[0129] The machine learning device 10b, like the machine learning devices 10 and 10a described in the first and second embodiments respectively, is a device that performs AI calculations using a trained machine learning model. It inputs predetermined input values ​​corresponding to the driving state of the motor Mb into the machine learning model and outputs output values ​​related to the phase of the motor Mb. In this embodiment, for example, the d-axis voltage command value v is used as an input value to the machine learning model. dc_SYC ** and q-axis voltage command value v qc_SYC ** And the current detection value i in the rotating coordinate system dc ,i qc And the first phase error estimate Δθ obtained by the phase error calculation unit 9b b ^ And the rotational frequency estimate ω obtained by the phase calculation unit 11b b ^ The following are used. The machine learning device 10b stores a pre-trained machine learning model based on the relationship between these values ​​and the phase error of the motor Mb. Using this machine learning model, the machine learning device 10b outputs the first phase error estimate Δθ as an output value related to the phase of the motor Mb. b ^ Corrected second phase error estimate Δθ b ^^ It can output.

[0130] The phase calculation unit 11b calculates the phase estimate θ for the phase of the motor Mb. b ^The following is calculated. In this embodiment, the phase calculation unit 11b calculates the second phase error estimate Δθ obtained by the machine learning device 10b. b ^^ Based on this, the phase estimate θ b ^ And the estimated rotational frequency ω of motor Mb b ^ The program calculates and outputs the results.

[0131] Specifically, the phase calculation unit 11b processes the second phase error estimate Δθ, which is the output of the machine learning device 10b. b ^^ The predetermined phase error command value Δθ c * (For example, Δθ) c * To track (=0), proportional and integral control are used to estimate the rotation frequency ω according to equation (26) below. b ^ The following is calculated. Also, according to equation (27), the phase estimate θ b ^ Perform the calculation.

[0132]

number

number

[0133] The basic operation of the position sensorless vector control method in the power converter 1b of this embodiment is as described above. Next, the details of the machine learning device 10b, which is a feature of this embodiment, will be described.

[0134] Figure 12 is a functional block diagram of a machine learning model used in a third embodiment of the present invention. The machine learning device 10b of this embodiment has a pre-trained machine learning model 100b. The machine learning model 100b is a deep learning model for AI computation using, for example, a neural network, similar to the machine learning models 100 and 100a described in the first and second embodiments, and it calculates and outputs output values ​​for a predetermined combination of input values ​​using AI computation. Specifically, as shown in Figure 12, for example, the machine learning model 100b is configured by combining an input layer 101b, a hidden layer 102b, and an output layer 103b.

[0135] Multiple types of input values ​​for AI calculation are input to the input layer 101b. In this embodiment, for example, the six types of input values ​​shown in Figure 12 (d-axis voltage command value v dc_SYC ** q-axis voltage command value v qc_SYC ** d-axis current detection value i dc q-axis current detection value i qc , First phase error estimate Δθ b ^ and rotation frequency estimate ω b ^ The combination of the above is input to the input layer 101b. Note that it is not necessary to input all six of these input values ​​to the input layer 101b. Alternatively, other variables, such as the active power value of the rotating coordinate system (i qc ×v qc_SYC ** ) and reactive power value (i dc ×v dc_SYC ** An input value further including at least one of the above may be input to the input layer 101b.

[0136] The hidden layer 102b is constructed by hierarchically combining multiple neurons 104b, and performs a predetermined weighting operation on each input value input to the input layer 101b, according to the hierarchy. By repeating this process, forward and backward propagation characteristics are multiplied to each input value within the hidden layer 102b. The result of the operation at the final layer of the hidden layer 102b is input to the output layer 103b. The configuration of each neuron 104b in the hidden layer 102b is the same as the configuration of each neuron 104 described in Figure 3 in the first embodiment.

[0137] The output layer 103b generates a predetermined output value from the calculation result of the final layer of the hidden layer 102b. In this embodiment, for example, one type of output value (second phase error estimate Δθ) is shown in Figure 12. b ^^ The output value generated by the output layer 103b is output from the machine learning model 100b as the output value of the machine learning device 10b.

[0138] Note that the machine learning model 100b described in Figure 12 is just one example, and other machine learning models may be used. As long as the machine learning device 10b can use any machine learning model with various structures, not just the neural network structure shown in Figure 12, to obtain a desired output value for any given input value, the machine learning device 10b can utilize such models.

[0139] Next, the learning method for the machine learning model used in the machine learning device 10b will be described. Similar to the first and second embodiments, the machine learning device 10b utilizes a pre-trained machine learning model 100b. To create this model, the machine learning model 100b is trained before the power converter 1b is put into operation, using, for example, a test motor having characteristics equivalent to the motor Mb in Figure 11.

[0140] In this embodiment, as in the first and second embodiments, the training data acquired when the test motor 13a2 is stably driven in the training data acquisition system 13 shown in Figure 4 can be used to train the machine learning model 100b. However, the test motor 13a2 in the test device 13a has characteristics equivalent to the motor Mb shown in Figure 11. The training signal is the second phase error estimate Δθ, which is the output value of the machine learning model 100b. b ^^ The detected phase error corresponding to Δθ b This is used.

[0141] Phase error detection value Δθ b This is the phase estimate θ b ^ and the phase θ of encoder 13a3 b The detected value is used to obtain the following equation (28).

[0142]

number

[0143] The machine learning model 100b is trained using the training data acquired as described above. The trained machine learning model 100b is sent to the machine learning device 10b in the power converter 1b shown in Figure 11 and used in the AI ​​calculations performed by the machine learning device 10b. As a result, the machine learning device 10b uses the trained machine learning model 100b to determine the d-axis voltage command value v dc_SYC ** q-axis voltage command value v qc_SYC ** d-axis current detection value i dc q-axis current detection value i qc , First phase error estimate Δθ b ^ and rotation frequency estimate ω b ^ From each input value, the output value is the second phase error estimate Δθ b ^^ It becomes possible to calculate this.

[0144] According to the third embodiment of the present invention described above, the same effects and advantages as those described in the second embodiment are achieved.

[0145] <Fourth Embodiment> Figure 13 is a configuration diagram of a power converter according to the fourth embodiment of the present invention. In the first to third embodiments described above, examples were explained in which a machine learning device within the power converter performs AI calculations using a machine learning model created in advance by machine learning on a computer. In contrast, this embodiment describes an example in which the machine learning device within the power converter can update the trained machine learning model at any time.

[0146] The power converter 1c shown in Figure 13 is connected to a motor M, which is an induction motor, and controls the drive of the motor M by controlling the AC power supplied to the motor M. The power converter 1c has the same configuration as the power converter 1 shown in Figure 1, which was described in the first embodiment. Note that each component except for the power converter 2 is implemented using, for example, the software of a microcontroller mounted on the power converter 1c.

[0147] In this embodiment, the machine learning device 10 is configured to communicate with an IoT controller 16 located outside the power converter 1c via wired or wireless connection. The IoT controller 16 is, for example, the computer used in the learning data acquisition system 13 shown in Figure 4, as described in the first embodiment. Note that an edge computer, cloud computer, or microprocessor with deep learning capabilities may also be used as the IoT controller 16.

[0148] The IoT controller 16 is used as a vector control controller 13a1 in the training data acquisition system 13 shown in Figure 4, and also functions as a deep learning device 15 as described in Figure 7. From the various input signals and training signals acquired by the training data acquisition system 13, a trained machine learning model is created using a deep learning network and can be transmitted to the power converter 1c at any time.

[0149] When the power converter 1c receives a trained machine learning model from the IoT controller 16, it replaces the machine learning model previously stored in the machine learning device 10 with the newly received model. This makes it possible to update the machine learning model even while the motor M is running.

[0150] <Fifth Embodiment> Figure 14 is a diagram showing the configuration of a power conversion device according to a fifth embodiment of the present invention. In this embodiment, an example is described in which information obtained during motor drive is transmitted from the power conversion device to a higher-level computer, and this information is used as an input signal to train a machine learning model.

[0151] The power converter 1d shown in Figure 14 is connected to a motor M, which is an induction motor, and controls the drive of the motor M by controlling the AC power supplied to the motor M. The power converter 1d has the same configuration as the power converter 1 shown in Figure 1, which was described in the first embodiment. Note that each component except for the power converter 2 is implemented using, for example, the software of a microcontroller mounted on the power converter 1d.

[0152] In this embodiment, as in the fourth embodiment described above, the machine learning device 10 is configured to communicate with the deep learning device 15, which is located outside the power converter 1d, by wired or wireless means. The deep learning device 15 has the same functions as the one shown in Figure 7 in the first embodiment, and in this embodiment it acts as a higher-level device of the power converter 1d.

[0153] In this embodiment, during the drive control of the motor M, an input signal to the machine learning device 10 is transmitted from the power converter 1d to the deep learning device 15. Upon receiving this signal, the deep learning device 15 performs unsupervised learning using the input signal to create a trained machine learning model that utilizes a deep learning network. The created machine learning model is then transmitted to the power converter 1d at any time.

[0154] When the power converter 1d receives a trained machine learning model from the deep learning device 15, it replaces the machine learning model previously stored in the machine learning device 10 with the newly received model. This makes it possible to update the machine learning model even while the motor M is running, similar to the fourth embodiment.

[0155] <Sixth Embodiment> Figure 15 is a configuration diagram of a power conversion device according to the sixth embodiment of the present invention. In this embodiment, an example of a power conversion device is described in which it is possible to select whether or not to use the calculation results obtained by a machine learning device for motor drive control.

[0156] The power converter 1e shown in Figure 15 is connected to a motor M, which is an induction motor, and controls the drive of the motor M by controlling the AC power supplied to the motor M. The power converter 1e includes a power converter 2, a coordinate transformation unit 5, a speed control calculation unit 6, a d-axis current flux setting unit 7, a vector control calculation unit 8, a frequency calculation unit 9, a machine learning device 10, a phase calculation unit 11, and a coordinate transformation unit 12, similar to those described in Figure 1 of the first embodiment, plus an output signal selection unit 17. However, the phase calculation unit 11 uses the slip frequency estimate value ω in the first embodiment. s ^^ and the estimated second rotation frequency ω r ^^ Instead, the selected slip frequency estimate ω is output from the output signal selection unit 17. s ^^^ and selected rotation frequency estimate ω r ^^^ Based on this, the phase estimate θ ^ And the command value ω1 for the primary frequency of motor M * The system calculates and outputs the following. Note that each component except for power converter 2 is implemented using, for example, the software of the microcontroller mounted on power converter 1e.

[0157] The output signal selection unit 17 receives the slip frequency command value ω calculated by the frequency calculation unit 9. s * and the first rotation frequency estimate ω r^ And the slip frequency estimate ω calculated by the machine learning device 10. s ^^ and the estimated second rotation frequency ω r ^^ The following is input. The output signal selection unit 17 selects the slip frequency command value ω s * Or slip frequency estimate ω s ^^ Select one of the following, and select the slip frequency estimate ω s ^^^ This is output to the phase calculation unit 11. Furthermore, the first rotation frequency estimate ω r ^ Alternatively, the estimated second rotation frequency ω r ^^ Select one of the following, and select the rotation frequency estimate ω r ^^^ This is output to the phase calculation unit 11.

[0158] The user of the power converter 1e can arbitrarily select which signal to output to the output signal selection unit 17. For example, it is preferable that the user can arbitrarily set whether to output the calculation result of the frequency calculation unit 9 or the machine learning device 10 to the phase calculation unit 11 by operating the operation panel provided on the power converter 1e or by operating various computers that can communicate with the power converter 1e via wired or wireless means.

[0159] In the fourth to sixth embodiments of the present invention described above, power converters 1c, 1d, and 1e were described, respectively, which are connected to a motor M that is an induction motor and perform drive control of the motor M, similar to the power converter 1 described in the first embodiment. However, the same control is possible with the power converter 1a described in the second embodiment and the power converter 1b described in the third embodiment. That is, a trained machine learning model can be created by an IoT controller 16 or a deep learning device 15 located outside the power converter, which is connected to motors Ma and Mb that are synchronous motors and perform drive control of the motors Ma and Mb, and this can be downloaded to and updated by a machine learning device located inside the power converter. Furthermore, an output signal selection unit 17 is provided inside these power converters, and either the calculation result of the frequency calculation unit 9 or the calculation result of the machine learning device 10 can be output from the output signal selection unit 17 to the phase calculation unit 11 according to the user's selection.

[0160] Furthermore, in the first embodiment of the present invention described above, the vector control calculation unit 8 determines the current detection values ​​i of the d axis and q axis. dc ,i qc The current command values ​​i for the d axis and q axis d * ,i q * The voltage correction values ​​Δv for the d axis and q axis are calculated according to equation (3) to track each of these changes. dc_IM Δv qc_IM The voltage correction value Δv is calculated, and further, according to equation (4), dc_IM Δv qc_IM The voltage reference value v dc_IM * ,v qc_IM * By adding these values, the voltage command values ​​for the d-axis and q-axis are obtained. dc_IM ** , v qc_IM ** An example of performing vector control calculations was described. Furthermore, similar examples of performing vector control calculations were described in each of the second to sixth embodiments. However, a different vector control calculation method was used to control the voltage command values ​​v of the d axis and q axis. dc_IM ** , v qc_IM** You may also try to find it.

[0161] Specifically, for example, the following vector control calculation method can be used. First, according to equation (29) below, the current command values ​​i for the d axis and q axis are calculated. d * ,i q * and current detection value i dc ,i qc From an intermediate current command value i d ** ,i q ** Create.

[0162]

number

[0163] Next, the primary frequency command value ω1 * And using the electrical circuit parameters of motor M, a vector control calculation represented by the following equation (30) is performed to obtain the voltage command values ​​v for the d axis and q axis. dc *** ,v qc *** We seek.

[0164]

number

[0165] The definitions of the constants in equations (29) and (30) are as follows: K pd1 : Proportional gain of current control on the d axis K id1 : Integral gain of current control on the d axis K pq1 : Proportional gain of current control on the q axis K iq1 Integral gain of current control on the q axis T d : Electrical time constant of the d axis (L s / R1) T q : Electrical time constant of the q axis (Ls / R1)

[0166] Alternatively, for example, the following vector control operation method may be used. First, according to the following formula (31), the current command values i d * , i q * and the current detection values i dc , i qc are used to create the voltage correction value Δv d_p * of the proportional operation component of the d-axis, the voltage correction value Δv d_i * of the integral operation component of the d-axis, the voltage correction value Δv q_p * of the proportional operation component of the q-axis, and the voltage correction value Δv q_i * of the integral operation component of the q-axis, respectively.

[0167]

Equation

[0168] Next, using the primary frequency command value ω1 * and the electrical circuit parameters of the motor M, perform the vector control operation represented by the following formula (32) to obtain the voltage command values v dc **** , v qc **** .

[0169]

Equation

[0170] The definitions of the respective constants in formulas (31) and (32) are as follows. K pd2 : Proportional gain of d-axis current control K id2 : Integral gain of d-axis current control K pq2 : Proportional gain of q-axis current control Kiq2 Integral gain of current control on the q axis

[0171] Furthermore, for example, according to equation (33) below, the current command values ​​i for the d axis and q axis are determined. d * ,i q * And, the q-axis current detection value i qc primary delayed signal i qctd And the primary frequency command value ω1 * Then, using the electrical circuit parameters of motor M, vector control calculations are performed to determine the voltage command values ​​v for the d axis and q axis. dc ***** ,v qc ***** You may also request this.

[0172]

number

[0173] In each of the first to sixth embodiments of the present invention described above, for example, silicon (Si) can be used as the material for the semiconductor element constituting the power converter 2. Alternatively, the power converter 2 may be constructed using wide-bandgap semiconductor elements such as silicon carbide (SiC) or gallium nitride (GaN).

[0174] In each of the first to sixth embodiments, the machine learning models used in the machine learning devices 10, 10a, and 10b, respectively, can be, for example, pre-trained deep learning networks. Alternatively, machine learning models that do not apply deep learning may be used.

[0175] The present invention is not limited to the embodiments or modifications described above, and can be implemented using any components without departing from its spirit. Furthermore, each embodiment or modification may be adopted individually, or multiple embodiments may be adopted in any combination. In other words, the present invention can achieve the effects described above by arbitrarily combining the features of each embodiment.

[0176] The embodiments and modifications described above are merely examples, and the present invention is not limited to these, as long as the features of the invention are not impaired. Furthermore, although various embodiments and modifications have been described above, the present invention is not limited to these. Other embodiments conceivable within the scope of the technical idea of ​​the present invention are also included within the scope of the present invention. [Explanation of symbols]

[0177] M, Ma, Mb... motor 1, 1a, 1b, 1c, 1d, 1e... Power conversion device 2…Power converter 3…DC power supply 4…Current detector 5... Coordinate transformation section 6, 6a, 6b... Speed ​​control calculation unit 7...d-axis current and magnetic flux setting section 7a…d-axis current setting section 8, 8a, 8b... Vector control calculation unit 9…Frequency calculation unit 9a, 9b...Phase error calculation section 10, 10a, 10b… Machine learning devices 11,11a,11b...Phase calculation section 12... Coordinate transformation section 13…Training data acquisition system 14…Training data 15…Deep learning device 16…IoT Controller 17…Output signal selection section

Claims

1. A phase calculation unit that calculates a phase estimate for the phase of the electric motor, A coordinate transformation unit converts the current value of the AC current flowing through the motor into a current value in a rotating coordinate system based on the phase estimate calculated by the phase calculation unit, A vector control calculation unit calculates a voltage command value for the electric motor based on the current value of the rotating coordinate system, A power converter that generates the AC current based on the voltage command value, A machine learning device having a trained machine learning model, which takes an input value including at least one of the voltage command value and the current value of the rotating coordinate system as input to the machine learning model and outputs an output value related to the phase, The power conversion device is characterized in that the phase calculation unit calculates the phase estimate value based on the output value.

2. In the power conversion device according to claim 1, The aforementioned motor is an induction motor, The power conversion device includes a frequency calculation unit that uses the voltage command value and the current value of the rotating coordinate system to calculate a slip frequency command value representing the command value of the slip frequency of the induction motor and a first rotation frequency estimate value representing the estimated value of the rotation frequency of the induction motor. The input value includes at least one of the voltage command value and the current value of the rotating coordinate system, the slip frequency command value and the first rotation frequency estimate, The machine learning device outputs an output value that includes, in response to the input value, a slip frequency estimate representing the estimated slip frequency and a second rotation frequency estimate obtained by correcting the first rotation frequency estimate. The power conversion device is characterized in that the phase calculation unit calculates the phase estimate based on the slip frequency estimate and the second rotation frequency estimate.

3. In the power conversion device according to claim 2, The power conversion device is characterized in that the input value further includes at least one of the active power value and the reactive power value of the rotating coordinate system.

4. In the power conversion device according to claim 2, The power conversion device is characterized in that the machine learning model has been trained using the slip frequency and rotation frequency of the induction motor in a stable driving state as training signals.

5. In the power conversion device according to claim 1, The aforementioned motor is a synchronous motor, The power conversion device includes a phase error calculation unit that calculates a first phase error estimate value, which represents an estimated value of the error between the control phase used for controlling the synchronous motor and the phase, using the voltage command value and the current value of the rotating coordinate system. The input value includes at least one of the voltage command value and the current value of the rotating coordinate system, and the first phase error estimate. The machine learning device outputs an output value that includes a second phase error estimate obtained by correcting the first phase error estimate for the input value. The power conversion device is characterized in that the phase calculation unit calculates the phase estimate value based on the second phase error estimate value.

6. In the power conversion device according to claim 5, The power conversion device is characterized in that the input value further includes at least one of the active power value and the reactive power value of the rotating coordinate system.

7. In the power conversion device according to claim 5, The power conversion device is characterized in that the machine learning model has been trained using the error when the synchronous motor is in a stable driving state as a training signal.

8. In the power conversion device according to any one of claims 1 to 7, The power conversion device is characterized in that the machine learning model is configured using a pre-trained deep learning network.

9. In the power conversion device according to any one of claims 1 to 7, The power conversion device is characterized in that the machine learning model is configured using a deep learning network that was trained while the electric motor was in operation.

10. In the power conversion device according to any one of claims 1 to 7, The machine learning device is characterized by replacing the machine learning model with another machine learning model created by a higher-level device connected to the power conversion device.

11. In the power conversion device according to claim 10, A power conversion device characterized by transmitting the input values ​​obtained when the electric motor is driven to the higher-level device, and receiving from the higher-level device another machine learning model created by the higher-level device through unsupervised learning using the input values.

12. In the power conversion device according to any one of claims 2 to 4, The system includes an output signal selection unit that selects either the slip frequency command value and the first rotation frequency estimate, or the slip frequency estimate and the second rotation frequency estimate, as an output signal to the phase calculation unit. The power conversion device is characterized in that the phase calculation unit calculates the phase estimate value based on the output signal selected by the output signal selection unit.

13. In the power conversion device according to any one of claims 5 to 7, The system includes an output signal selection unit that selects either the first phase error estimate or the second phase error estimate as the output signal to the phase calculation unit, The power conversion device is characterized in that the phase calculation unit calculates the phase estimate value based on the output signal selected by the output signal selection unit.

Images