Control system, learning device, inference device, and controller

Abstract

The control system includes a detection unit that detects the position of a workpiece moving on a conveyance path, a drive unit that drives a processing unit that performs processing on the workpiece, a controller (12) that controls the drive unit, and a remote I / O that outputs a detection signal from the detection unit to the controller (12). The controller (12) includes a Kalman filter (23) that estimates the future position of the workpiece based on the detection signal, and a timing generation unit (25) that generates processing timing by the processing unit based on the estimated future position of the workpiece. The processing timing is transmitted to the drive unit via the remote I / O.

Description

Technical Field

[0001] The present disclosure relates to a control system, a learning device, an inference device, and a controller.

Background Art

[0002] In a cutting process of an object to be processed in an apparatus composed of an unwinding shaft and a winding shaft such as a packaging machine or a printing machine, after an optical sensor reads a mark called a registration mark pre-printed on the object to be processed moving at high speed, cutting is performed by a device with a motor after passing a certain position. Here, since the reading position and the cutting position are separated, it is necessary to estimate the future position based on the device state at the time when the registration mark is detected and read.

[0003] Patent Document 1 discloses disposing a mark detector on the upstream side of a rotating member that cuts a conveyed web member, transmitting a detection signal of the mark detector to an arithmetic control unit, calculating a feed length from a speed command and the mark detection signal, and adjusting the phase of the timing of the cutting process.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] However, conventionally, the estimation of the future position is performed by linear extrapolation. Therefore, in an acceleration / deceleration region other than the constant speed region, the estimation accuracy is poor, and there is a drawback that the cut position is deviated. In addition, since the filter cannot be freely switched, an optimal extrapolation method cannot be selected according to different conditions, and there is a drawback that the quality of the produced product cannot be improved.

[0006] The present disclosure has been made in view of the above circumstances, and an object thereof is to obtain an estimation accuracy of a future position that exceeds linear extrapolation in an acceleration / deceleration region.

Means for Solving the Problem

[0007] To achieve the above object, the control system of the present disclosure includes a detection unit that detects the position of a workpiece moving on a conveyance path, a drive unit that drives a processing unit that performs processing on the workpiece, a controller that controls the drive unit, and a remote I / O that outputs a detection signal from the detection unit to the controller. The controller includes a Kalman filter that estimates the future position of the workpiece based on the detection signal, and a timing generation unit that generates processing timing by the processing unit based on the estimated future position of the workpiece. A linear extrapolator that estimates the future position of the workpiece by linear extrapolation, and a switch that selects either the Kalman filter or the linear extrapolator and outputs the estimated future position of the workpiece to the timing generation unit. The control system includes a remote I / O that transmits the processing timing to the drive unit via the remote I / O.

Advantages of the Invention

[0008] According to the present disclosure, by using a Kalman filter that estimates the future position of the workpiece, it is possible to obtain an estimation accuracy of a future position that exceeds linear extrapolation in an acceleration / deceleration region.

Brief Description of the Drawings

[0009]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Embodiments for Carrying Out the Invention

[0010] (First Embodiment) Hereinafter, the control system 100 according to the first embodiment of the present disclosure will be described with reference to the drawings. In each drawing, the same or equivalent parts are denoted by the same reference numerals.

[0011] FIG. 1 is a configuration diagram showing a control system 100 according to the first embodiment. The control system 100 includes an unwinding shaft 1, a winding shaft 2, a workpiece 3 which is a sheet-like member such as wrapping paper, printing paper, or film conveyed between the unwinding shaft 1 and the winding shaft 2, a cutter 4 for cutting the workpiece 3, a cylindrical rotor 5 provided with the cutter 4 and rotating along the conveyance direction on its outer peripheral surface, a servo motor 6 for rotationally driving the rotor 5, a servo amplifier 7 for driving the servo motor 6, a mark sensor 9 for reading a mark 8 provided on the surface of the workpiece 3 and generating a calculation start signal for the cut timing, an encoder 10 for detecting the movement of the workpiece 3, a remote I / O 11 for receiving the detection signals of the mark sensor 9 and the encoder 10, and a controller 12 for estimating the cutting position of the workpiece 3 and controlling the cutting. The controller 12, the remote I / O 11, and the servo amplifier 7 are connected by a communication line 13. Note that the cutter 4 is an example of the processing unit of the present disclosure, the servo motor 6 is an example of the driving unit of the present disclosure, and the mark sensor 9 and the encoder 10 are examples of the detection unit of the present disclosure.

[0012] The workpiece 3 is wound around the unwinding shaft 1 in a roll shape, and the workpiece 3 is fed out by the rotation of the winding shaft 2. The winding shaft 2 winds up the workpiece 3 fed out from the unwinding shaft 1. The workpiece 3 fed out from the unwinding shaft 1 is conveyed toward the rotor 5 provided with the cutter 4 by the rotation of the unwinding shaft 1 and the winding shaft 2.

[0013] At the ends of the surface of the workpiece 3, black marks 8 are provided at regular lengths in the conveying direction. The marks 8 indicate the cutting positions of the workpiece 3, and the dotted lines indicate the cutting lines when cutting with the cutter 4 from the marks 8. Note that the cutting lines are shown in the drawings for explanation purposes and are not provided on the actual surface of the workpiece 3.

[0014] The cutter 4 cuts the workpiece 3 and is provided parallel to the rotation axis on the outer peripheral surface of the rotor 5. The rotor 5 rotates along the conveying direction of the workpiece 3 at a position where the outer peripheral surface faces the workpiece 3 by the rotation of the servo motor 6. When the cutter 4 reaches a position facing the workpiece 3 due to the rotation of the rotor 5, the cutter 4 abuts against the workpiece 3 and cuts the workpiece 3.

[0015] The servo motor 6 is driven by the servo amplifier 7, and the servo amplifier 7 controls the servo motor 6 according to the stored operation parameters. In this embodiment, a configuration is disclosed in which the workpiece 3 is cut by rotating the rotor 5 with the cutter 4 using a rotary motor. However, the present invention is not limited to this, and for example, a configuration in which the cutter 4 is moved parallel along the conveying direction of the workpiece 3 by a linear motor may be used.

[0016] The remote I / O 11 receives signals input from the mark sensor 9 and the encoder 10. The mark sensor 9 is an optical sensor that is disposed above the conveyance path of the workpiece 3, away from the upstream side of the conveyance path from the rotor 5, and detects a mark 8 provided on the surface of the workpiece 3. The encoder 10 is disposed on the conveyance path of the workpiece 3 and continuously outputs pulses according to the movement of the workpiece 3. The remote I / O 11 communicates with the controller 12 and operates according to instructions sent from the controller 12. The remote I / O 11 transmits signals input from the mark sensor 9 and the encoder 10 to the controller 12. Further, the remote I / O 11 outputs a signal to the servo amplifier 7 according to an instruction from the controller 12.

[0017] The controller 12 is, for example, a PLC (Programmable Logic Controller). The controller 12 acquires the current position of the workpiece 3 by receiving the detection signals of the mark sensor 9 and the encoder 10. Based on the acquired current position, the controller 12 obtains the predicted future position of the workpiece 3, predicts the position where the workpiece 3 is to be cut, and outputs a signal instructing the cut timing of the workpiece 3 to the remote I / O 11.

[0018] Fig. 2 shows a functional configuration representing the internal processing of the controller 12. The controller 12 includes a pulse counter 21 that counts encoder pulses, a switch 22 that switches the output direction of the count value of the pulse counter 21, a Kalman filter 23 and a linear extrapolator 24 that are estimators for predicting the future position of the workpiece 3, and a timing generation unit 25 that generates a cut timing signal for cutting the workpiece 3.

[0019] The pulse counter 21 receives the mark detection signal that is the output of the mark sensor 9 and the encoder pulses that are the output of the encoder 10 from the remote I / O 11. The pulse counter 21 counts the encoder pulses with the timing when the mark detection signal is received as the start timing, acquires a current position signal indicating the current position of the workpiece 3, and outputs it to the switch 22.

[0020] The switch 22 selects and outputs the output destination of the input current position signal. A Kalman filter 23 and a linear extrapolator 24 are respectively connected as the output destinations of the switch 22. The switch 22 switches the output destination to the Kalman filter 23 or the linear extrapolator 24 according to the switching signal. The output timing of the switching signal is preset by the user.

[0021] The Kalman filter 23 and the linear extrapolator 24 estimate and calculate the future position of the workpiece 3 based on the current position signal input from the switch 22. The linear extrapolator 24 connects the current position and the past position with a straight line and estimates the future position by linearly extrapolating using the straight line. Linear extrapolation has the advantages of simple and fast calculation and not requiring much resources. However, in the case of a data pattern that changes curvilinearly, the future position cannot be accurately estimated. In contrast, the Kalman filter 23 can accurately estimate the future position even for a data pattern that changes curvilinearly.

[0022] FIG. 3 shows the details of the Kalman filter 23. The Kalman filter 23 includes a delay unit 31, a predictor 32, and a smoother 33. The delay unit 31 holds the state estimate value of the previous sampling for one sampling period and outputs it for the current sampling. Here, the state estimate value includes the smoothed value and the covariance matrix of the smoothing error. The predictor 32 calculates the predicted value and the covariance matrix of the prediction error using the smoothed value and the covariance matrix of the smoothing error obtained from the delay unit 31 one sampling period ago. The predicted value and the covariance matrix of the prediction error calculated here are output to the smoother 33. The smoother 33 acquires, as observation data, the current position signal obtained from the mark detection signal and the encoder pulse, and calculates the filter gain, the smoothed value, and the covariance matrix of the smoothing error. In calculating these filter gain, smoothed value, and covariance matrix of the smoothing error, in addition to the current position signal obtained from the mark detection signal and the encoder pulse, the processing result of the predictor 32 is used, and further, the processing result of the smoother 33 is returned to the delay unit 31. The future position predictor 34 calculates the future position after N samplings by extrapolation from the smoothed position and the smoothed velocity vector obtained from the smoother 33.

[0023] The workpiece 3 moving on the conveyance path moves at a speed according to a predetermined speed map. Here, as the speed map, for example, an acceleration period in which it is accelerated until it reaches a predetermined speed in the period immediately after the start of movement, a constant speed period in which the speed is maintained when the predetermined speed is reached, and a deceleration period in which the speed is decelerated are each provided. Here, the waveforms of the estimated values calculated using the Kalman filter 23 and the linear extrapolator 24 in the acceleration period are shown in FIG. 4. In FIG. 4, a shows the estimated value calculated using the linear Kalman filter as the Kalman filter 23, and b shows the estimated value calculated using the linear extrapolator 24. From this, in the region with acceleration and deceleration, the Kalman filter 23 generates a smoother estimated result than the estimated value calculated using the linear extrapolator 24. Therefore, in the acceleration and deceleration periods in which the position of the workpiece 3 changes curvilinearly, the future position can be accurately estimated by using the Kalman filter 23. On the other hand, in the constant speed period in which the position of the workpiece 3 changes linearly, a highly accurate estimation can also be performed using the linear extrapolator 24. Therefore, in the constant speed period, by using the linear extrapolator 24, the future position can be accurately estimated while considering the calculation speed, resource load, etc. Therefore, the user makes a setting to select the Kalman filter 23 in the acceleration period in which the workpiece 3 moves while accelerating and the deceleration period in which it moves while decelerating in setting the output timing of the switching signal, and makes a setting to select the linear extrapolator 24 in the constant speed period in which the workpiece 3 moves at a constant speed. Note that the acceleration and deceleration periods and the constant speed period may be determined by referring to the speed map, or may be determined by obtaining the speed, acceleration, or jerk from the encoder pulses.

[0024] The timing generation unit 25 generates the target timing for cutting the workpiece 3 based on the future predicted position information estimated by the Kalman filter 23 or the linear extrapolator 24. Specifically, the timing generation unit 25 calculates the arrival time to the cutting position of the workpiece 3 by dividing the estimated future predicted position by the conveyance speed of the workpiece 3, and further calculates after how many clocks the arrival time will be. The generated cut timing signal is output to the servo amplifier 7 through the remote I / O 11. The servo amplifier 7 controls the rotation of the rotor 5 based on the cut timing signal, and cuts the workpiece 3 with the cutter 4.

[0025] The controller 12, in terms of hardware, as shown in FIG. 5, includes a processor 41 that processes data according to a control program, a main memory unit 42 that functions as a work area for the processor, an auxiliary storage unit 43 for storing data over a long period, an input unit 44 that receives data inputs, an output unit 45 that outputs data, a communication unit 46 that communicates with other devices, a display unit 47, and a bus that interconnects these elements. The auxiliary storage unit 43 stores a control program for controlling the cut timing executed by the processor. The input unit 44 receives signals input from the mark sensor 9 and the encoder 10 transmitted from the remote I / O 11 and provides them to the processor 41.

[0026] (Second Embodiment) In order to estimate the future position more accurately, the set values of parameters such as the mechanical model used in the Kalman filter 23 and the Kalman gain need to match the physical quantities of the actual machine. In general, it is difficult to identify them for complex machines and they may need to be obtained experimentally. In this embodiment, by learning with the position, speed, and acceleration of the workpiece 3 as inputs, the optimal parameters of the Kalman filter 23 are obtained with the cutting accuracy as the reward. The learning is performed by machine learning of reinforcement learning (Q-learning).

[0027] FIG. 6 is a configuration diagram of a machine learning device related to the control system 100. The learning device 201 includes a data acquisition unit 202 and a model generation unit 203.

[0028] The data acquisition unit 202 acquires, as learning data, the parameters of the Kalman filter 23 as action A, and the position, velocity, and acceleration of the workpiece 3 as state S.

[0029] The model generation unit 203 learns the parameters of the Kalman filter 23 for obtaining the most accurate cutting result as the optimal action A based on the learning data including the parameters of the Kalman filter 23 as action A and including the position, velocity, and acceleration of the workpiece 3 as state S. That is, a learned model is generated that infers the parameters of the Kalman filter 23, which is the optimal action A, from the position, velocity, and acceleration of the workpiece 3, which is the state S of the control system 100.

[0030] As the learning algorithm used by the model generation unit 203, known algorithms such as supervised learning, unsupervised learning, and reinforcement learning can be used. As an example, the case of applying reinforcement learning will be described. In reinforcement learning, an agent (acting entity) in a certain environment observes the current state (parameters of the environment) and determines the action to be taken. The environment changes dynamically due to the agent's action, and the agent is given a reward according to the change in the environment. The agent repeats this and learns the action policy that can obtain the most rewards through a series of actions. As typical methods of reinforcement learning, Q-learning and TD-learning are known. For example, in the case of Q-learning, the general update formula for the action value function Q(s,a) is represented by Equation (1).

[0031]

Equation

[0032] In Equation (1), s t represents the state of the environment at time t, and a t represents the action at time t. Due to the action a t , the state becomes s t+1It changes to r t+1 represents the reward obtained by the change of that state, γ represents the discount rate, and α represents the learning coefficient. Note that γ is in the range of 0 < γ ≤ 1, and α is in the range of 0 < α ≤ 1. When the action A is the action a t becomes, and the state S becomes the state s t becomes, and the best action a t at the state s t at time t is learned.

[0033] In the update formula represented by Equation (1), if the action value Q of the action a with the highest Q value at time t + 1 is greater than the action value Q of the action a executed at time t, the action value Q is increased. Conversely, if the opposite is the case, the action value Q is decreased. In other words, the action value function Q(s, a) is updated so that the action value Q of the action a at time t approaches the best action value at time t + 1. As a result, the best action value in a certain environment is sequentially propagated to the action values in the previous environment.

[0034] As described above, when generating a learned model by reinforcement learning, the model generation unit 203 includes a reward calculation unit 204 and a function update unit 205.

[0035] The reward calculation unit 204 calculates the reward based on the parameters of the Kalman filter 23 that is the action A, and the position, speed, and acceleration of the workpiece 3 that is the state S. The reward calculation unit 204 calculates the reward r based on the cutting result of the workpiece 3 as the reward criterion, that is, the error range of how far the position where the workpiece 3 is cut is from the predetermined optimal cutting position. Therefore, the cutting result of the workpiece 3 being within the reference error range is used as the reward criterion. In the case of the reward increase criterion where the cutting result of the workpiece 3 is within the reference error range, the reward r is increased (for example, a reward of "1" is given). On the other hand, in the case of the reward decrease criterion where the cutting result of the workpiece 3 exceeds the reference error range, the reward r is reduced (for example, a reward of "-1" is given).

[0036] The function update unit 205 updates a function for determining an optimal action A according to the reward calculated by the reward calculation unit 204, and outputs it to the learned model storage unit 206. For example, in the case of Q-learning, the action value function Q(s t ,a t ) represented by Equation (1) is used as a function for calculating the optimal action A.

[0037] The above learning is repeatedly executed. The learned model storage unit 206 stores the action value function Q(s t ,a t ) updated by the function update unit 205, that is, stores the learned model.

[0038] Next, with reference to FIG. 7, the process of learning by the learning device 201 will be described. FIG. 7 is a flowchart regarding the learning process of the learning device 201.

[0039] In step S11, the data acquisition unit 202 acquires, as learning data, the parameters of the Kalman filter 23 which is the action A, and the position, speed, and acceleration of the workpiece 3 which is the state S.

[0040] In step S12, the model generation unit 203 calculates a reward based on the action A and the state S. Specifically, the reward calculation unit 204 acquires the action A and the state S, and determines whether to increase or decrease the reward based on a predetermined reward criterion.

[0041] When the reward calculation unit 204 determines to increase the reward (step S12: Yes), it increases the reward in step S13. On the other hand, when the reward calculation unit 204 determines to decrease the reward (step S12: No), it decreases the reward in step S14.

[0042] In step S15, the function update unit 205 updates the action value function Q(s t ,a t ) represented by Equation (1) stored in the learned model storage unit 206 based on the reward calculated by the reward calculation unit 204.

[0043] The learning device 201 repeatedly executes the steps from S11 to S15 above, and stores the generated action value function Q(s t , a t ) as a learned model.

[0044] The learning device 201 according to the present embodiment stores the learned model in the learned model storage unit 206 provided outside the learning device 201. However, the learned model storage unit 206 may be provided inside the learning device 201.

[0045] FIG. 8 is a configuration diagram of the inference device 301 related to the control system 100. The inference device 301 includes a data acquisition unit 302 and an inference unit 303.

[0046] The data acquisition unit 302 acquires the position, speed, and acceleration of the workpiece 3 in the state S.

[0047] The inference unit 303 infers the parameters of the Kalman filter 23, which is the optimal action A, using the learned model. That is, by inputting the state S acquired by the data acquisition unit 302 into this learned model, the optimal action A suitable for the state S can be inferred. Here, the state S input here is data including the current position, speed, and acceleration of the workpiece 3.

[0048] In the present embodiment, it has been described that the optimal action A is output using the learned model learned by the model generation unit 203 of the learning device 201 related to the control system 100. However, a learned model may be acquired from another control system 100, and the optimal action A may be output based on this learned model.

[0049] Next, with reference to FIG. 9, the process for obtaining the optimal action A, that is, the parameters of the optimal Kalman filter 23, using the learning device 201 and determining the optimal cutting position will be described.

[0050] In step S21, the data acquisition unit 302 acquires the state S.

[0051] In step S22, the inference unit 303 inputs the state S into the learned model stored in the learned model storage unit 206 to obtain the optimal action A. The inference unit 303 outputs the obtained optimal action A to the controller 12 of the control system 100 (step S23).

[0052] In step S24, the controller 12 determines the optimal cutting position of the workpiece 3 by using the output optimal action A, that is, the parameters of the Kalman filter 23 that can obtain a highly accurate cutting result.

[0053] In addition, in this embodiment, the case where reinforcement learning is applied to the learning algorithm used by the inference unit 303 has been described, but it is not limited thereto. For the learning algorithm, in addition to reinforcement learning, supervised learning, unsupervised learning, semi-supervised learning, etc. can also be applied.

[0054] Also, as the learning algorithm used in the model generation unit 203, deep learning that learns the extraction of the parameters of the Kalman filter 23 itself can be used, and machine learning can also be performed according to other known methods, such as neural networks, genetic programming, functional logic programming, support vector machines, etc.

[0055] Note that the learning device 201 and the inference device 301 may be connected to the controller 12 of the control system 100 via a network, for example, and may be devices separate from this control system 100. Also, the learning device 201 and the inference device 301 may be built into the controller 12. Furthermore, the learning device 201 and the inference device 301 may exist on a cloud server.

[0056] Further, the model generation unit 203 may learn the optimal action A using the learning data acquired from the plurality of control systems 100. Note that the model generation unit 203 may acquire the learning data from the plurality of control systems 100 used in the same area, or may learn the optimal action A using the learning data collected from the plurality of control systems 100 operating independently in different areas. Also, it is possible to add or remove the control system 100 that collects the learning data midway. Furthermore, the learning device 201 that has learned the optimal action A for a certain control system 100 may be applied to another control system 100 different from this, and the optimal action A for the other control system 100 may be relearned and updated.

[0057] In the above embodiment, since the model of the system is linear, a linear Kalman filter is used as the Kalman filter 23. On the other hand, when dealing with a non-linear model, by using a non-linear Kalman filter as the Kalman filter 23, accurate position estimation is possible. Also, when it is unknown whether the model is linear or non-linear, by using an extended Kalman filter as the Kalman filter 23, position estimation corresponding to both models is possible. Further, these plurality of Kalman filters may be provided as the Kalman filter 23 and switched by the switch 22. Also, this switching may be determined by obtaining the speed, acceleration, or jerk from the encoder pulses.

[0058] In the above embodiment, the machining of the workpiece 3 has been described as cutting, but it is not limited to this, and for example, it may be applied to printing on the workpiece 3.

[0059] The present disclosure can be variously embodied and modified without departing from the broad spirit and scope of the present disclosure. Also, the above-described embodiments are for explaining the present disclosure and do not limit the scope of the present disclosure. That is, the scope of the present disclosure is indicated by the claims rather than the embodiments. And various modifications made within the scope of the claims and within the scope of the meaning of the disclosure equivalent thereto are considered to be within the scope of the present disclosure.

Industrial Applicability

[0060] The present disclosure can be widely applied to a control system that controls a processing position for performing processing on a workpiece based on the position of the workpiece moving on a conveyance path.

Explanation of Signs

[0061] 1 Pay-out shaft, 2 Take-up shaft, 3 Workpiece, 4 Cutter, 5 Rotor, 6 Servo motor, 7 Servo amplifier, 8 Mark, 9 Mark sensor, 10 Encoder, 11 Remote I / O, 12 Controller, 13 Communication line, 21 Pulse counter, 22 Switch, 23 Kalman filter, 24 Linear extrapolator, 25 Timing generator, 31 Delayer, 32 Predictor, 33 Smoother, 34 Future position predictor, 41 Processor, 42 Main storage unit, 43 Auxiliary storage unit, 44 Input unit, 45 Output unit, 46 Communication unit, 47 Display unit, 100 Control system, 201 Learning device, 202 Data acquisition unit, 203 Model generation unit, 204 Reward calculation unit, 205 Function update unit, 206 Learned model storage unit, 301 Inference device, 302 Data acquisition unit, 303 Inference unit.

Claims

1. A detection unit that detects the position of a workpiece moving along a transport path, A drive unit that drives a processing unit that performs processing on the workpiece, A controller that controls the aforementioned drive unit, It includes a remote I / O that outputs a detection signal from the detection unit to the controller, The aforementioned controller, A Kalman filter that estimates the future position of the workpiece based on the detection signal, A timing generation unit that generates a machining timing by the machining unit based on the estimated future position of the workpiece, It includes a mechanism that transmits the processing timing to the drive unit via the remote I / O, Control system.

2. The aforementioned controller, A linear extrapolator that estimates the future position of the workpiece by linear extrapolation, The system includes a switch that selects either the Kalman filter or the linear extrapolator and outputs the estimated future position of the workpiece to the timing generation unit. The control system according to claim 1.

3. The aforementioned controller, The switch is instructed to select the Kalman filter when the workpiece is moving at an accelerating or decelerating speed, and to select the linear extrapolator when the workpiece is moving at a constant speed. The control system according to claim 2.

4. The aforementioned controller, The movement state of the workpiece is determined from the speed, acceleration, or jerk of the workpiece, and a selection is instructed to the switch. The control system according to claim 3.

5. The workpiece is a sheet-like member, The processing unit is a cutter for cutting the sheet-like member. The control system according to any one of claims 1 to 4.

6. The detection unit is a mark detection sensor that generates a mark detection signal by detecting a mark provided on the workpiece. The control system according to any one of claims 1 to 4.

7. A data acquisition unit that acquires training data including the position, velocity, and acceleration of a workpiece and the parameters of a Kalman filter in the control system of claim 1, A model generation unit generates a trained model for inferring the parameters of a Kalman filter to obtain the most accurate cutting result from the position, velocity, and acceleration of the workpiece using the aforementioned training data. A learning device equipped with the following features.

8. A data acquisition unit in the control system of claim 1 that acquires the current position, velocity, and acceleration of a workpiece, An inference unit that outputs the parameters of a Kalman filter from the current position, velocity, and acceleration of a workpiece, using a learning model to infer the parameters of a Kalman filter for obtaining the most accurate cutting result from the position, velocity, and acceleration of the workpiece. An inference device equipped with the following features.

9. A detection signal from a detection unit that detects the position of a workpiece moving along a transport path is input via remote I / O, and a controller that controls a drive unit that drives a processing unit to perform processing on the workpiece, A Kalman filter that estimates the future position of the workpiece based on the detection signal, A timing generation unit that generates a machining timing by the machining unit based on the estimated future position of the workpiece, It includes a mechanism that transmits the processing timing to the drive unit via the remote I / O, controller.

Images