Machine learning device and control device

Abstract

A machine learning device according to the present disclosure comprises: a state observation unit that acquires shape data indicating a shape to which a plate-shaped workpiece is to be cut by an industrial machine provided with a machining head to which a gap sensor is attached, data related to a path of interest, which is a movement path along which the machining head included in the industrial machine moves over a cut part of the workpiece that has been cut by the cutting, a gap amount when the machining head is closest to the workpiece when moving along the path of interest, and a retracted height of the machining head when moving along the path of interest; a training data creating unit that creates training data to be used for machine learning training on the basis of the data acquired by the state observation unit; and a learning unit that performs machine learning processing based on the training data to generate a trained model for estimating the retracted height of the machining head on the basis of the shape data and the path of interest.

Description

Machine learning device and control device

[0001] The present disclosure relates to a machine learning device and a control device.

[0002] Conventionally, there are industrial machines that process sheet metal, such as laser cutting machines, plasma cutting machines, and gas cutting machines. These industrial machines emit energy from the processing head, such as laser, plasma, or gas. When cutting sheet metal into a closed shape, the cut part may not fall but may get caught on something, causing it to tilt and protrude from the sheet metal. If the processing head interferes with this protruding part, it can cause the processing head to malfunction.

[0003] A conventional technique for avoiding this interference is to retract the machining head in the vertical direction of the plate workpiece when positioning it, and then return it to the next machining position. For example, see Patent Document 1.

[0004] Japanese Patent Application Publication No. 07-016773

[0005] When avoiding interference, if the tool is retracted to an excessive retraction height, unnecessary machining time and energy are required. Therefore, technology that can determine the most efficient retraction amount is desired for interference avoidance operations.

[0006] The machine learning device disclosed herein solves the above problem by using machine learning to learn the correct retraction height from shape data related to machining and the retraction height corresponding to the movement path of the machining head.

[0007] One aspect of the present disclosure is a machine learning device that includes: shape data indicating the shape of a plate-shaped workpiece cut by an industrial machine equipped with a machining head having a gap sensor attached; data related to a target path, which is a movement path along which a machining head equipped on the industrial machine moves over the cut portion of the workpiece cut by the cutting; data related to a target path, which is a movement path along which the machining head provided on the industrial machine moves over the cut portion of the workpiece cut by the cutting; a learning data creation unit that creates learning data to be used in machine learning learning processing based on the data acquired by the state observation unit; and a learning unit that performs machine learning processing based on the learning data and generates a trained model for estimating the retraction height of the machining head based on the shape data and the target path.

[0008] FIG. 1 is a schematic hardware configuration diagram of a machine learning device according to a first embodiment of the present disclosure. FIG. 2 is a block diagram showing the schematic functions of a machine learning device according to the first embodiment. FIG. 3 is a schematic diagram showing an example of a path along which a machining head moves over a workpiece. FIG. 4 is a schematic diagram showing a state in which a machining head is retracted. FIG. 5 is a block diagram showing the schematic functions of a machine learning device according to a second embodiment. FIG. 6 is a block diagram showing the schematic functions of a machine learning device according to a third embodiment. FIG. 7 is a schematic hardware configuration diagram of a control device according to a fourth embodiment of the present disclosure. FIG. 8 is a block diagram showing the schematic functions of a control device according to the fourth embodiment.

[0009] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. In the following description, components having the same or similar functions will be denoted by the same reference numerals. Duplicate descriptions of those components may be omitted.

[0010] In this application, "based on XX" means "based on at least XX," and includes cases where it is based on other elements in addition to XX. Furthermore, "based on XX" is not limited to cases where XX is used directly, but also includes cases where it is based on XX that has been calculated or processed. "XX" is any element (for example, any information).

[0011] First Embodiment Fig. 1 is a schematic hardware configuration diagram showing the main parts of a machine learning device according to a first embodiment of the present disclosure. The machine learning device 1 according to this embodiment operates in learning mode to create a trained model to be used by a control device 3. The machine learning device 1 according to this embodiment can be implemented on a computer such as a personal computer, a cell computer, a fog computer 6, or a cloud server 7. This embodiment shows an example in which the machine learning device 1 is implemented on a computer connected via a network to a control device that controls industrial machinery, the fog computer 6, the cloud server 7, or the like.

[0012] The CPU 11 included in the machine learning device 1 according to this embodiment is a processor that provides overall control of the machine learning device 1. The CPU 11 reads a system program stored in the ROM 12 via the bus 22 and controls the entire machine learning device 1 in accordance with the system program. The RAM 13 temporarily stores temporary calculation data, display data, various data acquired from the outside, and the like.

[0013] The non-volatile memory 14 is composed of, for example, a battery-backed memory (not shown) or an SSD (Solid State Drive), and retains its stored state even when the machine learning device 1 is powered off. The non-volatile memory 14 stores programs and data read from an external device 72 via the interface 15, programs and data input via the input device 71, and programs and data acquired from the control device 3 that controls the industrial machine 4 or other devices via the network 5. The programs and data stored in the non-volatile memory 14 may be expanded into the RAM 13 when executed / used. In addition, various system programs such as known analysis programs are written in the ROM 12 in advance.

[0014] The interface 15 is an interface for connecting the CPU 11 of the machine learning device 1 to an external device 72 such as a USB device. For example, system programs, data, etc. are read from the external device 72. In addition, programs, data, etc. created or edited within the machine learning device 1 can be stored in external storage means via the external device 72.

[0015] The interface 20 is an interface for connecting the CPU 11 of the machine learning device 1 to a wired or wireless network 5. The network 5 may communicate using technologies such as serial communication such as RS-485, Ethernet (registered trademark), optical communication, wireless LAN, Wi-Fi (registered trademark), Bluetooth (registered trademark), etc. The machine learning device 1, a control device 3 that controls at least one industrial machine 4, a fog computer 6, a cloud server 7, etc. may be connected to the network 5.

[0016] The display device 70 displays data and the like obtained as a result of executing various data and programs loaded into the memory, output via the interface 17. An input device 71, which is composed of at least one input device such as a keyboard, a pointing device, a voice input device, an imaging device, etc., passes commands, data, etc. based on user operations to the CPU 11 via the interface 18.

[0017] The industrial machine 4 is a processing machine such as a laser processing machine, a plasma processing machine, or a gas processing machine. These processing machines are equipped with a processing head. The control device 3 controls the processing head to move along a predetermined path commanded by a processing program or the like. When processing, the control device 3 controls the processing head to output, for example, a laser. The workpiece is then cut using the laser. During processing, a gap sensor provided in the processing head measures the distance between the workpiece and the processing head. Gap control is then performed based on the measured distance. The machine learning device 1 is configured to be able to acquire data such as the processing program used for control, coordinate values ​​of the processing head of the industrial machine 4, and the distance between the workpiece and the processing head from the control device 3, for example, via a network 5. When the machine learning device 1 is operating in learning mode, the control device 3, which controls the industrial machine 4, may retract the processing head using, for example, a processing head retraction function according to conventional technology.

[0018] 2 is a schematic block diagram illustrating functions of the machine learning device 1 according to the first embodiment of the present disclosure. Each function of the machine learning device 1 according to this embodiment is realized by the CPU 11 included in the machine learning device 1 shown in FIG. 1 executing a system program and controlling the operation of each unit of the machine learning device 1.

[0019] The machine learning device 1 of this embodiment includes a state observing unit 100, a learning data creating unit 110, a learning unit 140, and an output unit 180. In addition, the RAM 13 to the nonvolatile memory 14 of the machine learning device 1 are provided in advance with a model storage unit 160, which is an area for storing the trained model created by the learning unit 140.

[0020] The state observing unit 100 acquires, as state data, shape data indicating the shape of the workpiece to be cut by the industrial machine 4, and data related to a target path, which is a movement path along which a machining head provided in the industrial machine 4 moves over the cut portion of the workpiece cut by the cutting process. The state observing unit 100 also acquires the gap amount when the machining head comes closest to the workpiece when moving along the target path, and the retraction height of the machining head when moving along the target path.

[0021] The shape data acquired by the state observing unit 100 may be data related to a machining path that forms a closed shape for cutting a workpiece. The machining path data may be series data of coordinate values ​​of the relative machining path between the workpiece and the machining head, as instructed by the machining program. Such data is obtained by analyzing the machining program executed by the control device 3. The coordinate values ​​included in the machining path may be relative coordinate values ​​based on a predetermined position of the closed shape. The coordinate values ​​included in the machining path may also be machine coordinate values ​​that indicate an absolute position on the industrial machine 4. Furthermore, the coordinate values ​​may also be coordinate values ​​that indicate a position related to characteristics unique to the industrial machine 4. Examples of such coordinate values ​​include relative coordinate values ​​with respect to a predetermined position, such as a support table that supports the workpiece. Note that these coordinate values ​​may be used as machining path data as is, or may be converted into coordinate values ​​in a continuous vector format and used as machining path data. In the latter format, the same shape can be represented by the same data regardless of the orientation of the machining shape on the workpiece.

[0022] The target path of the machining head acquired by the state observing unit 100 may be data relating to the movement path of the machining head moving over the cut portion of the workpiece acquired as shape data. The data relating to the movement path included in the target path may be series data of coordinate values ​​including at least the coordinate values ​​of the position where the machining head starts moving within the closed shape of the machining path associated with the shape data and the coordinate values ​​of the position where the movement within the closed shape ends when each path is projected onto the XY plane. Such data is obtained by analyzing the machining program executed by the control device 3. The coordinate values ​​included in the target path may be relative coordinate values ​​based on a predetermined position on the closed shape. They may also be machine coordinate values ​​indicating an absolute position on the industrial machine 4. They may also be coordinate values ​​indicating a position related to characteristics unique to the industrial machine 4. Examples of such coordinate values ​​include relative coordinate values ​​with respect to a predetermined position, such as a support table that supports the workpiece.

[0023] FIG. 3 is a schematic diagram showing an example of a path along which a machining head moves over a workpiece. FIG. 3 illustrates an example of machining a plate-shaped workpiece 500 while the machining head moves over it. In FIG. 3, dotted arrows indicate paths along which the machining head moves without machining the workpiece. Solid arrows indicate paths along which the machining head moves while machining the workpiece. Each path in FIG. 3 shows how the machining head moves as seen from the workpiece when the control device that controls the industrial machine 4 executes the machining program. The machining program includes blocks that indicate commands to the control device 3 and the industrial machine 4, and each block is assigned a sequence number starting with N1. The characters written near each path in FIG. 3 indicate the sequence number that commands the movement along that path. In the example of FIG. 3, the machining head moves over the workpiece 500 in response to a command from block N3, and one closed shape is machined in response to commands from blocks N4 to N6. As a result, the workpiece at the portion indicated by reference numeral 501 is cut. After that, the machining head moves to the next machining start position in response to a command from block N7, and the next closed shape is machined in response to commands from blocks N8 to N11. This cuts the workpiece at the portion indicated by reference numeral 502. Then, the machining head moves outside the workpiece in response to a command from block N12.

[0024] At this time, the state observing unit 100 acquires the machining path along which the machining head moves in response to commands from blocks N4 to N6 as first shape data. It also acquires the machining path along which the machining head moves in response to commands from blocks N8 to N11 as second shape data. Meanwhile, blocks N7 and N12 pass over the cut portion 501 formed by the machining paths N4 to N6 when viewed from the positive direction of the Z axis. Therefore, the state observing unit 100 acquires, as data related to the first target path, the coordinate values ​​of the position where the machining head starts moving (the start position of the movement by block N7) and the coordinate values ​​of the position where the movement ends (the position where the movement path of N7 intersects with the machining path of N5) over the cut portion 501 formed by the machining paths N4 to N6 on the path along which the machining head moves in response to commands from block N7. In addition, the coordinate values ​​of the position where the movement of the machining head starts over the cutting portion 501 on the path along which the machining head moves in response to instructions from block N12 (the position where the movement path of N12 intersects with the machining path of N5) and the coordinate values ​​of the position where the movement ends (the position where the movement path of N12 intersects with the machining path of N4) are obtained as data relating to the second target path.

[0025] The gap amount obtained by the state observing unit 100 when the machining head is closest to the workpiece is measured when the machining head moves along the target path. FIG. 4 is a schematic diagram showing the state in which the machining head retracts. In FIG. 4, a plate-shaped workpiece 500 is supported on a support base 510 installed on a table 505. Also, a cutting portion 501 has been cut off from the workpiece 500 after a predetermined processing. FIG. 4 shows the state in which the machining head 520, to which a gap sensor 521 is attached, moves above the workpiece 500 and above the cutting portion 501. That is, during this movement, the machining head moves along the target path. At this time, the gap sensor 521 attached to the machining head 520 detects the smallest gap amount Gap when the cutting portion 501 passes over the portion protruding above the workpiece. min The state observing unit 100 detects this gap amount Gap min is acquired as the gap amount when the machining head provided on the industrial machine 4 comes closest to the workpiece when the machining head moves along the target path.

[0026] The retract height of the machining head when moving along the target path, which is acquired by the state observing unit 100, can be acquired while the machining head is moving along the target path. In the example of FIG. 4, the gap sensor 521 attached to the machining head 520 detects the smallest gap amount Gap min Then, the gap amount Gap min At the timing when the tip position P of the processing head 520 is detected, i The coordinate value of (x i , y i , z i At this time, the state observing unit 100 detects the position P i Z coordinate value z i is acquired as the withdrawal height of the machining head when moving on the target path.

[0027] The state observation unit 100 outputs the shape data thus acquired, data relating to the target path, data relating to the gap amount when the machining head is closest to the workpiece, and the retraction height of the machining head when moving along the target path to the learning data creation unit 110.

[0028] The learning data creation unit 110 creates learning data to be used for learning by the learning unit 140 based on the shape data acquired by the state observation unit 100, data related to the target path, the gap amount when the machining head is closest to the workpiece, and the retraction height of the machining head when moving along the target path. The learning data may be, for example, data that the learning unit 140 calculates to learn the correlation of data calculated based on the data related to the gap amount and the retraction height for the shape data and the data related to the target path. When such learning is performed by the learning unit 140, the learning data creation unit 110 creates training data using, for example, the data related to the shape data and the target path as input data and data related to an appropriate retraction height created based on the gap amount and the retraction height when the machining head is closest to the workpiece as label data. The learning data may also be data that evaluates the retraction height for the shape data and the data related to the target path based on the data related to the gap amount. When such learning is performed by the learning unit 140, the learning data creation unit 110 creates learning data in which, for example, shape data and data related to a focus path are used as input data, the retraction height is used as action data, and data related to the gap amount when the machining head is closest to the workpiece is used as judgment data to be used for a predetermined judgment. The learning data created by the learning data creation unit 110 is designed in accordance with the learning algorithm performed by the learning unit 140 and the model used. The learning data creation unit 110 outputs the created learning data to the learning unit 140.

[0029] The learning unit 140 performs machine learning based on the learning data created by the learning data creation unit 110 to generate a trained model. The model used by the learning unit 140 may be, for example, a model using a convolutional neural network or the like. It may also be an encoder-decoder model. It may also be a model such as a multilayer perceptron or a recurrent neural network. The learning unit 140 stores the generated trained model in the model storage unit 160.

[0030] The output unit 180 outputs the trained model generated by the learning unit 140. The output unit 180 may be configured to output data estimated based on the trained model generated by the learning unit 140. The output unit 180 may be configured to output the trained model based on an operation instruction from a user. The output unit 180 may be configured to output the trained model to, for example, an external device 72. The output unit 180 may also be configured to output the trained model to another computer such as a fog computer 6, a cloud server 7, or a control device 3 via the network 5.

[0031] The machine learning device 1 having the above configuration generates a predetermined trained model based on shape data indicating the shape of the cut portion, data related to a target path along which the machining head moves over the cut portion, data related to the gap amount when the machining head moves over the target path, and data related to the retraction height of the machining head when moving over the target path. This trained model is used to estimate the appropriate retraction height when the machining head moves over the cut portion. This is expected to enable retraction operations to be performed at a retraction height that is neither too high nor too low for the cut portion, thereby reducing processing time. Furthermore, when the cut portion has a shape that will not fall, the estimated appropriate retraction height is a position that is the height of the workpiece surface plus a predetermined margin, so retraction operations are not performed, thereby also reducing processing time.

[0032] Second Embodiment A machine learning device according to a second embodiment of the present disclosure will be described below. The machine learning device 1 according to this embodiment learns relationships between state data using supervised learning. The machine learning device 1 according to this embodiment has the same hardware configuration as the machine learning device 1 according to the first embodiment.

[0033] 5 is a schematic block diagram illustrating functions of the machine learning device 1 according to the second embodiment of the present disclosure. Each function of the machine learning device 1 according to this embodiment is realized by the CPU 11 included in the machine learning device 1 shown in FIG. 1 executing a system program and controlling the operation of each unit of the machine learning device 1.

[0034] The machine learning device 1 of this embodiment includes a state observing unit 100, a learning data creating unit 110, a learning unit 140, and an output unit 180. In addition, the RAM 13 to the nonvolatile memory 14 of the machine learning device 1 are provided in advance with a model storage unit 160, which is an area for storing the trained model created by the learning unit 140.

[0035] Similar to the state observing unit 100 according to the first embodiment, the state observing unit 100 according to the present embodiment acquires shape data indicating the shape of the workpiece to be cut by the industrial machine 4, data relating to a target path which is a movement path along which a machining head provided in the industrial machine 4 moves over the cut portion of the workpiece cut by the cutting process, data relating to a gap amount when the machining head is closest to the workpiece when moving along the target path, and data relating to a retraction height of the machining head when moving along the target path. The state observing unit 100 outputs the acquired shape data, data relating to the target path, data relating to the gap amount when the machining head moving along the target path is closest to the workpiece, and data relating to a retraction height of the machining head moving along the target path to the learning data creation unit 110.

[0036] The learning data creation unit 110 according to this embodiment creates learning data to be used for learning by the learning unit 140, based on the shape data acquired by the state observation unit 100, data related to the target path, data related to the gap amount when the machining head comes closest to the workpiece when moving along the target path, and data related to the retraction height of the machining head when moving along the target path. The learning data creation unit 110 includes a retraction height calculation unit 112.

[0037] The retraction height calculation unit 112 calculates a target retraction height based on the gap amount when the machining head is closest to the workpiece when moving along the noted path, which is acquired by the state observation unit 100, and the retraction height of the machining head when moving along the noted path. The target retraction height is a retraction height that is calculated based on the acquired data and is considered to be neither too much nor too little. The target retraction height may be, for example, the retraction height of the machining head when moving along the noted path, minus the gap amount when the machining head is closest to the workpiece when moving along the noted path, and then add a predetermined margin to the result. The target retraction height calculated in this way is a height that can be moved to a position on the noted path that is away from the plate material of the workpiece by a predetermined margin in the part that protrudes from the plate material of the workpiece. For example, in the example of FIG. 4, the retraction height calculation unit 112 calculates a target retraction height based on the position P i Z coordinate value z i From the minimum gap amount Gap min and adding a predetermined margin m (z i -Gap min + m) to the target evacuation height z ap It is calculated as follows.

[0038] The learning data creation unit 110 then creates training data using the shape data acquired by the state observing unit 100 and the data related to the focused route as input data, and the target evacuation height calculated by the evacuation height calculation unit 112 as label data. The learning data creation unit 110 outputs the created learning data to the learning unit 140.

[0039] The learning unit 140 according to this embodiment performs supervised learning using a predetermined model based on the learning data created by the learning data creation unit 110. Then, when shape data and data related to the focus route are input, a trained model is generated that outputs an estimated appropriate evacuation height that is neither excessive nor insufficient. The model used by the learning unit 140 may be, for example, a convolutional neural network model. It may also be an encoder-decoder model. It may also be a model such as a multilayer perceptron or a recurrent neural network. By using such a model, a model that estimates an appropriate evacuation height can be generated from the input shape data and data related to the focus route. The learning unit 140 stores the generated trained model in the model storage unit 160.

[0040] The output unit 180 according to this embodiment outputs the trained model generated by the training unit 140. The output unit 180 may output the trained model based on an operation instruction from a user. The output unit 180 may output the trained model to, for example, an external device 72. The output unit 180 may also output the trained model to another computer, such as the fog computer 6, the cloud server 7, or the control device 3, via the network 5.

[0041] The machine learning device 1 having the above configuration generates a trained model that estimates an appropriate retraction height based on shape data indicating the cut shape and data related to the target path moving over the cut portion. Using this trained model, it is expected that retraction operations can be performed at a retraction height that is neither too high nor too low for the cut portion, thereby shortening the processing time. When the cut portion has a shape that will not fall, the estimated appropriate retraction height is a position that is the height of the workpiece surface plus a predetermined margin, so no retraction operation is performed, thereby shortening the processing time.

[0042] Third Embodiment A machine learning device according to a third embodiment of the present disclosure will be described below. The machine learning device 1 according to this embodiment learns relationships between state data using reinforcement learning. The machine learning device 1 according to this embodiment has the same hardware configuration as the machine learning device 1 according to the first embodiment.

[0043] 6 is a schematic block diagram illustrating functions of the machine learning device 1 according to the third embodiment of the present disclosure. Each function of the machine learning device 1 according to this embodiment is realized by the CPU 11 of the machine learning device 1 shown in FIG. 1 executing a system program and controlling the operation of each unit of the machine learning device 1.

[0044] The machine learning device 1 of this embodiment includes a state observing unit 100, a learning data creating unit 110, a learning unit 140, and an output unit 180. In addition, the RAM 13 to the nonvolatile memory 14 of the machine learning device 1 are provided in advance with a model storage unit 160, which is an area for storing the trained model created by the learning unit 140.

[0045] Like the state observing unit 100 according to the first embodiment, the state observing unit 100 according to the present embodiment acquires shape data indicating the shape of the workpiece to be cut by the industrial machine 4, data related to a target path which is a movement path along which a machining head provided in the industrial machine 4 moves over the cut portion of the workpiece cut by the cutting process, the gap amount when the machining head comes closest to the workpiece when moving along the target path, and the retraction height of the machining head when moving along the target path. The state observing unit 100 outputs the acquired shape data, data related to the target path, data related to the gap amount when the machining head comes closest to the workpiece when moving along the target path, and data related to the retraction height of the machining head when moving along the target path to the learning data creation unit 110.

[0046] The learning data creation unit 110 according to this embodiment creates learning data to be used for learning by the learning unit 140 based on the shape data, data related to the noted path, data related to the gap amount when the machining head is closest to the workpiece when moving along the noted path, and data related to the retraction height when the machining head moves along the noted path, acquired by the state observing unit 100. The learning data creation unit 110 creates learning data in which the shape data and data related to the noted path acquired by the state observing unit 100 are used as state data indicating the state of the industrial machine 4, data related to the retraction height when the machining head moves along the noted path as action data, and data related to the gap amount when the machining head is closest to the workpiece when moving along the noted path as judgment data. The learning data creation unit 110 outputs the created teacher data to the learning unit 140.

[0047] The learning unit 140 according to this embodiment performs reinforcement learning using a model related to a predetermined value function, based on the learning data created by the learning data creation unit 110. The learning unit 140 includes a reward calculation unit 142.

[0048] The reward calculation unit 142 calculates the reward based on the gap amount when the machining head is closest to the workpiece when an action is taken to move the machining head along the target path at a certain retraction height in the state of the industrial machine 4 indicated by the shape data and the data related to the target path. This reward calculation may be performed using a formula that calculates a larger positive reward, for example, the smaller the gap amount when the machining head is closest to the workpiece. Alternatively, the formula may calculate a smaller reward, for example, the larger the gap amount when the machining head is closest to the workpiece. Furthermore, if the gap amount when the machining head is closest to the workpiece is zero or if the target path is interrupted above the cutting portion, the machining head may be considered to have come into contact with the cutting portion and a large negative reward may be calculated.

[0049] The learning unit 140 according to this embodiment updates the value of taking an action of moving the machining head along the noted path at a certain retraction height in the state of the industrial machine 4 indicated by the shape data and the data related to the noted path, based on the retraction height calculated by the retraction calculation unit 142. The model storage unit 160 stores a trained model related to a value function that calculates the value of selecting each action in each state. The trained model related to the value function may use, for example, an action value table. Alternatively, a model such as a convolutional neural network may be used. By updating this value function, the learning unit 140 updates the value of the action of moving the machining head along the noted path at a certain retraction height in the state of the industrial machine 4 indicated by the shape data and the data related to the noted path. When determining (estimating) the appropriate retraction height for the machining head moving along the noted path using this value function, the retraction height with the highest value among actions (retraction heights) that can be taken in the current state is determined (estimated) as the appropriate retraction height.

[0050] The output unit 180 according to this embodiment outputs a trained model related to the value function generated by the learning unit 140. The output unit 180 may be configured to output data estimated based on the trained model generated by the learning unit 140. The output unit 180 may be configured to output the trained model based on an operation instruction from a user. The output unit 180 may be configured to output the trained model to, for example, an external device 72. Furthermore, the output unit 180 may be configured to output the trained model to another computer such as a fog computer 6, a cloud server 7, or a control device 3 via the network 5.

[0051] The machine learning device 1 having the above configuration generates a trained model related to a value function used to estimate an appropriate retraction height based on shape data indicating the cut shape and data related to a target path moving over the cut portion. Using this trained model, it is expected that retraction operations can be performed at a retraction height that is neither too high nor too low for the cut portion, thereby shortening the processing time. When the cut portion has a shape that will not fall, the appropriate retraction height determined is a position that is the height of the workpiece surface plus a predetermined margin, so no retraction operation is performed, thereby shortening the processing time.

[0052] [Fourth Embodiment] A control device according to a fourth embodiment of the present disclosure will be described below. FIG. 7 is a schematic hardware configuration diagram showing the main parts of a control device according to the fourth embodiment of the present disclosure. The control device 3 according to this embodiment uses a trained model generated by the machine learning device 1 according to the first to third embodiments. The control device 3 can be implemented on a control device that controls industrial machinery such as a laser processing machine, a plasma processing machine, or a gas processing machine. In this embodiment, an example is shown in which the control device 3 is implemented as a control device that controls an industrial machine 4 that is a laser processing machine.

[0053] The CPU 311 included in the control device 3 of the present disclosure is a processor that performs overall control of the control device 3. The CPU 311 reads a system program stored in the ROM 312 via the bus 322 and controls the entire control device 3 in accordance with the system program. The RAM 313 temporarily stores temporary calculation data, display data, various data input from outside, and the like.

[0054] The nonvolatile memory 314 is configured, for example, by a memory backed up by a battery (not shown) or an SSD (Solid State Drive), and retains its stored state even when the power to the control device 3 is turned off. The nonvolatile memory 314 stores control programs and data read from the external device 372 via the interface 315, data and control programs input via the input device 371, various data acquired from the industrial machine 4, and the like. The control programs and data stored in the nonvolatile memory 314 may be expanded into the RAM 313 when executed / used. Furthermore, various system programs such as known analysis programs are written in the ROM 312 in advance.

[0055] The interface 315 is an interface for connecting the CPU 311 of the control device 3 to an external device 372 such as a USB memory, CompactFlash (registered trademark), or SD card. For example, control programs and various data used to control the industrial machine 4 can be read from the external device 372. In addition, control programs and various data edited in the control device 3 can be stored in the external device 372.

[0056] A PLC (Programmable Logic Controller) 316 outputs signals to the industrial machine 4 and peripheral devices of the industrial machine 4 (for example, a tool changer, an actuator such as a robot, a sensor attached to the industrial machine 4, etc.) via an I / O unit 317 to control them according to a sequence program stored in the control device 3. The PLC 316 also receives signals from various switches on an operation panel provided on the main body of the industrial machine 4 and peripheral devices, and after performing the necessary signal processing, passes the signals to the CPU 311. Depending on the configuration of the industrial machine 4, the laser oscillator 360 may also be controlled by the PLC 316.

[0057] The display device 370 displays data and the like obtained as a result of executing various data and programs loaded into memory, output via the interface 318. Furthermore, an input device 371, which is composed of at least one input device such as a keyboard, a pointing device, a voice input device, an imaging device, etc., passes commands, data, etc. based on user operations to the CPU 311 via the interface 319.

[0058] The interface 320 is an interface for connecting the CPU 311 of the control device 3 to a wired or wireless network 5. The network 5 may communicate using technologies such as serial communication such as RS-485, Ethernet (registered trademark), optical communication, wireless LAN, Wi-Fi (registered trademark), Bluetooth (registered trademark), etc. The machine learning device 1, other control devices 3, fog computers 6, cloud servers 7, etc. are connected to the network 5, and data is exchanged between the network 5 and the control devices 3.

[0059] An axis control circuit 330 for controlling the axes of the industrial machine 4 receives a command from the CPU 311 to move the axis by a predetermined amount and outputs the axis command to a servo amplifier 340. The servo amplifier 340 receives this command and drives a servo motor 350 that moves the axis of the machine tool. The servo motor 350 for the axis has a built-in position / speed detector, and position / speed feedback signals from this position / speed detector are fed back to the axis control circuit 330, thereby performing position / speed feedback control. Note that while the hardware configuration diagram in FIG. 7 shows only one axis control circuit 330, servo amplifier 340, and servo motor 350, in reality, there are as many as the number of axes of the industrial machine 4 to be controlled. For example, a laser processing machine has three linear axes, namely, an X-axis, a Y-axis, and a Z-axis, which move the laser oscillator 360 and the workpiece relative to each other.

[0060] In order to control the laser oscillator 360 provided in the industrial machine 4, the oscillator control circuit 335 receives a laser output control command from the CPU 311 and outputs it to the laser oscillator 360. Although only one oscillator control circuit 335 and one laser oscillator 360 are shown in the hardware configuration diagram of Fig. 7, in reality, there are provided as many of them as there are provided in the industrial machine 4 to be controlled.

[0061] 8 is a schematic block diagram illustrating functions of the control device 3 according to the fourth embodiment of the present disclosure. Each function of the control device 3 according to this embodiment is realized by the CPU 311 of the control device 3 shown in FIG. 7 executing a system program and controlling the operation of each part of the control device 3.

[0062] The control device 3 of this embodiment includes a control unit 400, a state observation unit 420, and an estimation unit 440. A machining program 410 used to control the operation of the industrial machine 4 is pre-stored in the RAM 313 to the non-volatile memory 314 of the control device 3. Furthermore, the RAM 313 to the non-volatile memory 314 of the control device 3 are provided with a model storage unit 460, which is an area that stores the trained model generated by the machine learning device 1 according to any of the first to third embodiments.

[0063] The control unit 400 sequentially reads each block of the machining program 410 and analyzes the commands in the read blocks. Then, based on the analysis results, it controls each part of the industrial machine 4. Each block of the machining program 310 includes a movement command for the servo motor 350 that drives each axis of the industrial machine 4, a command to turn on / off laser output from the laser oscillator 360 of the industrial machine 4, and the like. The control unit 400 creates movement command data for the servo motor 350 based on, for example, the movement command. Then, it controls the servo motor 350 that drives each axis of the industrial machine 4 based on the created movement command data. Furthermore, the control unit 400 creates data to control the output signal for the laser oscillator 360 based on a command to turn on / off laser output from the laser oscillator 360. Then, it controls the operation of the laser oscillator 360 based on the created data.

[0064] When the machining head provided in the industrial machine 4 moves over the cut portion of the workpiece cut by cutting, the control unit 400 controls the machining head to retract in the Z-axis direction. At this time, the control unit 400 outputs to the state observation unit 420 shape data indicating the shape of the cut portion of the workpiece and data related to a target path, which is a movement path along which the machining head moves over the cut portion of the workpiece cut by cutting.

[0065] The state observing unit 420 acquires shape data indicating the shape of the cut portion of the workpiece and data relating to a target path, which is a movement path along which the machining head moves over the cut portion of the workpiece cut by the cutting process, from the control unit 400. The state observing unit 420 then outputs the acquired shape data indicating the shape of the cut portion of the workpiece and data relating to a target path, which is a movement path along which the machining head moves over the cut portion of the workpiece cut by the cutting process, to the estimation unit 440.

[0066] The estimation unit 440 performs an estimation process for the appropriate retraction height using a trained model stored in the model storage unit 460, based on shape data indicating the shape of the cut portion of the workpiece input from the state observation unit 420 and data related to a target path, which is a movement path along which the machining head moves over the cut portion of the workpiece cut by the cutting process. The model storage unit 460 stores a trained model generated in advance by the machine learning device 1 according to the first to third embodiments. This trained model is, for example, a trained model based on supervised learning, and outputs the retraction height by inputting shape data indicating the shape of the cut portion of the workpiece and data related to a target path, which is a movement path along which the machining head moves over the cut portion of the workpiece cut by the cutting process. In this case, the estimation unit 440 uses the output of the trained model as the estimation result of the appropriate retraction height. Furthermore, this trained model is, for example, a trained model related to a value function based on reinforcement learning, and calculates the value of the retraction height that can be taken in that state using shape data indicating the shape of the cut portion of the workpiece and data related to the target path, which is the movement path along which the machining head moves over the cut portion of the workpiece cut by the cutting process, as state data. In this case, the estimation unit 440 determines the retraction height that calculates the highest value based on the state data as the estimated result of the appropriate retraction height. The estimation unit 440 outputs the estimated appropriate retraction height to the control unit 400.

[0067] Then, the control unit 400 retracts the machining head of the industrial machine 4 to the appropriate retraction height estimated by the estimation unit 440 .

[0068] The control device 3 having the above configuration estimates the appropriate retraction height based on shape data indicating the cut shape and data related to the target path moving over the cut portion. Then, based on the estimation result, it performs a retraction operation at a retraction height that is neither too high nor too low for the cut portion. As a result, it becomes possible to shorten the machining time. When the cut portion has a shape that will not fall, the estimated appropriate retraction height is a position that is the height of the workpiece surface plus a predetermined margin, so no retraction operation is performed, and the machining time is also shortened.

[0069] [Other Embodiments] In the above-described embodiment, the control device 3 is configured to include the model storage unit 460. However, the model storage unit 460 may be provided on another device, such as the fog computer 6 or the cloud server 7. In this case, the control device 3 references the model storage unit 460 via the network 5. With such a configuration, in an environment where multiple control devices 3 are installed, it becomes possible to share trained models among the multiple control devices 3 and to collectively manage trained models.

[0070] Although the embodiments of the present disclosure have been described in detail above, the present disclosure is not limited to the individual embodiments described above. Various additions, substitutions, modifications, partial deletions, etc. are possible in these embodiments without departing from the gist of the invention or the idea and intent of the present disclosure derived from the content described in the claims and their equivalents. For example, in the above-described embodiments, the order of each operation and the order of each process are shown as examples and are not limited to these. The same applies when numerical values ​​or mathematical expressions are used in the description of the above-described embodiments.

[0071] Below are notes relating to embodiments of the present disclosure. (Supplementary Note 1) A machine learning device (1) according to one aspect of the present disclosure includes: shape data indicating a shape to be cut of a plate-shaped workpiece (500) in an industrial machine (4) equipped with a gap sensor (521) in a machining head (520); data related to a target path, which is a movement path along which a machining head (520) equipped in the industrial machine (4) moves over a cut portion (501) of the workpiece (500) cut by the cutting; a state observation unit (100) that acquires a gap amount when the machining head (520) comes closest to the workpiece (500) while moving along the target path; a learning data creation unit (110) that creates learning data to be used in machine learning learning processing based on the data acquired by the state observation unit (100); and a learning unit (140) that performs machine learning processing based on the learning data and generates a trained model for estimating the retraction height of the machining head (520) based on the shape data and the target path.

[0072] (Supplementary Note 2) The learning data creation unit (110) included in the machine learning device (1) according to another aspect of the present disclosure includes a retraction height calculation unit (112), and the retraction height calculation unit (112) calculates a target retraction height of the machining head (520) when moving along the target path from the gap amount and the retraction height of the machining head (520) when moving along the target path, the learning data creation unit (110) creates teacher data as learning data, with the shape data and data related to the target path as input data and the target retraction height as label data, and the learning unit (140) performs supervised learning based on the learning data, and generates a trained model that estimates an appropriate retraction height using the shape data and the movement path as input.

[0073] (Supplementary Note 3) The learning unit (140) included in a machine learning device (1) according to another aspect of the present disclosure includes a reward calculation unit (142), and the learning data creation unit (110) creates learning data in which data related to the shape data and the noted path is state data, data related to a retraction height when the machining head (520) moves along the noted path is behavior data, and data related to the gap amount is judgment data, the reward calculation unit (142) calculates a reward based on the learning data, and the learning unit (140) performs reinforcement learning based on the learning data and the reward to generate a learned model related to a value function that calculates the value of the retraction height in a state in which the shape data and the movement path are observed. (Supplementary Note 4) The shape data observed by the state observation unit (100) included in a machine learning device (1) according to another aspect of the present disclosure is data indicating a machining path including absolute coordinate values ​​of a cut workpiece.

[0074] (Supplementary Note 5) A control device (3) according to one aspect of the present disclosure includes a trained model generated by a machine learning device (1), a state observation unit (420) that acquires shape data indicating a shape to be cut of a plate-shaped workpiece (500) in an industrial machine (4) that includes a machining head (520) to which a gap sensor (521) is attached, and data related to a target path that is a moving path along which the machining head (520) of the industrial machine (4) moves over a cut portion (501) of the workpiece (500) cut by the cutting process, an estimation unit (440) that performs estimation processing using the trained model based on the shape data and the data related to the target path to estimate an appropriate retraction height, and a control unit (400) that performs retraction control of the machining head (520) of the industrial machine (4) based on the estimated appropriate retraction height.

[0075] 1 Machine learning device 3 Control device 4 Industrial machine 5 Network 6 Fog computer 7 Cloud server 11 CPU 12 ROM 13 RAM 14 Non-volatile memory 15, 17, 18, 20 Interface 22 Bus 70 Display device 71 Input device 72 External device 100 State observation unit 110 Learning data creation unit 112 Evacuation height calculation unit 140 Learning unit 142 Reward calculation unit 160 Model storage unit 180 Output unit 311 CPU 312 ROM 313 RAM 314 Non-volatile memory 315, 318, 319, 320 Interface 316 PLC 317 I / O unit 322 Bus 330 Axis control circuit 335 Oscillator control circuit 340 Servo amplifier 350 Servo motor 360 Laser oscillator 370 Display device 371 Input device 372 External device 400 Control unit 410 Machining program 420 State observation unit 440 Estimation unit 460 Model storage unit 500 Workpiece 501, 502 Cutting portion 505 Table 510 Support base 520 Machining head 521 Gap sensor

Claims

1. A machine learning device comprising: a state observation unit that acquires shape data indicating the shape of a plate-shaped workpiece cut by an industrial machine equipped with a processing head equipped with a gap sensor, data related to a target path which is a movement path along which the processing head equipped on the industrial machine moves over the cut portion of the workpiece cut by the cutting process, the gap amount when the processing head comes closest to the workpiece when moving along the target path, and the retraction height of the processing head when moving along the target path; a learning data creation unit that creates learning data to be used in machine learning learning processing based on the data acquired by the state observation unit; and a learning unit that performs machine learning processing based on the learning data and generates a trained model for estimating the retraction height of the processing head based on the shape data and the target path.

2. The machine learning device according to claim 1, wherein the learning data creation unit includes a retraction height calculation unit, which calculates a target retraction height of the machining head when moving along the target path from the gap amount and the retraction height of the machining head when moving along the target path, the learning data creation unit creates teacher data as learning data, with the shape data and data related to the target path as input data and the target retraction height as label data, and the learning unit performs supervised learning based on the learning data, and generates a trained model that estimates an appropriate retraction height using the shape data and the movement path as input.

3. The machine learning device according to claim 1, wherein the learning unit includes a reward calculation unit, the learning data creation unit creates learning data in which data relating to the shape data and the focus path is state data, data relating to the retraction height when the machining head moves along the focus path is behavior data, and data relating to the gap amount is judgment data, the reward calculation unit calculates a reward based on the learning data, and the learning unit performs reinforcement learning based on the learning data and the reward to generate a learned model related to a value function that calculates the value of the retraction height in a state in which the shape data and the movement path are observed.

4. The machine learning device according to claim 2 or 3, wherein the shape data observed by the state observation unit is data indicating a machining path including absolute coordinate values ​​of the cut workpiece.

5. A control device comprising: a trained model generated by the machine learning device according to any one of claims 1 to 4; a state observation unit that acquires shape data indicating the shape of a plate-shaped workpiece cut by an industrial machine having a machining head equipped with a gap sensor, and data relating to a target path, which is a moving path along which a machining head of the industrial machine moves over the cut portion of the workpiece cut by the cutting process; an estimation unit that performs estimation processing using the trained model based on the shape data and the data relating to the target path, and estimates an appropriate retraction height; and a control unit that performs retraction control of the machining head of the industrial machine based on the estimated appropriate retraction height.

Images