Motor control device

Abstract

To provide a motor control device which can more optimally improve efficiency in real time by incorporating physical restriction in learning algorithm when an output signal which directly becomes the optimal efficiency is derived in terms of learning by neural network structure.SOLUTION: A motor control device 1 controls a motor 6, comprises a neural network compensator 11 which derives an output signal which becomes the optimal efficiency by inputting an input signal to repeat learning by propagation and reverse propagation, and derives the output signal in terms of learning from the input signal to minimize a teacher signal by using the input signal as a current peak command value ip* and a current peak value ip, using the output signal as a current phase command value θi*, and using a value based on an inclination Δip / Δθi* as the teacher signal.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a motor control device that controls the operation of a motor by utilizing neural network control. [Background technology]

[0002] Interior permanent magnet (IPM) motors (permanent magnet synchronous motors) have a structure in which permanent magnets are placed inside the rotor, and can also use reluctance torque, making it easy to achieve high efficiency, so they have been widely used in home appliances, industrial equipment, automotive applications, etc. In addition, with the recent development of AI technology, its introduction in the field of motor control has also been considered.

[0003] For example, Patent Document 1 proposes a learning device and a learning method for optimizing the PI gain of a current controller in a motor current control system by learning using the amount of overshoot, undershoot, and rise time of a current in response to a step-like torque command as a reward. Also, for example, Patent Document 2 proposes a machine learning method capable of learning an optimal current command for a motor. In this document, a motor current command value is derived by learning using the motor torque, motor current, and motor voltage as a reward.

[0004] Furthermore, for example, Patent Document 3 proposes a device that uses a neural network to derive a primary voltage and a phase angle and control an induction machine. However, all of the configurations described in these documents have the problem that it is difficult to minimize losses and prevent efficiency degradation with good responsiveness in response to fluctuations in motor parameters due to product variations and aging of the motor.

[0005] Therefore, for example, Patent Document 4 proposes a motor control device that uses neural network control to learn and derive an output signal that provides optimal efficiency in real time, and performs compensation to optimize the overall efficiency of the motor (mechanical output / power input). [Prior art documents] [Patent documents]

[0006] [Patent Document 1] JP 2017-34844 A [Patent Document 2] JP 2018-14838 A [Patent Document 3] Patent No. 3054521 [Patent Document 4] JP 2022-42871 A Summary of the Invention [Problem to be solved by the invention]

[0007] However, in conventional compensation using neural network control, the learning method is to learn by feeding data ad-hoc, so it deviates from the optimal value according to the control device and external environmental factors. In other words, if the parameters change due to manufacturing variations in the motor or aging, it cannot operate at the operating point where loss is minimized, and the overall efficiency (mechanical output / power input) decreases.

[0008] In addition, even if inexpensive parts with large variations are used, it is not possible to operate at the operating point where losses are minimized, and the overall efficiency also decreases. This leads to an increase in current and heat, shortening the lifespan of the elements. There was also the problem of shortening the operating time to prevent inverter failure.

[0009] The present invention has been made to solve the above-mentioned conventional technical problems, and aims to provide a motor control device that can achieve more optimal efficiency improvement in real time by incorporating physical constraints into the learning algorithm when an output signal that directly results in optimal efficiency is derived through learning using a neural network structure. [Means for solving the problem]

[0010] The motor control device of the present invention is provided with a neural network compensator that receives an input signal and repeats learning by forward propagation and backward propagation to derive an output signal for improving the efficiency of the motor, and controls the motor based on the output signal derived by the neural network compensator. The neural network compensator converts the input signal into a current peak command value i p * and the current peak value i p The output signal is the current phase command value θ i * And with a slope of Δi p / Δθ i * The present invention is characterized in that a value based on the above is used as a teacher signal, and an output signal is derived from an input signal through learning so as to minimize the teacher signal.

[0011] The motor control device of the present invention is the above-mentioned invention, p / Δθ i * The sign function sgn(Δi p / Δθ i * ) is used as a teacher signal for the neural network compensator.

[0012] The motor control device of the invention of claim 3 is the invention of claim 1, wherein the neural network compensator receives a mechanical angular velocity command value ω * and the estimated mechanical angular velocity ω are further input, and the compensation current value i N is further derived as an output signal.

[0013] The motor control device of the invention according to claim 4 is characterized in that in the above invention, the neural network compensator is of a single configuration including an input layer, an intermediate layer, and an output layer.

[0014] The motor control device of the present invention according to claim 5 is characterized in that in each of the above inventions, the motor is a permanent magnet synchronous motor.

[0015] The motor control device of the invention of claim 6 is characterized in that it comprises a motor drive unit that drives and controls a motor in the inventions of claims 1 to 4, and a motor control unit that controls the motor by the motor drive unit based on the output signal of the neural network compensator. Effect of the Invention

[0016] According to the present invention, a control device is provided with a neural network compensator that receives an input signal and repeats learning by forward propagation and backward propagation to derive an output signal for improving the efficiency of a motor, and controls a motor based on the output signal derived by the neural network compensator. The neural network compensator converts the input signal into a current peak command value i p * and the current peak value i p The output signal is the current phase command value θ i * And with a slope of Δi p / Δθ i * As a result, even if there is product variation in the motor, or if the motor parameters vary due to magnetic saturation, deterioration over time, or temperature changes, loss is minimized in real time, a decrease in efficiency is prevented, and the motor can be driven with minimal loss.

[0017] This makes it possible to use inexpensive motors with greater variation, and also significantly reduces the amount of work required to adapt the parameters, thereby enabling costs to be reduced and achieving so-called robustness.

[0018] In particular, the neural network compensator has a slope Δi p / Δθ i * The value based on the above is used as a teacher signal, and the output signal is derived from the input signal by learning so as to minimize the teacher signal. p / Δθ i *By incorporating these physical constraints into the learning algorithm of the neural network, it becomes possible to drive the motor with more optimal efficiency. It also becomes more compatible with a speed controller that does not generate overshoot.

[0019] According to the invention of claim 2, in the above invention, the inclination Δi p / Δθ i * The sign function sgn(Δi p / Δθ i * ) is used as the teacher signal for the neural network compensator, it is possible to improve the simplicity of calculations in the neural network compensator, the speed of learning, and the stability.

[0020] In addition, as in the invention of claim 3, the neural network compensator receives as an input signal a mechanical angular velocity command value ω * and the estimated mechanical angular velocity ω are further input, and the compensation current value i N By further deriving the above as an output signal, it becomes possible to improve the speed response of the motor in real time to disturbances. This also makes it possible to reduce the number of design steps for the upper control device that generates the speed command.

[0021] In particular, since the neural network compensator is an integrated form of neural networks for improving motor efficiency and dynamic characteristics, a single configuration such as the invention of claim 4 makes it possible to design efficiency improvement and dynamic characteristic improvement in a unified manner, and it is also possible to reduce the network size and calculation load compared to when each output is generated by a separate neural network.

[0022] This makes it possible to improve the computational efficiency and reduce power consumption while improving the stability of the system, and also reduces costs because a computation device with large computational resources is not required. Furthermore, since learning that contributes to both efficiency improvement and dynamic characteristic improvement is performed, it is possible to further improve the performance of efficiency improvement and dynamic characteristic improvement compared to the case where a neural network for efficiency improvement and a neural network for dynamic characteristic improvement are provided separately.

[0023] Furthermore, the above-mentioned inventions are effective for a permanent magnet synchronous motor such as that of the invention of claim 5. Specifically, as in the invention of claim 6, a motor drive unit that drives and controls the motor, and a motor control unit that controls the motor by the motor drive unit based on the output signal of the neural network compensator are further provided to control the motor. [Brief description of the drawings]

[0024] [Figure 1] 1 is a block diagram of a motor control device according to an embodiment of the present invention (first embodiment). [Diagram 2] FIG. 2 is a block diagram of a neural network compensator of the motor controller of FIG. 1. [Diagram 3] FIG. 3 is a diagram illustrating an example of the internal structure of the neural network compensator of FIG. 2. [Figure 4] FIG. 1 is a diagram showing optimal phase conditions on the dq axes for explaining the relationship between the current peak value and the current phase at which the motor achieves optimal efficiency. [Diagram 5] FIG. 13 is a diagram showing an optimum phase condition on the ip-θi axis for explaining the relationship between the current peak value and the current phase at which the motor achieves optimum efficiency. [Figure 6] 1 is a diagram illustrating a response waveform to a loss in a motor control device to which the present invention is applied. FIG. [Figure 7] 1A to 1C are diagrams illustrating response waveforms of current (id, iq, ip) and phase (θi) of a motor control device to which the present invention is applied. [Figure 8]FIG. 13 is a diagram illustrating a response waveform to a loss after a physical parameter of the motor control device to which the present invention is applied is changed. [Figure 9] 1A to 1C are diagrams illustrating response waveforms of current (id, iq, ip) and phase (θi) after a change in physical parameters of a motor control device to which the present invention is applied. [Figure 10] FIG. 13 is a diagram illustrating the ratio of each evaluation value to an optimal value before and after a parameter change. [Figure 11] FIG. 11 is a block diagram of a motor control device according to another embodiment to which the present invention is applied (Embodiment 2). [Figure 12] FIG. 12 is a block diagram of a neural network compensator of the motor controller of FIG. [Figure 13] FIG. 13 is a diagram illustrating an example of the internal structure of the neural network compensator of FIG. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0025] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. EXAMPLES

[0026] (1) Motor control device 1 1 is a block diagram showing the configuration of a motor control device 1 according to one embodiment of the present invention. The motor control device 1 according to this embodiment includes an inverter circuit 9 as a motor drive unit, and a motor control unit 3, and is configured to convert and generate AC power of a predetermined frequency and supply it to a motor 6. The motor 6 is a three-phase embedded permanent magnet motor that drives an electric compressor used in an air conditioner of an electrically powered vehicle such as an electric vehicle or a hybrid vehicle. In this embodiment, an Interior Permanent Magnet Synchronous Motor (IPMSM) is used, and is driven by the inverter circuit 9 according to a voltage command generated by the motor control unit 3.

[0027] (2) Inverter circuit 9 (motor drive unit) The inverter circuit 9 is configured by connecting multiple (six) switching elements in a bridge configuration. Each switching element of the inverter circuit 9 is switched by a PWM signal generated by a PWM signal generator 8 of the inverter control unit 3, which will be described later.

[0028] (3) Motor control unit 3 In this embodiment, the motor control unit 3 calculates the estimated mechanical angular velocity ω' of the motor 6. m and the mechanical angular velocity command value ω * Based on the deviation from the d-axis voltage command value Vd * , q-axis voltage command value Vq * These d-axis voltage command values ​​Vd * , q-axis voltage command value Vq * Finally, a PWM signal generator 8 is used to generate a PWM signal for switching each switching element of the inverter circuit 9, and the motor 6 is driven by sensorless vector control. Note that the motor 6 may be controlled by using a position sensor instead of the sensorless control.

[0029] The motor control unit 3 of this embodiment is composed of a microcomputer, which is an example of a computer equipped with a processor, and has as its functions a neural network compensator 11, a speed controller 12, a polar coordinate converter 13, a current controller 14, a decoupling compensator 16, a phase voltage command calculator 7, a PWM signal generator 8, a dq-axis current converter 10, a three-phase current estimator 17, a magnet position estimator 18, and a rotation speed calculator 19, etc.

[0030] The three-phase current estimator 17 estimates each phase voltage output by the phase voltage command calculator 7, that is, the U-phase voltage command value V * u , V-phase voltage command value V * v , W-phase voltage command value V * w (Six PWM signals generated by the PWM signal generator 8 may be used.) From the phase current of one phase flowing through the inverter circuit 9 detected by one shunt resistor, each phase current (U-phase current i u , V phase current i v, W phase current i w ) is estimated (single-shunt current detection method). Other possible methods for detecting the phase current include the two-shunt current detection method, which uses two shunt resistors to detect two-phase currents, the three-shunt current detection method, which uses three shunt resistors to detect three-phase currents, and the Hall CT current detection method, which detects phase currents using a Hall CT.

[0031] In this embodiment, the magnet position estimator 18 estimates the phase currents output by the three-phase current estimator 17, i.e., the U-phase current i u , V phase current i v , W phase current i w From the estimated electrical angle θ ' e The estimated electrical angle θ ' e In addition to these, the U-phase voltage command value V * u , V-phase voltage command value V * v , W-phase voltage command value V * w The d-axis voltage command value Vd * , q-axis voltage command value Vq * Also, the d-axis voltage command value Vd * , q-axis voltage command value Vq * and d-axis current i d , q-axis current i q In addition, these U-phase currents i u , V phase current i v , W phase current i w , U-phase voltage command value V * u , V-phase voltage command value V * v , W-phase voltage command value V * w , d-axis voltage command value Vd * , q-axis voltage command value Vq * , d-axis current i d , q-axis current i q The electrical angle estimation value θ is calculated by using any one of the above, a combination of the above, or all of the above. ' e may be estimated.

[0032] The rotation speed calculator 19 calculates the estimated electrical angle θ ' e From the above, the estimated mechanical angular velocity ω ' m Furthermore, the dq-axis current converter 10 estimates the electrical angle estimated value θ ' e From d-axis current i d and the q-axis current i q In addition, the estimated electrical angle θ output by the magnet position estimator 18 is calculated. ' e is further input to the phase voltage command calculator 7, and is also converted into the d-axis current i d and q-axis current i q and the estimated mechanical angular velocity ω output by the rotation speed calculator 19. ' m is input to the decoupling compensator 16.

[0033] In addition, the estimated mechanical angular velocity ω ' m is further input to a subtractor 21. The subtractor 21 receives the mechanical angular velocity command value ω * is input, and the mechanical angular velocity command value ω m * From the estimated mechanical angular velocity ω ' m In addition, when the motor 6 is controlled using a position sensor as described above, the mechanical angular velocity (ω m ) is the estimated mechanical angular velocity ω ' m is input to the subtractor 21 instead of

[0034] The deviation calculated by the subtractor 21 is input to the speed controller 12. The speed controller 12 calculates IP and the current peak value i p From the relation between the current and torque, the current peak command value i p * Note that the calculation is not based on the formula, but on the current peak value i pUsing a map created offline based on the relationship between current and torque, the current peak command value i p * Alternatively, when using an equation, the parameters may be identified or estimated online to improve accuracy.

[0035] The polar coordinate converter 13 receives the current peak command value i p * The polar coordinate converter 13 receives a current phase command value θ i * The neural network compensator 11 will be described in detail later.

[0036] The polar coordinate converter 13 converts these current phase command values ​​θ i * and the current peak command value i p * to d-axis current command value i d * and the q-axis current command value i q * The polar coordinate converter 13 derives the d-axis current command value i d * and the q-axis current command value i q * Derive:

[0037]

number

[0038] The d-axis current command value i output by the polar coordinate converter 13 d * and the q-axis current command value i q * is input to the subtractors 22 and 23, respectively. The subtractors 22 and 23 receive the d-axis current i d and q-axis current i q are input, and the deviations are calculated in the subtractors 22 and 23, respectively.

[0039] The deviations output by the subtractors 22 and 23 are input to the current controller 14. The current controller 14 performs a PI calculation using the deviations to obtain a d-axis voltage command value V d * and the q-axis voltage command value V q * These d-axis voltage command values ​​V d * and the q-axis voltage command value V q * After the interference between the d and q axes is cancelled by the decoupling compensator 16 (in FIG. 1, the output is V ' d * , V ' q * ) is input to the phase voltage command calculator 7. Incidentally, this decoupling compensator 16 may be omitted.

[0040] The phase voltage command calculator 7 calculates the d-axis voltage command value V ' d * and the q-axis voltage command value V ' q * , the electrical angle estimate value θ output by the magnet position estimator 18 ' e to U-phase voltage command value V u * , V-phase voltage command value V v * , W-phase voltage command value V w * The PWM signal generator 8 generates a voltage command value V u * , V v * , V w * A PWM signal for switching (PWM control) each switching element of the inverter circuit 9 is generated from the inverter circuit 9. Then, each phase voltage Vu, Vv, Vw is applied to the motor 6 from the inverter circuit 9, thereby achieving sensorless vector control of the motor 6 in this embodiment.

[0041] (4) Neural network compensator 11 (Example 1) Next, the neural network compensator 11 in Fig. 1 will be described in detail with reference to Fig. 2 and Fig. 3. Fig. 2 is a block diagram of the neural network compensator 11 of this embodiment, and Fig. 3 is a diagram showing the internal structure of the neural network compensator 11. The neural network compensator 11 receives an input signal and repeats learning through forward propagation and backward propagation to derive an output signal for improving the efficiency of the motor 6, i.e., an output signal that makes the motor 6 optimally efficient.

[0042] In addition, the output signal for improving the efficiency of the motor 6 is a current phase command value θ i * The input signal is the current value of the motor 6, i p and the current peak command value i p * (Current peak value i p (command value).

[0043] In addition, the teacher signal to be minimized in the present invention is sgn(Δi p / Δθ i * ) The neural network compensator 11 of the embodiment of FIG. 1 and FIG. 2 uses the teacher signal sgn(Δi p / Δθ i * In this application, the output signal is derived from the input signal by learning so as to minimize the current phase command value θ i * In order to derive the current peak command value i p * and the current peak value i p , the teacher signal to be minimized is sgn(Δi p / Δθ i * ) is used.

[0044] Here, as shown in Fig. 4, the required torque T e For the current peak value i p The current phase θ that minimizes i As shown in Figure 5, when the efficiency is optimal, the slope Δip / Δθ i * becomes zero (0). This is a physical constraint in the present invention. In addition, in this embodiment, in consideration of the simplicity of calculation, the speed of learning, and the improvement of stability, the slope Δi p / Δθ i * The sign function sgn is applied to the control time, and learning is performed using only the sign (-, 0, +) of the slope at the control time. That is, the neural network compensator 11 repeats forward propagation and backward propagation to obtain the sign function sgn(Δi p / Δθ i * ) is the optimal current phase command value θ i * Derive:

[0045] In FIG. 1, the current peak command value i p * and the optimal current phase command value θ derived by the neural network compensator 11 through learning. i * Thus, the speed of the motor 6 is controlled in a state of optimum efficiency.

[0046] Next, the neural network compensator 11 of the embodiment will be described in more detail with reference to the internal structure shown in FIG. p / Δθ i * The sign function sgn(Δi p / Δθ i * )) is minimized by the compensation amount (output signal: current phase command value θ i * ) to the current input signal (current peak command value i p * and the current peak value i p ) is derived through learning.

[0047] Here, the loss (copper loss) is the square of the current peak value i p 2 The torque is proportional to the current peak value i p The current peak value ip is the current phase θ i Therefore, the neural network compensator 11 determines the current peak value i within the range that satisfies the required torque condition. p The current phase θ at which i is derived by updating the information every calculation cycle.

[0048] The internal structure of the neural network (NN) compensator 11 in this embodiment is as shown in Fig. 3 (two inputs, one output). There are multiple intermediate layers 27 between the input layer 24 and the output layer 26, and each intermediate layer 27 is composed of multiple neurons 28, which are threshold values, and the output layer 26 is also composed of neurons 29, which are threshold values.

[0049] The input layer 24 and neurons 28 in the first hidden layer 27 are connected by synapses (shown as a whole at 31), which are the connection weights between the input layer 24 and the hidden layer 27 in question, the neurons 28 in the previous and next hidden layers 27 are connected by synapses (shown as a whole at 32), and the neuron 28 in the last hidden layer 27 and neuron 29 in the output layer 26 are connected by a synapse (shown as a whole at 33).

[0050] The connection equation of each of the neurons 28 and 29 is expressed as equation (II). The neural network compensator 11 of the embodiment has an input layer (2 inputs), a second layer (first hidden layer: 10 neurons), a third layer (last hidden layer: 10 neurons), and an output layer (1 output).

[0051]

number

[0052] Here, in formula (II), y is the result, x is the input, and w i are weights, all of which are vectors. b is the threshold (bias), and σ is the activation function. The weights w i Then, the update formula (learning formula) for the threshold b is given by formula (III).

[0053]

number

[0054] Here, α and β in formula (III) are learning rates, and E is a loss function (current peak value i p The calculation procedure of the neural network compensator 11 is as follows: Step 1: forward propagation calculation (shown in formula (IV)), Step 2: back propagation calculation (shown in formula (V)), and Step 3: gradient descent calculation (shown in formula (VI)) are performed in one sampling period (example of a three-layer structure).

[0055]

number

[0056]

number

[0057]

number

[0058] In addition, the subscripts in parentheses () in formulas (IV) to (VI) indicate the connection between the current layer and the next layer. Also, the subscripts in brackets [] indicate the weights transmitted from which layer to which layer. Subscript 1 is 2×10, subscript 2 is 10×10, and subscript 3 is 10×1.

[0059] In addition, W in formulas (IV) to (VI) [1] , W [2] , W [3] are the weights of each layer, B [1] , B [2] , B [3] are the thresholds for each layer, and α and β are the learning rates (hyperparameters) for updating the weights. [1] , A [2] is the neuron's connection equation, and σ is the activation function. Furthermore, θ i *is the current phase command value θ i * , L is the loss function, and the subscript T is the transpose matrix.

[0060] The neural network compensator 11 receives the teacher signal sgn(Δi p / Δθ i * ) is minimized by forward and backward propagation, and the input signal (current peak command value i p * and the current peak value i p ) to the output signal (current phase command value θ i * ) is derived by learning. Then, the current phase command value θ i * is input to the polar coordinate converter 13.

[0061] (5) Evaluation of the motor control device 1 in Figure 1 Next, the effect of optimal efficiency control of the motor 6 using the neural network compensator 11 of this embodiment (FIGS. 1 to 3) will be described with reference to Figures 6 to 10. Figure 6 shows a time response waveform for losses when there is no change in physical parameters. The percentages shown in the figure indicate the ratio to the optimal value (Optimal). The solid line shows the case of the motor control device 1 of the embodiment (NN-based MTPA), and the dashed line shows the case of motor control by conventional MTPA without using a neural network compensator.

[0062] From FIG. 6, it can be seen that the motor control device 1 of this embodiment (optimum efficiency ratio 100%) and the conventional general MTPA (100%) are almost equivalent and have optimal efficiency. d and q-axis current i q , and current peak value i p and current phase θ i The time response waveforms of are shown in Fig. 7. It can be seen from Fig. 7 that the current levels of each control match the optimal conditions.

[0063] Next, we compare and verify the results when the physical parameters are changed. Here, taking into consideration practical applications, we assume that the physical parameters after the change cannot be used. In other words, with MTPA, only the parameters before the change can be used, and the design is based on those values. Figure 8 shows the time response waveform for losses. As can be seen from Figure 8, with conventional MTPA (dashed line), optimality is lost due to changes in physical parameters, with the loss level deteriorating by 3.7% from the optimal efficiency.

[0064] In contrast, in the motor control device 1 of this embodiment (solid line), it is clear that the loss can be maintained at an optimal value (0% degradation) by learning using neural network control without requiring physical parameters. d and q-axis current i q , and current peak value i p and current phase θ i The time response waveforms of are shown in Fig. 9. From Fig. 9, it can be seen that in the conventional MTPA, due to changes in the physical parameters, each value deviates from the optimal condition (dashed line), but in the motor control device 1 of the embodiment, each value approximately coincides with the optimal value (solid line).

[0065] Figure 10 shows a comparison of the conventional MTPA and the motor control device 1 of the embodiment (neural network compensation) before and after the parameter change, with each value summarized as a percentage of the optimal value. In the figure, (a) shows the results before the physical parameters were changed, and (b) shows the results after the changes. As can be seen from this figure, after the parameters were changed, the motor control device 1 of the embodiment took all values ​​close to the optimal values, confirming its superiority over general MTPA.

[0066] As described above, the input signal of the neural network compensator 11 is converted into the current peak command value i p * and the current peak value i p The output signal is the current phase command value θ i * And with a slope of Δi p / Δθ i *As a result, even if there is product variation in motor 6 or if motor parameters vary due to magnetic saturation, deterioration over time, or temperature changes, loss is minimized in real time, a decrease in efficiency is prevented, and the motor can be driven with minimal loss.

[0067] This makes it possible to adopt an inexpensive motor 6 with greater variation, and also significantly reduces the amount of work required for parameter adaptation, thereby enabling cost reduction and achieving so-called robustness.

[0068] In particular, the neural network compensator 11 calculates the slope Δi p / Δθ i * The value based on the above is used as a teacher signal, and the output signal is derived from the input signal by learning so as to minimize the teacher signal. p / Δθ i * By incorporating these physical constraints into the learning algorithm of the neural network, it becomes possible to drive the motor with more optimal efficiency. It also increases the affinity with a speed controller that does not generate overshoot. In other words, if a speed overshoot occurs, unnecessary power consumption occurs due to it, and learning is performed in a state containing noise. Therefore, in the present invention, by configuring the speed controller with a control system that does not generate overshoot, optimal learning becomes possible.

[0069] In the embodiment, the slope Δi p / Δθ i * The sign function sgn(Δi p / Δθ i * ) is used as the teacher signal for the neural network compensator 11, which makes it possible to simplify the calculations in the neural network compensator 11, speed up learning, and improve stability. In particular, the present invention is effective for the motor 6 that is a permanent magnet synchronous motor as in the embodiment. EXAMPLES

[0070] Next, a motor control device 1 according to another embodiment of the present invention will be described with reference to Figures 11 to 13. Figure 11 is a block diagram showing the configuration of the motor control device 1 according to this embodiment (embodiment 2). In Figures 11 to 13, the same reference characters as in Figures 1 to 10 denote the same or similar functions, and descriptions of common parts will be omitted.

[0071] In this embodiment, the mechanical angular velocity command value ω * and the estimated mechanical angular velocity ω is input to the speed controller 12. In this embodiment, the three-phase current estimator 17, magnet position estimator 18, rotation speed calculator 19, subtractor 21, speed controller 12, etc. form a feedback control system in the motor control unit 3.

[0072] In addition, the speed controller 12 of this embodiment performs IP calculation and current peak value i p From the relation between the current and torque, the current peak command value i p * The feedback control output value i for deriving pf * Note that the calculation is not based on the above formula, but on the current peak value i p Using a map created offline based on the relationship between torque and feedback control output value i pf * Alternatively, when using an equation, the parameters may be identified or estimated online to improve accuracy.

[0073] In this embodiment, an adder 20 is provided to add the feedback control output value i pf * is input to the adder 20. The adder 20 also receives the compensation current value i N is input, and the feedback control output value i pf * , the compensation current value iN is added to the current peak command value i p * is derived.

[0074] The polar coordinate converter 13 receives the current peak command value i p * is input as the other input. One input of the polar coordinate converter 13 is the current phase command value θi * The circuitry after the polar coordinate converter 13 is the same as that of the embodiment described above.

[0075] (6) Neural network compensator 11 (Example 2) Next, the neural network compensator 11 of this embodiment will be described in detail with reference to Figures 12 and 13. Figure 12 is a block diagram of the neural network compensator 11, speed controller 12 and polar coordinate converter 13 of this embodiment, and Figure 13 shows the internal structure of the neural network compensator 11.

[0076] The neural network compensator 11 of this embodiment is a single configuration that integrates learning functions for improving the efficiency and dynamic characteristics of the motor 6, and derives an output signal for improving the efficiency of the motor 6 and an output signal for improving the dynamic characteristics of the motor 6 by inputting an input signal and repeating learning through forward propagation and backward propagation.

[0077] In this embodiment, the output signal for improving the efficiency of the motor 6 is the current phase command value θ i * The output signal for improving the dynamic characteristics of the motor 6 is the compensation current value i N The input signal is the current value of the motor 6, i p and the current peak command value i p * (Current peak value i p ), the mechanical angular velocity ω as the angular velocity value of the motor 6 and the mechanical angular velocity command value ω * (Command value of mechanical angular velocity ω).

[0078] In addition, there are two teacher signals to be minimized. One is the aforementioned sgn(Δi p / Δθ i * ) The other is the feedback control output value i pf * The zero squared error for (i pf * -0) 2 , that is, (i pf * ) 2 In this embodiment, the neural network compensator 11 calculates each teacher signal sgn(Δi p / Δθ i * ) and (i pf * ) 2 An output signal is derived from an input signal through learning so as to minimize the following: In this application, the minimization naturally includes zero.

[0079] That is, the current phase command value θi that provides optimal efficiency * In order to derive the current peak command value i p * and the current peak value i p , the teacher signal to be minimized is sgn(Δi p / Δθ i * ) is used. Furthermore, the compensation current value i N To derive the above, the input signal includes the mechanical angular velocity command value ω * and mechanical angular velocity ω, the teacher signal to be minimized is the feedback control output value i pf * The zero squared error for (i pf * ) 2 Use the following.

[0080] The neural network compensator 11 of this embodiment is also a multi-layered neural network compensator, and repeats learning by forward propagation and backward propagation to generate a teacher signal sgn(Δi p / Δθ i * ) is the optimal current phase command value θ i* and (i pf * ) 2 The optimal compensation current value i that minimizes (zeroes) N Derive:

[0081] In this embodiment, the target input is the current peak value i p and mechanical angular velocity ω, and is considered to be a control method more suitable for controlling the motor 6 consisting of the permanent magnet synchronous motor of the embodiment. In particular, it is effective in cases where the current torque τ cannot be detected or when used in a speed controller. p The optimal current phase command value θ i * and the optimal compensation current value i calculated together with the mechanical angular velocity ω N As a result, the motor 6 is accurately controlled with optimal efficiency and good speed responsiveness.

[0082] Next, the neural network compensator 11 of this embodiment will be described in more detail with reference to the internal structure shown in FIG. p / Δθ i * ) and (i pf * ) 2 The compensation amount (output signal) is derived from the current input signal through learning so as to minimize

[0083] That is, the current peak value i p Current peak command value i p * (Required current peak value) compensation amount (current phase command value θ i * : output signal) to the currently available information (current peak command value i p * , current peak value i p : Input signal) is updated every calculation period, and the feedback control output value i pf * (required feedback control output value) compensation amount (compensation current value i N: output signal) to the currently available information (mechanical angular velocity command value ω * , mechanical angular velocity ω: input signal) by updating the information every calculation period.

[0084] The internal structure of the neural network (NN) controller 11 in this embodiment is as shown in Fig. 13 (four inputs, two outputs). In this case, there are also multiple intermediate layers 27 between the input layer 24 and the output layer 26, and each intermediate layer 27 is composed of multiple neurons 28, which are threshold values, and the output layer 26 is also composed of neurons 29, which are threshold values ​​(two in this case).

[0085] The input layer 24 and neurons 28 in the first hidden layer 27 are connected by synapses (shown as a whole at 31), which are the connection weights between the input layer 24 and the hidden layer 27 in question, the neurons 28 in the previous and next hidden layers 27 are connected by synapses (shown as a whole at 32), and the neuron 28 in the last hidden layer 27 and neuron 29 in the output layer 26 are connected by a synapse (shown as a whole at 33).

[0086] The coupling equation for each neuron 28, 29 is the above-mentioned equation (II), and the weight w i The update formula (learning formula) of the threshold θ is the above-mentioned formula (III). The neural network compensator 11 of this embodiment has an input layer (4 inputs), a second layer (first hidden layer: 10 neurons), a third layer (last hidden layer: 10 neurons), and an output layer (2 outputs).

[0087] In this embodiment, the neural network compensator 11 calculates the training signal sgn(Δi p / Δθ i * ) and (i pf * ) 2 The synapses 31 to 33 are updated by forward propagation and backward propagation so as to minimize the input signal (current peak command value i p * and the current peak value i p ) to the output signal (current phase command value θ i* ) is derived through learning, and the input signal (mechanical angular velocity command value ω * and mechanical angular velocity ω) to output signal (compensation current value i N ) is derived through learning.

[0088] The current phase command value θ i * is input as is to the polar coordinate converter 13. On the other hand, the compensation current value i N is the feedback control output value i output by the speed controller 12 at the adder 20. pf * Then, the compensation current value i N The feedback control output value i compensated by pf * is the current peak command value i p * (Output signal: compensation current value i N The value obtained by

[0089] Here, if a neural network compensator for improving efficiency and a neural network compensator for improving dynamic characteristics are provided separately, the total number of synapses in each neural network compensator will be 260, the total number of neurons will be 42, the total number of multiplications (forward propagation) will be approximately 260, and the total number of multiplications (backpropagation) will be approximately 3,000.

[0090] On the other hand, in the neural network compensator 11 of this embodiment, these are integrated (integrated neural network compensator), so that the neurons 28 in the intermediate layer 27 and each of the synapses 31 to 33 can be shared, and the number of synapses is 160, the number of neurons is 22, the number of multiplications (forward propagation) is approximately 180, and the number of multiplications (backward propagation) is approximately 1,600, which is a reduction of just over half.

[0091] As in the previous embodiment, this embodiment also minimizes losses in real time, prevents efficiency degradation, and enables motor drive with minimum losses even when there is product variation in the motor 6, or when motor parameters fluctuate due to magnetic saturation, aging, or temperature changes, making it possible to drive the motor with minimum losses. This not only makes it possible to reduce costs, but also makes it possible to achieve so-called robustness.

[0092] In particular, in this case, the neural network compensator 11 also has a slope Δi p / Δθ i * The value based on the above is used as a teacher signal, and the output signal is derived from the input signal by learning so as to minimize the teacher signal. p / Δθ i * By incorporating these physical constraints into the neural network learning algorithm, it becomes possible to drive the motor with more optimal efficiency. It also becomes more compatible with speed controllers that do not generate overshoot.

[0093] Also in this embodiment, the slope Δi p / Δθ i * The sign function sgn(Δi p / Δθ i * ) is used as the teacher signal for the neural network compensator 11, it is possible to improve the simplicity of calculations in the neural network compensator 11, the speed of learning, and stability.

[0094] Furthermore, in this embodiment, the neural network compensator 11 receives as an input signal a mechanical angular velocity command value ω * and the estimated mechanical angular velocity ω are further input, and the compensation current value i N is further derived as an output signal, it is possible to improve the speed responsiveness of the motor 6 in real time to disturbances. This also makes it possible to reduce the number of design steps for the upper control device that generates the speed command.

[0095] In particular, in this embodiment, the neural network compensator 11 is an integrated form of neural networks for improving the efficiency and dynamic characteristics of the motor 6, making it possible to design a single configuration that takes into consideration both efficiency improvement and dynamic characteristic improvement in a unified manner, and also making it possible to reduce the network size and calculation load compared to when each output is generated by a separate neural network.

[0096] As a result, it is possible to improve the computational efficiency and reduce power consumption while improving the stability of the system, and since a computation device with large computational resources is not required, it is also possible to reduce costs. Furthermore, since learning that contributes to both efficiency improvement and dynamic characteristic improvement is performed, it is possible to further improve the performance of efficiency improvement and dynamic characteristic improvement compared to the case where a neural network for efficiency improvement and a neural network for dynamic characteristic improvement are provided separately.

[0097] Incidentally, the controlled object of the motor control device 1 of the present invention is not limited to the permanent magnet synchronous motor shown in the embodiment, except for the invention of claim 5. [Explanation of symbols]

[0098] 1 Motor control device 3 Motor control unit 6 Motor 7-phase voltage command calculator 8 PWM signal generator 9 Inverter circuit (motor drive section) 10 dq axis current transducer 11 Neural Network Compensator 12 Speed ​​Controller 13 Polar Coordinate Converter 14 Current Controller 16 Decoupling Controller 17 Three-phase current estimator 18 Magnet position estimator 19 Rotational speed calculator 24 Input layer 26 Output layer 27 Middle Class 28, 29 Neurons 31~33 Synapse

Claims

1. A motor control device includes a neural network compensator that receives an input signal and repeats learning by forward propagation and backward propagation to derive an output signal for improving motor efficiency, and controls the motor based on the output signal derived by the neural network compensator, The neural network compensator comprises: The input signal is converted into a current height command value i p * and the current peak value i p year, The output signal is converted into a current phase command value θ i * At the same time, Tilt Δi p / Δθ i * a value based on the above is used as a teacher signal, and the output signal is derived from the input signal through learning so as to minimize the teacher signal.

2. The inclination Δi p / Δθ i * The sign function sgn(Δi p / Δθ i * 2. The motor control device according to claim 1, wherein the neural network compensator is configured to have a teacher signal based on the above-mentioned.

3. The neural network compensator receives a mechanical angular velocity command value ω * and the estimated mechanical angular velocity ω are further input, and a compensation current value i N 2. The motor control device according to claim 1, further comprising:

4. 4. The motor control device according to claim 3, wherein the neural network compensator has a single structure including an input layer, an intermediate layer, and an output layer.

5. 5. The motor control device according to claim 1, wherein the motor is a permanent magnet synchronous motor.

6. a motor driving unit that drives and controls the motor; a motor control unit that controls the motor by the motor drive unit based on the output signal of the neural network compensator; 5. The motor control device according to claim 1, further comprising:

Images