Control device for a vibrating actuator and a vibrating drive device having the same, interchangeable lenses, imaging device, automatic stage

Abstract

To provide a control device for a vibration type actuator having a control amount output unit that is different from a conventional PID controller as a main control amount output unit.SOLUTION: A control device for a vibration type actuator includes a control unit having: a first control amount output unit having a first learned model that is machine-learned so as to output a first control amount for moving a contact body relative to a vibrator; and a second control amount output unit having a second learned model that is machine-learned so as to output a second control amount being data having the same data format as the first control amount when inputting a second velocity detected when the contact body is moved relative to a vibrator by the first control amount. The parameters of the first learned model and the parameters of the second learned model are updated based on a control deviation that is a difference between the first control amount and the second control amount output within the same sampling cycle as the first control amount.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a control device for a vibration type actuator, a vibration type driving device having the same, an interchangeable lens, an imaging device, and an automatic stage.

Background Art

[0002] A vibration type motor will be described as an example of a vibration type actuator. The vibration type motor generates high-frequency vibration in an electro-mechanical energy conversion element such as a piezoelectric element coupled to an elastic body by applying an alternating voltage to the electro-mechanical energy conversion element. And it is a non-electromagnetic drive type motor configured to extract the vibration energy as continuous mechanical motion.

[0003] The vibration type motor has excellent motor performance such as small size, light weight, high precision, and high torque during low-speed driving compared with an electromagnetic drive type motor. On the other hand, since it has non-linear motor characteristics, it is difficult to model, and the controllability changes according to driving conditions and temperature environment, so ingenuity is required for the control system. Also, there are many control parameters such as frequency, phase difference, and voltage amplitude, and the adjustment is complicated.

[0004] FIG. 15(a) is a control block diagram of a vibration type driving device by conventional general PID control (see Patent Document 1). An alternating voltage (alternating current signal) of two phases (A phase, B phase) is output from a driving circuit to which a control amount described later is input. By controlling the frequency (1 / period), phase difference, and voltage amplitude (see FIG. 15(b)) of the two-phase alternating voltage output from the driving circuit, the detection speed of the vibration type motor can be controlled. Note that the voltage amplitude is variable depending on the pulse width input from a PID controller to the driving circuit, which will be described later. Also, (1), (2), and (3) in FIG. 15(b) represent the period, phase difference, and voltage amplitude of the two-phase alternating voltage output from the driving circuit, respectively.

[0005] The position deviation, which is the difference (target position - detected position) between the target position generated by the position command unit and the detected position of the vibration-type motor detected by the position detection unit, is input to the PID controller (control variable output unit). Then, the control variables (frequency, phase difference, and pulse width) calculated using PID calculations according to the position deviation input to the PID controller are sequentially output from the PID controller at each control sampling period and input to the drive circuit. Then, a two-phase AC voltage is output from the drive circuit, and the speed of the vibration-type actuator is controlled by the two-phase AC voltage output from the drive circuit. Position feedback control is then performed. The control sampling period will be simply referred to as the "sampling period" below.

[0006] Figure 15(c) schematically shows the frequency-speed characteristics of a vibration motor. Specifically, Figure 15(c) shows that the slope of the frequency-speed characteristics is large at high-speed frequencies (f1) and small at low-speed frequencies (f2). The slope of the speed curve of a vibration motor differs depending on the speed range in which it is used, making it difficult to adjust the PID control gain. For example, it shows that the slope differs between high-speed frequency f1 and low-speed frequency f2. Figure 15(d) schematically shows the phase difference-speed characteristics of a vibration motor, comparing the phase difference-speed characteristics in the low-speed range (f2) and the high-speed range (f1).

[0007] As shown in Figures 15(c) and 15(d), the frequency-speed characteristic slope and phase difference-speed characteristic of a vibration actuator differ depending on the speed range in which it is used, so the control performance changes due to the phase difference. Also, when the ambient temperature changes, for example, from room temperature to low temperature, the resonant frequency shifts to the higher frequency side based on the temperature characteristics of the piezoelectric element. In that case, the speed and slope will be different when driven at the same drive frequency, so the control performance will also change depending on the ambient temperature. Furthermore, the speed and slope will differ due to individual differences in vibration motors, so the control performance will also change depending on the individual unit. In addition, the control performance will change over time. It is necessary to adjust the PID control gains (proportional gain, integral gain, and differential gain of PID control) considering all of these changing factors and design the system to ensure gain margin and phase margin. [Prior art documents] [Patent Documents]

[0008] [Patent Document 1] Japanese Patent Publication No. 2016-144262 [Overview of the project] [Problems that the invention aims to solve]

[0009] Therefore, there has been a need for a control device for a vibration-type actuator that has a control variable output unit different from that of conventional PID controllers as its main control variable output unit. The present invention aims to provide a control device for a vibration-type actuator that has a control variable output unit different from that of conventional PID controllers as its main control variable output unit. [Means for solving the problem]

[0010] One aspect of the present invention is a control device for a vibration-type actuator that moves a contact body in contact with a vibrator relative to the vibrator by vibration generated by the vibrator, comprising: a first control variable output unit having a first trained model that has been machine-trained to output a first control variable for moving the contact body relative to the vibrator when a first speed for moving the contact body relative to the vibrator is input; and a second control variable output unit having a second trained model that has been machine-trained to output a second control variable which is data in the same data format as the first control variable when a second speed detected when the contact body is moved relative to the vibrator by the first control variable is input, wherein the parameters of the first trained model and the parameters of the second trained model are updated based on a control deviation which is the difference between the first control variable and the second control variable output within the same sampling period as the first control variable. [Effects of the Invention]

[0011] According to the present invention, it is possible to provide a control device for a vibration-type actuator that has a control variable output unit different from that of conventional PID controllers as its main control variable output unit. [Brief explanation of the drawing]

[0012] [Figure 1] This is a control block diagram of the vibration-type drive device of the first embodiment. [Figure 2] This diagram illustrates the driving principle of a linear-driven vibration actuator. [Figure 3] This is a perspective view illustrating the lens drive mechanism of a lens barrel. [Figure 4] This figure shows the first trained model, the second trained model, and the neural network configuration of the trained model in the first embodiment. [Figure 5] This is a flowchart of the control process using machine learning and a trained model in the first embodiment. [Figure 6] It is a flowchart or the like when Adam is used as an optimization method for the parameters of a neural network. [Figure 7] (a) compares the calculation results of Adam, RMSprop, Momentum, and SGD using the learning model of the first embodiment and the learning data obtained by actual measurement. (b) is an example of learning the control amount (phase difference) by Adam. [Figure 8] It is a flowchart of adaptive control that operates in the adaptive control unit of the present invention. [Figure 9] It is a timing chart for explaining (a) batch learning in the machine learning unit and (b) adaptive control in the adaptive control unit (control unit). [Figure 10] It is the result when feedback control is performed with a predetermined target position pattern using an unlearned model, and shows the effect of the adaptive control of the present invention. [Figure 11] It compares the results when feedback control is performed with a predetermined target position pattern in the present invention and the conventional PID control. [Figure 12] It is a control block diagram of a vibration type driving device in another form of the first embodiment. [Figure 13] It is a control block diagram of a vibration type driving device in another form of the first embodiment. [Figure 14] It is a control block diagram of a vibration type driving device in another form of the first embodiment. [Figure 15] It is an example showing a control device for a vibration type motor by conventional general PID control. [Figure 16] It is a control block diagram of a vibration type driving device in the second embodiment (when a learned model and a PID controller are used in combination). [Figure 17] It is a control block diagram of a vibration type driving device in another form of the second embodiment. [Figure 18] It is a control block diagram of a vibration type driving device in the third embodiment (when a learned model and a PID controller are used in combination). [Figure 19]It is a control block diagram of a vibration type drive device in another form of the third embodiment. [Figure 20] It is a (a) plan view showing the appearance and (b) schematic diagram of the internal configuration of an imaging device which is an application example of a control device of a vibration type actuator of the fifth embodiment. [Figure 21] It is a perspective view showing the appearance of a microscope which is an application example of a control device of a vibration type actuator in another form of the fifth embodiment.

Mode for Carrying Out the Invention

[0013] (First Embodiment) FIG. 1 is a diagram showing a vibration type drive device 17 of the present embodiment. The vibration type drive device 17 has a control device 15 and a vibration type motor 13 (vibration type actuator). The control device 15 has an adaptive control unit 10 (control unit) for controlling the vibration type actuator, a machine learning unit 12, a drive unit 11, a position detection unit 14 (position detection means), and a speed detection unit 16 (speed detection means). In FIG. 1, what is obtained by removing the vibration type actuator 13 from the vibration type drive device 17 is the control device 15.

[0014] The vibration type actuator 13 has a vibrator 131 and a contact body 132. The position detection unit 14 detects the relative position of the contact body 132 with respect to the vibrator 131 (hereinafter referred to as "relative position"). The relative position detected by the position detection unit 14 is hereinafter referred to as "detection position". The speed detection unit 16 detects the relative speed of the contact body 132 with respect to the vibrator 131 (hereinafter referred to as "relative speed"). The relative speed detected by the speed detection unit 16 is hereinafter referred to as "detection speed".

[0015] As the position detection unit 14, so-called absolute encoders, incremental encoders, etc. are used, but not limited thereto. The speed detection unit 16 is not limited to those that directly detect speed information (speed sensors), and may be those that indirectly detect speed information by calculating position information.

[0016] The control unit 10 is configured to generate a signal to control the drive of the vibrator 131 (the relative movement of the contact body 132 with respect to the vibrator 131). Specifically, the target speed (first speed) and the position deviation are input to a trained model, and the output phase difference and frequency are used as the control variable (first control variable) of the vibrating actuator 13.

[0017] The target speed (first speed) is the speed at which the detection speed (second speed) follows when the contact body 132 is moved relative to the oscillator 131. The position deviation is the difference between the target position (first position) and the detection position (second position). The target position (first position) is the position at which the detection position (second position) follows when the contact body 132 is moved relative to the oscillator 131. The target speed may also be generated by differentiating the target position with respect to time. The target position may also be generated by integrating the target speed.

[0018] Furthermore, in addition to phase difference and frequency, pulse width may also be used to change the voltage amplitude as the control variable (the first control variable and the second control variable described later). As described later, the first control variable is not limited to just phase difference and frequency. It may be one of phase difference, frequency, and voltage amplitude, or a combination of two of phase difference, frequency, and voltage amplitude. It may also be all of phase difference, frequency, and voltage amplitude. It may also be a combination of one or more of phase difference, frequency, and voltage amplitude with a control variable other than phase difference, frequency, and voltage amplitude.

[0019] The control unit 10 includes a speed command unit 101 (speed command means) that generates and commands a target speed, and a position command unit 102 (position command means) that generates and commands a target position. The control unit 10 also includes a first learned model 103, a second learned model 107, and an adaptive learning unit 108.

[0020] The drive unit 11 includes an AC signal generation unit 104 (AC signal generation means) and a boost circuit 105. The speed command unit 101 generates a target speed for each time period. The position command unit 102 generates a target position for each time period. The difference between the target position and the detected position detected by the position detection unit 14 is calculated as the position deviation.

[0021] Here, the target speed and target position are, for example, output as one command value per sampling period from each generation means. The sampling period refers to one cycle from the acquisition of the position deviation in Figure 1, through the output of the control variable, the application of AC voltage to the oscillator, the output of the detected speed and detected position, to just before the acquisition of the position deviation begins. In the above cycle, the position or speed of the vibrating actuator is feedback controlled.

[0022] Using the target velocity and position deviation, the first trained model 103 calculates and outputs the first control variable (phase difference, frequency). The first trained model 103, the second trained model 107, and the trained model 106 have the neural network (hereinafter also referred to as "NN") configuration shown in Figure 4. The NN consists of an input layer X layer, a hidden layer H layer, and an output layer Z layer. In this embodiment, the target velocity is set as input x1 and the position deviation as input x2 as input data, and the phase difference is set as output z1 and the frequency as output z2 as output data.

[0023] The input layer has two neurons (X1, X2), the hidden layer has seven neurons (H1, H2, ..., H7), and the output layer has two neurons (Z1, Z2). A common sigmoid function (Figure 4(b)) was used as the activation function. The number of neurons in the hidden layer is not limited to seven, but a range of 3 to 20 is preferable. This is because there is a trade-off between learning accuracy and learning speed: fewer neurons in the hidden layer decrease learning accuracy but learning converges faster (faster learning speed), while more neurons in the hidden layer improve learning accuracy but learning speed slows down. In addition, while sigmoid functions and ReLU (ramp function) are generally used as the activation function for the output layer, a linear function (Figure 4(c)) was used to accommodate the negative sign of the phase difference, which is the controlled variable.

[0024] The weight (first weight) connecting the input layer neuron (first neuron) and the hidden layer neuron (second neuron) was defined as wh. The threshold value for the hidden layer neuron (second neuron) was defined as θh. The weight (second weight) connecting the hidden layer neuron (second neuron) and the output layer neuron (third neuron) was defined as wo. The threshold value for the output layer neuron (third neuron) was defined as θo. The weights and threshold values ​​are those learned by the machine learning unit 12 described later. The trained neural network can be considered as a collection of common feature patterns extracted from the time-series data of the relative velocity and control quantity of the vibration-type actuator. Therefore, the output is a value obtained by a function with the weights and threshold values ​​as variables (parameters).

[0025] The second trained model 107 sets the detected speed detected by the speed detection unit 16 as x1 and the target deviation (zero) as x2 as input data, and sets the phase difference as z1 and the frequency as z2 as output data. Note that the target deviation may be an offset value other than zero. The control variable t (first control variable) output from the first trained model 103 is used as the ground truth data, and the difference between the control variable t and the control variable z (second control variable) output from the second trained model 107 is calculated. Then, error data based on this difference, the control deviation (tz), is input to the adaptive learning unit 108.

[0026] As will be described later, the adaptive learning unit 108 updates the parameters (weights and thresholds) of the neural network (NN) at each sampling period using stochastic gradient descent (SGD), which is a type of backpropagation method. Note that the parameter (weights and thresholds) are updated for both the first trained model 103 and the second trained model 107, and the same parameters are applied at the same time.

[0027] After the parameters are updated, the control quantity based on the updated parameters is output in the next sampling period, and the vibration actuator is controlled. Note that the update frequency does not necessarily have to be every sampling period; it may be a predetermined period, such as twice or three times the sampling period.

[0028] The first control variables (phase difference, frequency) output from the first trained model 103, which is a neural network (NN), are input to the AC signal generation unit 104, which controls the speed and driving direction of the vibration actuator. The AC signal generation unit 104 generates a two-phase AC signal based on the first control variables.

[0029] The boost circuit 105 includes, for example, a coil and a transformer. The AC voltage boosted to a desired drive voltage by the boost circuit 105 is applied to the piezoelectric element of the vibrator 131 to drive the contact body 132. An example of a vibratory actuator applicable to the present invention will be described with reference to the drawings. The vibratory actuator of this embodiment includes a vibrator and a contact body.

[0030] Figure 2 illustrates the driving principle of a linear-driven (direct-acting) vibratory actuator as an example of a vibratory actuator. The vibratory actuator 13 shown in Figure 2(a) has an elastic body 203 and a piezoelectric element 204, which is an electromechanical energy conversion element, bonded to the elastic body 203, and a contact body 132 driven by the vibrator 131. By applying an AC voltage to the piezoelectric element 204, two vibration modes as shown in Figures 2(c) and (d) are generated, causing the contact body 132, which makes pressurized contact with the projection 202, to move in the direction of the arrow.

[0031] Figure 2(b) shows the electrode pattern of the piezoelectric element 204. For example, the piezoelectric element 204 of the oscillator 131 has electrode regions that are divided into two equal parts in the longitudinal direction. Also, the polarization direction in each electrode region is the same direction (+). Of the two electrode regions of the piezoelectric element 204, the electrode region located on the right in Figure 2(b) is subjected to an AC voltage (VB), and the electrode region located on the left is subjected to an AC voltage (VA).

[0032] If VB and VA are AC voltages with frequencies near the resonant frequency of the first vibration mode and in phase, the entire piezoelectric element 204 (the two electrode regions) will expand at one moment and contract at another. As a result, the oscillator 131 will experience vibration of the first vibration mode as shown in Figure 2(c) (hereinafter referred to as "upward vibration"). This causes a displacement in the upward direction (Z direction) of the projection 202.

[0033] Furthermore, if VB and VA are AC voltages with frequencies near the resonant frequency of the second vibration mode and with a phase difference of 180°, then at a certain moment, the electrode region on the right side of the piezoelectric element 204 will contract while the electrode region on the left side will expand. At another moment, the opposite relationship will occur. As a result, the oscillator 131 will experience vibration of the second vibration mode (hereinafter referred to as "feed vibration") as shown in Figure 2(d). This causes a displacement in the projection 202 in the driving direction (feed direction, X direction).

[0034] Therefore, by applying an AC voltage having a frequency near the resonant frequencies of the first and second vibration modes to the electrodes of the piezoelectric element 204, a vibration resulting from the combination of the first and second vibration modes can be excited.

[0035] In this way, by combining the two vibration modes, the projection 202 performs elliptical motion in a cross-section perpendicular to the Y direction (a direction perpendicular to the X and Z directions) in Figure 2(d). This elliptical motion drives the contact body 132 in the direction of the arrow in Figure 2(a). The direction in which the contact body 132 and the vibrator 131 move relative to each other, that is, the direction in which the contact body 132 is driven by the vibrator 131 (here, the X direction), is referred to as the driving direction. Furthermore, the amplitude ratio R of the second vibration mode with respect to the first vibration mode (amplitude of feed vibration / amplitude of thrust vibration) can be changed by changing the phase difference of the two-phase AC voltage input to the two equally divided electrodes.

[0036] This vibration-type actuator makes it possible to change the speed of the contacting body by changing the amplitude ratio of the vibration.

[0037] In the above description, the case in which the vibrator 131 is stationary (fixed) and the contact body 132 is moving (driven) was described as an example, but the present invention is not limited to this embodiment. The contact body and the vibrator only need to have their relative contact points change. For example, the contact body may be stationary (fixed) and the vibrator may be moving (driven). In other words, in the present invention, "driven" means changing the relative position of the contact body with respect to the vibrator, and it is not necessarily required that the absolute position of the contact body (for example, the position of the contact body relative to the position of the housing enclosing the contact body and the vibrator) change.

[0038] In the above description, a linear-driven (direct-acting) vibration actuator was used as an example. That is, the example described was when the vibrator 131 or contact body 132 moves (is driven) in a linear direction, but the present invention is not limited to this embodiment. The contact body and vibrator only need to have their contact points relatively changing. For example, the vibrator and contact body may move in a rotational direction. An example of a vibration actuator in which the vibrator and contact body move in a rotational direction is a ring-type (rotating) vibration actuator having a ring-shaped vibrator.

[0039] Vibration actuators are used, for example, in driving the autofocus of cameras.

[0040] Figure 3 is a perspective view illustrating the lens drive mechanism of the lens barrel. The lens holder drive mechanism using a vibrating actuator comprises a vibrator, a lens holder, and a first guide bar and a second guide bar, which are arranged in parallel and slidably hold the lens holder.

[0041] In this embodiment, we will describe the case where the second guide bar is a contact body, the second guide bar is fixed, and the vibrator and lens holder move together as a single unit.

[0042] The oscillator generates a relative moving force between the oscillator and a second guide bar, which is in contact with the elastic projection, through the elliptical motion of the oscillator's projection, which is produced by the application of a driving voltage to the electromechanical energy conversion element. This allows the lens holder, which is fixed integrally with the oscillator, to move along the first and second guide bars.

[0043] Specifically, the contact body drive mechanism 300 mainly comprises a lens holder 302, a lens 306, a vibrator 131 to which a flexible printed circuit board is coupled, a pressure magnet 305, two guide bars 303 and 304, and a base body (not shown). Here, the vibrator 131 will be used as an example to explain the mechanism.

[0044] The first guide bar 303 and the second guide bar 304 are arranged parallel to each other, with both ends of each guide bar held and fixed by a base (not shown). The lens holder 302 has a cylindrical holder portion 302a, a holding portion 302b for holding and fixing the transducer 131 and the pressurizing magnet 305, and a first guide portion 302c that fits with the first guide bar 303 to act as a guide.

[0045] The pressurizing magnet 305, which constitutes the pressurizing means, has a permanent magnet and two yokes positioned at both ends of the permanent magnet. A magnetic circuit is formed between the pressurizing magnet 305 and the second guide bar 304, and an attractive force is generated between these members. The pressurizing magnet 305 is positioned at a distance from the second guide bar 304, and the second guide bar 304 is positioned to be in contact with the vibrator 131.

[0046] The aforementioned attractive force applies pressure between the second guide bar 304 and the transducer 131. Two protrusions on the elastic body make pressurized contact with the second guide bar 304 to form a second guide section. The second guide section forms a guide mechanism using magnetic attraction, and although external forces may cause the transducer 131 and the second guide bar 304 to be pulled apart, this is dealt with as follows.

[0047] In other words, the anti-detachment part 302d provided on the lens holder 302 contacts the second guide bar 304, thereby causing the lens holder 302 to return to the desired position.

[0048] By applying a desired AC voltage signal to the transducer 131, a driving force is generated between the transducer 131 and the second guide bar 304, and this driving force drives the lens holder 302.

[0049] A position sensor (not shown in Figure 3), attached to the second guide bar 304 or the vibrator 131, detects the relative position and relative velocity of the vibrator 131 or the second guide bar 304 with respect to the second guide bar 304 or the vibrator 131. The detected position and detected velocity are then output from the position sensor. The detected position is fed back to the control unit 10 as a position deviation, and the vibration actuator is feedback controlled to track the target position at each unit time. The detected velocity is input to the machine learning unit 12 and used as training data along with the control quantities (phase difference, frequency) output from the control unit 10. Training data is data consisting of pairs of input data and output data (correct data).

[0050] This embodiment describes a two-phase drive control device that drives a piezoelectric element, which is an electromechanical energy conversion element, in two phases. However, the present invention is not limited to two-phase drive and can also be applied to vibration actuators with two or more phases.

[0051] Next, the machine learning unit 12 will be described in detail. The learning model 106 has a neural network configuration (NN configuration) (see Figure 4) that takes the detected velocity from the velocity detection unit 16 and the target deviation as inputs and outputs the phase difference and frequency. The target deviation is a value set so that the position deviation follows when the contact body 132 is moved relative to the oscillator 131.

[0052] The target deviation, in this context, is a value with the same dimensions (data format) as the position deviation. The target deviation is typically set to zero, but an offset value may be provided to compensate for mechanical play.

[0053] Furthermore, the learning model 106 may also accept a speed deviation, which is the difference between the target speed and the detected speed, as input instead of the target deviation. We discovered that when a speed deviation is provided, motor characteristics that could not be obtained conventionally can be learned as a secondary effect. Specifically, when a speed deviation is provided, features corresponding to the frequency response (so-called transmission characteristics) of the vibration-type actuator were learned based on the relationship between the various vibration components included in the speed deviation and the control variable. Then, the weight values ​​and thresholds of the NN related to the input position deviation are learned to appropriate values, and the control system can be compensated.

[0054] The control variables (phase difference, frequency) output from the control unit 10 are used as ground truth data and are compared with the control variables output from the untrained or training model 106 to calculate the error. In this example, phase difference and frequency are used as control variables, but combinations of pulse width and frequency, or pulse width and phase difference can also be used as control variables. Furthermore, the output layer of the NN may consist of one or more neurons, and may be designed to allow free combination selection from phase difference, frequency, and pulse width.

[0055] Figure 5 is a flowchart of adaptive control using machine learning and trained models in this embodiment. In step 1 (S1), initial values ​​are set for the weights and thresholds of the first trained model 103 and the second trained model 107 of the control unit 10 based on a random function (untrained state). In step 2 (S2), the two untrained models (untrained NNs) control the vibration actuator.

[0056] In step 3 (S3), time-series data of the first control quantity (phase difference, frequency) output from the first trained model 103 and the relative velocity (detected velocity) detected by the velocity detection unit 16 are acquired as training data while the vibration actuator is being driven. In step 4 (S4), the control quantity of the training data is used as the ground truth data, and an optimization calculation is performed using machine learning with the training model 106. Optimization means adjusting the parameters of the NN so that the output from the NN based on the input to the NN approaches the training data, and is not limited to adjusting the parameters of the NN so that the output from the NN based on the input to the NN matches the training data. The training model 106 has the same NN configuration as the first trained model 103 and the second trained model 107 used for adaptive control. The weights and thresholds of the NN are optimized by machine learning, and the parameters of the first trained model 103 and the second trained model of the control unit 10 are updated. In step 5 (S5), adaptive control of the vibration actuator is performed using the first trained model 103 and the second trained model, whose weights and thresholds have been updated.

[0057] After adaptive control, the process returns to step 3 (S3) to acquire training data in order to respond to changes in driving conditions and temperature environment. Batch learning is performed as the method for acquiring training data, with training conducted while the drive is stopped.

[0058] Figure 9 is a timing chart illustrating batch learning in the machine learning unit 12 and adaptive control in the control unit. The horizontal axis represents time, and the vertical axis represents the target position pattern given as a command value for feedback control of the vibration actuator.

[0059] Figure 9(a) shows an example of batch learning performed while the vibration actuator is stopped (when the contact body is not moving relative to the vibrator). In this example, time-series data of the relative velocity (detected velocity) and the controlled variable detected during the operation period of the vibration actuator are acquired as training data, and the machine learning and NN parameters (weights, thresholds) are updated using the stop period. Note that it is not necessarily required to perform machine learning during every stop period; for example, it is also possible to perform learning only when changes in the temperature environment or driving conditions are detected.

[0060] Figure 9(b) shows an example of adaptive control performed while driving a vibrating actuator (when the contact body is moved relative to the vibrator), updating the parameters of the NNs of the first trained model 103 and the second trained model at the sampling period. Unlike conventional batch learning, the present invention uses two trained models simultaneously, allowing the model itself to automatically adapt to changes at the sampling period, even when the model and the actual machine characteristics do not match due to temperature environment or load fluctuations.

[0061] The machine learning described in Step 4 (S4) above will be further explained using Figure 6.

[0062] Figure 6 shows a flowchart of the process when Adam is used as the optimization algorithm for NN parameters. Steps 1-2 (S1-S2) are the same as those described in Figure 5, except that the controlled variable output from the NN is only the phase difference.

[0063] In step 3 (S3), the first control variable (n) and velocity (n), which are time-series learning data as shown in Figure 6(b), are acquired. The first control variable (n) and velocity (n) are measurement data when the vibration actuator 13 is controlled by an untrained model, and velocity (n) is the velocity detected by the velocity detection unit 16 when the vibration actuator 13 is driven by the first control variable (n). The number of samples n for each of the first control variable (phase difference) and velocity is 3400. This is actual measurement data when driven for 0.34 sec at a control sampling rate (hereinafter simply referred to as "sampling rate") (1 / sampling period) of 10 kHz.

[0064] Furthermore, training data does not necessarily need to be acquired at the sampling rate; decimation can save memory and shorten training time.

[0065] This invention takes velocity (n) as input to a learning model 106, and compares the output z(n), which is the result calculated (derived) by the learning model 106, with the first control variable (n), t(n), which corresponds to the correct data of the learning data. Then, it calculates the error e(n). Specifically, the error e(n) is given by error e(n) = (t(n) - z(n)). 2 So, in step 4 (S4), the error E (=Σe(n)=Σ(t(n)-z(n)) for 3400 items was calculated. 2 The first loop calculates the result, and the error gradients ∇E of the weights (wh, wo) and thresholds (θh, θo) are calculated respectively.

[0066] Next, using the error gradient ∇E, the parameters are optimized using Adam, one of the optimization algorithms, as shown below.

[0067]

number

[0068] wt is the parameter update amount, ∇E is the error gradient, vt is the moving average of the error gradient, st is the moving average of the square of the error gradient, η is the learning rate, and ε is the division by zero prevention constant.

[0069] The parameters used were η=0.001, β1=0.9, β2=0.999, and ε=10e-12. With each iteration of the optimization operation, the weights and thresholds are updated, and the output z(n) of the learned model approaches the control variable (n) of the ground truth data, so the error E decreases.

[0070] Figure 6(c) shows the progression of the error E based on the number of calculation loops. Note that other optimization methods may also be used.

[0071] Figure 7(a) compares the calculation results of Adam, RMSprop, Momentum, and SGD using the learning model of this embodiment and the measured training data. From the viewpoint of the number of calculations, stability, and final error, Adam yielded the best results.

[0072] Figure 7(b) shows an example of training a controlled variable (phase difference) using Adam. The output z of the trained model after the first loop differs significantly from the controlled variable t of the ground truth data. After repeating the calculation, the output z of the trained model after 5000 loops almost matches the controlled variable t of the ground truth data. In this training example, optimization was performed with 5000 loops, but it is desirable to adjust the number of loops appropriately depending on the convergence rate.

[0073] Next, the adaptive control in step 5 (S5), which is a feature of the present invention, will be explained using Figure 8. Figure 8 shows a flowchart of the adaptive control that operates in the control unit of the present invention. In step 4 (S4), the weights and thresholds of the NN learned in the machine learning unit are set as the initial values ​​for the first trained model 103 and the second trained model 107. Step 5 (S5) shows the adaptive control that is performed at each sampling period while the vibration actuator is being driven, and in S5, the parameters are updated using the two trained models.

[0074] The first trained model 103 is input with the target velocity and position deviation. The second trained model 107 is input with the detected velocity and target deviation. Here, the target deviation is set to 0 (zero), but other values ​​such as offset values ​​or velocity deviations may also be input. Next, the control of the vibration-type actuator for one sampling period (Δt) is performed using the control variable t (first control variable) calculated by the first trained model 103 (S5-1).

[0075] Using the first controlled variable t as the ground truth data, the control deviation (tz) is calculated as the difference between the first controlled variable t and the second controlled variable z (second controlled variable), which is calculated by the second trained model 107 and has the same data format as the first controlled variable t. The error data e(t) is then obtained (S5-2). Here, the second controlled variable z is the controlled variable output from the second trained model 107 within the same sampling period as the first controlled variable t. Next, the obtained error data e(t) = (t(t) - z(t)) 2 The error gradient ∇e is calculated using (S5-3).

[0076] Note that "t" in (t) represents a predetermined timing and is different from the first controlled variable t. The error gradient ∇e is obtained using the derivatives of the activation functions of the hidden and output layers of the NN and the input / output values ​​of each layer. Next, the weights and thresholds of the NN are calculated using stochastic gradient descent (SGD), one of the backpropagation methods (S5-4). Note that, in addition to SGD, other optimization algorithms such as the steepest descent method or Newton's method may also be used.

[0077] Finally, the weights and thresholds of the calculation results are updated at each sampling period (S5-5).

[0078] The weights and thresholds are updated for both the first trained model 103 and the second trained model 107, with the same parameters applied at the same time. Then, the process returns to S5-1 at the next sampling period, and this control loop is constantly repeated while the system is running. The parameters of the first trained model and the second trained model may be updated at a longer period than the sampling period, for example, at an integer multiple of the sampling period, rather than at the sampling period.

[0079] Figure 10 shows the results when feedback control is performed using an untrained model with a predetermined target position pattern, demonstrating the effect of adaptive control according to the present invention. Feedback control is performed with a predetermined target position pattern, and the target speed is a trapezoidal drive with a maximum of 50 mm / s, performing a reciprocating motion with a 5 mm stroke including positioning. The horizontal axis represents time (sec), and the vertical axis represents the target position in units of encoder pulses (left axis) and the position deviation in μm units (right axis).

[0080] Figure 10(a) shows the control results without adaptive control. Both use an untrained model in which the NN weights and thresholds are set using a random function. The present invention can significantly improve the controllability of an untrained model by performing adaptive control. Without adaptive control, as an untrained model, it is naturally unable to follow the target speed. The detection position deviates significantly from the target position, the position deviation is large, and it becomes uncontrollable. Therefore, it is difficult to obtain high-quality training data using a completely untrained model, and conventionally it was necessary to use a trained model with a certain level of performance guarantee, or to obtain training data using a different controller such as PID control. The adaptive control of the present invention can solve this problem, and training data can be obtained without problems even when using an untrained model.

[0081] Figure 10(b) shows the control results obtained by the adaptive control of the present invention. Immediately after driving, the detected speed fluctuates relative to the target speed, but it can be seen that it is tracking the target speed after only 30ms. This is because, as mentioned above, the parameters (weights, thresholds) of the two trained models are updated by the adaptive learning unit at each sampling period (0.1ms), and adaptive control is performed to approach the target speed and target position.

[0082] Figure 11 compares the results of feedback control performed with a predetermined target position pattern in the control and PID control methods of the present invention. The target speed is a trapezoidal drive with a maximum speed of 50 mm / s, and the pattern involves a reciprocating motion with a 5 mm stroke including positioning.

[0083] The horizontal axis represents time (sec), and the vertical axis, from top to bottom, represents the phase difference control amount (deg), detection speed (mm / s) and speed deviation (mm / s), target position (encoder pulse count: 8000 pls per mm) and position deviation (μm).

[0084] Figure 11(a) shows the results obtained by PID control. Here, the frequency of the AC signal applied to the vibration actuator 13 (driving frequency) was fixed at 93 kHz, and control was performed using only the phase difference. Of the speed deviation and position deviation, the position deviation in particular tends to be larger in the acceleration and deceleration regions. This is because the inertia of the driven body driven by the vibration actuator has an effect. It can also be seen that it takes a long time to reach a complete stop. It is possible to reduce the position deviation by further increasing the PID gain, but in order to ensure robustness against changes in driving conditions (used frequency range 91 kHz to 95 kHz) and ambient temperature, a PID control gain with a certain gain margin and phase margin was applied.

[0085] Figure 11(b) shows the results of adaptive control using the trained models of the present invention (first trained model 103 and second trained model 107). Similarly, the drive frequency of the AC signal supplied to the vibration actuator was fixed at 93 kHz, and adaptive control was performed using a trained model that outputs a phase difference. It was found that the position deviation was improved in all regions during acceleration, deceleration, and stopping and settling by applying the present invention.

[0086] Figure 11(c) shows the result of adaptive control using a trained model that outputs phase difference and frequency. The drive frequency starts at 93kHz, and the control variable is manipulated in parallel with the phase difference. The frequency control variable is an unsigned absolute value, and the drive frequency is always manipulated towards the lower frequency side with respect to 93kHz. When there are two or more control variables as in this example, there are countless combinations of phase difference and frequency that can be obtained in machine learning to achieve a predetermined speed. Therefore, the learning result will vary depending on the characteristics of the training data acquired.

[0087] In this embodiment, the parameters of the neural network (NN) were set using a random function, and the NN parameters that showed the best results in terms of position deviation and power were selected by comparing multiple learning results. Alternatively, for example, the ratio of frequency to phase difference could be defined before learning. It was found that by using phase difference and frequency as control variables, the speed range of the vibration-type actuator could be expanded, and the position deviation could be improved by PID control. Note that Figures 11(b) and 11(c) show the change in the phase difference of a virtual PID control (top row of Figure 11(b) and top row of Figure 11(c)). This is the output of a control variable for which PID calculations were performed based on the position deviation, although it is not directly used to control the vibration-type actuator.

[0088] By utilizing these PID control variables, it is possible to detect anomalies in the control of the trained model. Specifically, by comparing the control variable output by the trained model with the PID control variable, if there is a large deviation from a predetermined range, it is possible to predict that the NN parameters have deviated from normal values, and the parameters can be reset. Although this function is not an essential configuration for obtaining the effects of the present invention, it can improve reliability from the perspective of guaranteeing the performance of adaptive control using trained models.

[0089] The above describes the configuration of the control device of the present invention. The control unit 10 and the machine learning unit 12 are composed of, for example, digital devices such as a CPU and a PLD (including an ASIC), and elements such as an A / D converter. The AC signal generation unit 104 of the drive unit 11 has, for example, a CPU, a function generator and a switching circuit, and the boost circuit is composed of, for example, a coil, a transformer and a capacitor. The control unit and the drive unit are not limited to being composed of one element or circuit, but may be composed of multiple elements or circuits. Furthermore, each process may be executed by any of the elements or circuits.

[0090] A CPU may be a processor or device that executes instructions to perform the actions shown in the flowchart. Instructions may be stored in memory or circuitry (not shown). Memory may be volatile or non-volatile. Examples of memory include random-access memory (RAM), read-only memory (ROM), or flash devices.

[0091] Figure 12 shows a control device for a vibration actuator in another embodiment of the present invention. The difference from the configuration in Figure 1 is that it does not include a machine learning unit, and control is performed only by the control unit 10. Of course, by performing batch learning in response to changes in ambient temperature and driving conditions using the machine learning unit, robust and good control performance can be obtained. However, even with control performed only by the control unit as in this embodiment, it is possible to perform control that follows a certain degree of target speed and target position, as shown in the results in Figure 10.

[0092] Figure 13 shows a control device for a vibratory motor actuator in another embodiment of the present invention. The difference from the configuration in Figure 12 is that the input to the first trained model 103 is the target speed and speed deviation, rather than the target speed and position deviation.

[0093] Figure 14 shows a control device for a vibratory motor actuator in another embodiment of the present invention. The difference from the configuration in Figure 13 is that the input to the first trained model 103 is only the target speed.

[0094] (Second embodiment) Another embodiment of the control unit shown in Figure 1 of the present invention will be described. Figure 16 is a control block diagram of a vibration-type drive device in a second embodiment (when a learned model and a PID controller are used in combination). In this control block, the position feedback control of the vibration-type actuator 13 is performed using a PID controller 1401, a first learned model 103, and a second learned model 107, etc. The position deviation is input to the PID controller 1401, and the result of the PID calculation is output from the PID controller 1401. Note that, instead of a PID controller, for example, a P controller, PI controller, PD controller, etc. are also applicable. The target speed and the PID-calculated position deviation are input to the first learned model 103.

[0095] By applying this embodiment, the gain of the position deviation input to the first trained model 103 can be adjusted, thereby enabling finer adjustments to the control system.

[0096] Figure 17 is a control block diagram of a vibration-type drive device in another embodiment of the second embodiment.

[0097] In this control block, velocity feedback control of the vibration actuator 13 is performed using a PID controller 1401, a first trained model 103, and a second trained model 107. The velocity deviation is input to the PID controller 1401, and the result of the PID calculation is output from the PID controller 1401. In addition to the PID controller, other controllers such as a P controller, PI controller, or PD controller can also be used. The first trained model 103 is input to the target velocity and the velocity deviation calculated using the PID calculation.

[0098] By applying this embodiment, the gain of the velocity deviation input to the first trained model 103 can be adjusted, thereby enabling finer adjustment of the control system.

[0099] (Third embodiment) Another embodiment of the control unit shown in Figure 1 of the present invention will be described.

[0100] Figure 18 is a control block diagram of the third embodiment (when a trained model and a PID controller are used in combination). In this control block diagram, the position feedback control of the vibration actuator 13 is performed by adding the output from the PID controller 1501 and the control variable (first control variable) output from the first trained model 103. The position deviation is input to the PID controller 1501, and the phase difference and frequency calculated using PID are output.

[0101] Note that configurations other than PID controllers are also acceptable; for example, P control, PI control, and PD control can also be applied. Furthermore, a phase compensator may be placed after the PID controller. The first trained model 103 receives the target velocity and position deviation as input, although the position deviation may be set to zero. The first trained model 103 outputs a phase difference and frequency, and these are added together, respectively, to the phase difference and frequency output from the PID arithmetic unit 1501. Alternatively, a machine learning unit 12 may be provided to perform machine learning using the added control quantities and the detected velocity detected by the velocity detection unit 16.

[0102] By applying this embodiment, gain adjustment of position deviation can be performed using PID control, allowing for finer adjustment of the control system. Furthermore, since it can be compared with the control results of the PID controller, anomalies in adaptive control can be detected, and the reliability of the control device can be guaranteed.

[0103] Figure 19 is a control block diagram of a vibration-type drive device in another embodiment of the third embodiment. In this control block diagram, the speed feedback control of the vibration-type actuator 13 is performed by adding the output from the PID controller 1501 and the control variable (first control variable) output from the first learned model 103. The speed deviation is input to the PID controller 1501, and the phase difference and frequency calculated using PID are output.

[0104] Note that configurations other than PID controllers are also acceptable; for example, P control, PI control, and PD control can also be applied. Furthermore, a phase compensator may be placed after the PID controller. The first trained model 103 receives the target speed and speed deviation as input, although the speed deviation may be set to zero. The first trained model 103 outputs the phase difference and frequency, and these are added together, respectively, to the phase difference and frequency output from the PID arithmetic unit 1501.

[0105] Alternatively, a machine learning unit 12 may be provided to perform machine learning using the added control amount and the detected speed detected by the speed detection unit 16.

[0106] By applying this embodiment, gain adjustment of the speed deviation can be performed using PID control, allowing for finer adjustment of the control system. Furthermore, since it can be compared with the control results of the PID controller, anomalies in adaptive control can be detected, and the reliability of the control device can be guaranteed.

[0107] (Fourth embodiment) In previous embodiments, if the control device has a first control variable output unit having a first trained model and a second control variable output unit having a second trained model, the machine learning unit may be removed from the control device. Although such a control device has the disadvantage that the first and second trained models cannot be subjected to machine learning again, in vibration-type drive devices where there is little need to perform machine learning again, there is the advantage that the configuration is simplified by removing the learning unit.

[0108] In previous embodiments, the system may include a memory unit that stores the parameters (first weight, second weight, second neuron threshold, and third neuron threshold) that the trained model possessed. The trained model may then be machine-learned by replacing the parameters of the trained model with the parameters stored in the memory unit.

[0109] Furthermore, in the above embodiments, an environmental sensor for detecting environmental conditions may be included. The trained model may then be trained using machine learning when the environmental sensor detects a change in the environment. The environmental sensor may be at least one of a temperature sensor and a humidity sensor.

[0110] (Fifth embodiment) In the first embodiment, an example was described in which the control device for the vibration actuator is used to drive the autofocus lens (driven object) of the imaging device, but the applications of the present invention are not limited to this. For example, as shown in Figure 20, it can also be used to drive the lens (driven object) or image sensor (driven object) during image stabilization. Figure 20(a) is a plan view (top view) showing the external appearance of the imaging device 60. Figure 20(b) is a schematic diagram of the internal configuration of the imaging device 60.

[0111] The imaging device 60 is broadly composed of a main body 61 and a lens barrel 62 that is detachably attached to the main body 61. The main body 61 includes an image sensor 63, such as a CCD sensor or CMOS sensor, which converts the optical image formed by light passing through the lens barrel 62 into an image signal, and a camera control microcontroller 64 that controls the overall operation of the imaging device 60. Multiple lenses L, such as a focus lens and a zoom lens, are arranged in predetermined positions in the lens barrel 62. The lens barrel 62 also incorporates an image blur correction device 50, which has a disc member 56 and a vibrator 131 provided on the disc member 56, and an image blur correction lens 65 is positioned in a hole formed in the center of the disc member 56. The image blur correction device 50 is positioned so that the image blur correction lens 65 can be moved in a plane perpendicular to the optical axis of the lens barrel 62. In this case, by driving the transducer 131 using the control device 15 of the present invention, the transducer 131 and the disc member 56 move relative to the contact body 132 fixed to the lens barrel, and the image blur correction lens 65 (driven body) is driven.

[0112] Furthermore, the control device of the present invention can also be used to drive a lens holder (driven object) for moving a zoom lens. Therefore, the control device of the present invention can be mounted not only on the imaging device but also on interchangeable lenses for driving the lens (driven object).

[0113] Furthermore, the control device for the vibration actuator shown in the first embodiment can also be used to drive a stage (driven object). For example, as shown in Figure 21, it can be mounted on an automated stage of a microscope.

[0114] The microscope in Figure 21 includes an imaging unit 70 containing an image sensor and optical system, and an automatic stage 71 mounted on a base and having a stage 72 (driven object) that is moved by a vibrating actuator. The object to be observed is placed on the stage 72, and a magnified image is captured by the imaging unit 70. If the observation range is wide, the stage 72 is moved by driving the vibrating actuator using the control device 15 of the first or second embodiment. This moves the object to be observed in the X and Y directions in the figure, and multiple images are acquired. The acquired images are combined using a computer (not shown) to obtain a single high-resolution image with a wide observation range. [Explanation of symbols]

[0115] 10 Adaptive Control Unit (Control Unit) 11 Drive unit 12 Machine Learning Department 13. Vibration-type actuator (vibration-type motor) 14 Position detection unit (position detection means) 15 Control device 16 Speed ​​detection unit (speed detection means) 101 Speed ​​command section (speed command means) 102 Position command unit (position command means) 103 First trained model (first control variable output unit) 104 AC signal generation unit (AC signal generation means) 105 Boost Circuit 106 Learning Model (Third Control Variable Output Unit) 107 Second trained model (second control variable output unit) 108 Adaptive Learning Department 131 Oscillator 132 Contact body

Claims

1. A control device for a vibration-type actuator that moves a contact body in contact with the vibrator relative to the vibrator by vibration generated in the vibrator, A first control variable output unit having a first trained model that has been trained by machine learning to output a first control variable for moving the contact body relative to the vibrator when a first speed for moving the contact body relative to the vibrator is input, The control unit includes a second control variable output unit having a second trained model that has been machine-trained to output a second control variable, which is data in the same data format as the first control variable, when a second velocity detected when the contact body is moved relative to the oscillator by the first control variable is input, A control device for a vibration-type actuator, characterized in that the parameters of the first trained model and the parameters of the second trained model are updated based on a control deviation, which is the difference between the first controlled variable and the second controlled variable output within the same sampling period as the first controlled variable.

2. The first trained model described above is The first speed and, The machine learning system is configured to output the first control variable when a value based on a position deviation, which is the difference between a first position for moving the contact body relative to the vibrator and a second position detected when the contact body is moved relative to the vibrator by the first control variable, is input. The second pre-trained model described above is The second speed and, The control device for a vibration-type actuator according to claim 1, characterized in that it is machine-trained to output the second control amount when a first target deviation, which is data in the same data format as the value based on the position deviation, is input.

3. The first trained model described above is The first speed and, When a value based on the velocity deviation, which is the difference between the first velocity and the second velocity, is input, the system is trained to output the first control variable. The second pre-trained model described above is The second speed and, The control device for a vibration-type actuator according to claim 1, characterized in that it is machine-trained to output the second control amount when a second target deviation, which is data in the same data format as the value based on the speed deviation, is input.

4. The control device for a vibration actuator according to any one of claims 1 to 3, characterized in that the parameters of the first trained model and the parameters of the second trained model are updated based on the control deviation, the error gradient obtained based on the control deviation, and the error gradient and an optimization algorithm.

5. The control device for a vibration-type actuator according to any one of claims 1 to 4, characterized in that the parameters of the first trained model and the parameters of the second trained model are updated at a period that is an integer multiple of the sampling period.

6. It has a PID controller, The control device for a vibration-type actuator according to claim 2, characterized in that the value based on the position deviation is the position deviation, or the output from the PID controller when the position deviation is input to the PID controller.

7. It has a PID controller, The control device for a vibration-type actuator according to claim 3, characterized in that the value based on the speed deviation is the speed deviation, or the output from the PID controller when the speed deviation is input to the PID controller.

8. The first trained model and the second trained model have a neural network configuration comprising an input layer having one or more first neurons, a hidden layer having multiple second neurons, and an output layer having one or more third neurons. The control device for a vibration-type actuator according to any one of claims 1 to 7, characterized in that the parameters of the first trained model and the parameters of the second trained model include a plurality of first weights connecting the first neuron and the second neuron, a plurality of second weights connecting the second neuron and the third neuron, a threshold for the second neuron, and a threshold for the third neuron.

9. The system includes a machine learning unit having a third control variable output unit that has a learning model that outputs a third control variable in the same data format as the first control variable when the second speed is input, The parameters of the learning model are updated by machine learning based on time-series data of the control deviation, which is the difference between the first control variable and the second control variable detected within the same sampling period as the first control variable. The control device for a vibration-type actuator according to claim 1, characterized in that the parameters of the first trained model and the parameters of the second trained model are updated by the parameters of the trained model.

10. The system includes a machine learning unit having a learning model that outputs a third control variable in the same data format as the first control variable when the second velocity and a third target deviation in the same data format as the position deviation are input. The parameters of the learning model are updated by machine learning based on time-series data of the control deviation, which is the difference between the first control variable and the second control variable detected within the same sampling period as the first control variable. The control device for a vibration-type actuator according to claim 2, characterized in that the parameters of the first trained model and the parameters of the second trained model are updated by the parameters of the trained model.

11. The system includes a machine learning unit having a learning model that outputs a third control variable in the same data format as the first control variable when the second velocity and a fourth target deviation in the same data format as the velocity deviation are input. The parameters of the learning model are updated by machine learning based on time-series data of the control deviation, which is the difference between the first control variable and the second control variable detected within the same sampling period as the first control variable. The control device for a vibration-type actuator according to claim 3, characterized in that the parameters of the first trained model and the parameters of the second trained model are updated by the parameters of the trained model.

12. The control device for a vibration-type actuator according to claim 9, characterized in that the parameters of the learning model are machine-learned based on time-series data of the control deviation, time-series data of the error gradient obtained from the time-series data of the control deviation, and the time-series data of the error gradient and an optimization algorithm.

13. The control device for a vibratory actuator according to any one of claims 9 to 12, characterized in that the parameters of the learning model are updated by machine learning in cases other than when the contact body is moved relative to the vibrator.

14. The control device for a vibratory actuator according to any one of claims 9 to 12, characterized in that the parameters of the learning model are updated by machine learning when the contact body is moved relative to the vibrator.

15. It has an environmental sensor that detects the environmental conditions, The control device for a vibration-type actuator according to any one of claims 9 to 14, characterized in that the parameters of the learning model are updated by machine learning when the environmental sensor detects a change in the environment.

16. The control device for a vibration-type actuator according to claim 15, characterized in that the environmental sensor is at least one of a temperature sensor and a humidity sensor.

17. It has a storage unit for storing the parameters of the learning model, The control device for a vibration-type actuator according to any one of claims 9 to 16, characterized in that the parameters of the learning model are updated by machine learning by replacing them with the parameters of the learning model that have been updated by machine learning and are stored in the memory unit.

18. The learning model comprises an input layer having one or more first neurons, a hidden layer having multiple second neurons, and an output layer having one or more third neurons. It has a neural network configuration that has, The control device for a vibration-type actuator according to any one of claims 9 to 17, characterized in that the parameters of the learning model include a plurality of first weights connecting the first neuron and the second neuron, a plurality of second weights connecting the second neuron and the third neuron, a threshold for the second neuron, and a threshold for the third neuron.

19. The control device for a vibration-type actuator according to claim 4 or 12, characterized in that the optimization algorithm is one of Adam, Momentum, RMSprop, and SGD.

20. The control device for a vibrating actuator according to any one of claims 1 to 19, characterized in that the first control variable has at least one of phase difference, frequency, and pulse width.

21. The speed command unit that commands the first speed, A control device for a vibration-type actuator according to any one of claims 1 to 20, comprising a speed detection unit for detecting the second speed.

22. The position command unit that commands the first position, The control device for a vibration-type actuator according to claim 2, further comprising a position detection unit for detecting the second position.

23. The vibrator has a drive unit that outputs an AC voltage that generates vibrations, The control device for a vibration-type actuator according to any one of claims 1 to 22, characterized in that the first control amount is input to the drive unit.

24. A drive unit outputs an AC voltage that generates vibrations to the vibrator, It has a PID controller, The control device for a vibration-type actuator according to claim 2, characterized in that the drive unit receives the sum of the first control amount and the output from the PID controller when the position deviation is input to the PID controller.

25. A drive unit outputs an AC voltage that generates vibrations to the vibrator, It has a PID controller, The control device for a vibration-type actuator according to claim 3, characterized in that the drive unit receives the sum of the first control amount and the output from the PID controller when the speed deviation is input to the PID controller.

26. A vibration-type actuator that moves a contact body that is in contact with the vibrator relative to the vibrator by vibrations generated in the vibrator, A vibration-type drive device characterized by comprising a control device for a vibration-type actuator according to any one of claims 1 to 25.

27. The vibration-type drive device according to claim 26, A replacement lens characterized by having a lens that is driven by moving the contact body relative to the vibrator.

28. The vibration-type drive device according to claim 26, An imaging device characterized by having an image sensor driven by moving the contact body relative to the vibrator.

29. The vibration-type drive device according to claim 26, An automatic stage characterized by having a stage that is driven by moving the contact body relative to the vibrator.

Images