MACHINE LEARNING DEVICE, FORECASTING DEVICE AND CONTROL DEVICE

Abstract

Machine learning device (30), comprising: a state monitoring unit (301) which takes as input data manual feed state information comprising a manual feed impulse waveform at a time point of a manual feed operation during each manual feed operation performed in a machine tool (10) suitable for manual feed; a marking detection unit (302) that determines marking data indicating a distance by which a moving part of the machine tool (10) has moved within a predetermined time immediately after the manual feed impulse waveform of the manual feed state information contained in the input data; and a learning unit (303) that performs supervised learning using the input data determined by the state observation unit (301) and the mark data determined by the mark detection unit (302) and generates a learned model (250).

Description

BACKGROUND OF THE INVENTION Area of ​​the invention

[0001] The present invention relates to a machine learning device, a prediction device and a control device. Related technology

[0002] In a machine tool, to prevent collisions of a moving part of a tool or the like, a technique is known which performs a fault check based on the position of the moving part at a previous time, calculated from a machine program, and an outline model in which the outline of the moving part and the outline of a stationary part are stored in advance. If a fault is detected, the moving part slows down and stops, or an alarm is generated. See, for example, patent document 1.

[0003] JP 4 221 016 B2 (Patent Document 1) OVERVIEW OF THE INVENTION

[0004] However, some machine tools have a moving part that is suitable for manual feed. In this case, it is difficult to predict the future position of the moving part, since the machining program is not executed and the user moves the moving part in real time.

[0005] In this case, a method for predicting the future position of the moving part could involve calculating its future position and performing a fault check based on the assumption that the current pulse signal state (i.e., the operation of the manual feed) is maintained using the pulse signal generated by a handle operated by the user during manual feed. However, if the user does not continuously rotate the handle, or even if the user carefully rotates the handle near a disturbing object such as a workpiece or table, as described above, the calculated future position could indicate that the handle is being disturbed by the object, and an alarm could be generated.Because an alarm is generated excessively, the user's intended activity is consequently interrupted, which can decrease work efficiency.

[0006] Therefore, it is desirable to prevent a collision of the moving part without generating an excessive alarm during manual feed.

[0007] The problem is solved by a machine learning device with the features of claim 1, by a prediction device with the features of claim 4, and by a control device with the features of claim 8. (1) An aspect of a machine learning device (30) according to this disclosure comprises: a state monitoring unit (301) which, as input data, determines manual feed state information comprising a manual feed pulse waveform at a time point of a manual feed operation during each manual feed operation performed in a machine tool (10) suitable for manual feed; a mark detection unit (302) which determines mark data indicating a distance by which a moving part of the machine tool (10) has moved within a predetermined time immediately following the manual feed pulse waveform of the manual feed state information contained in the input data;and a learning unit (303) that performs supervised learning using the input data determined by the state monitoring unit (301) and the mark detection data acquired by the mark detection unit (302) and generates a learned model (250). (2) One aspect of a prediction device (20) according to this disclosure comprises a learned model (250) generated by the machine learning device (30) described in (1); an input unit (201) that inputs the manual feed state information of the manual feed currently being performed with respect to a machine tool (10) suitable for manual feed; and a prediction unit (202) that inputs the manual feed state information received from the input unit (201) into the learned model (250) and predicts a movement distance of a moving part of the machine tool (10) after a predetermined time from a current time based on the manual feed state information. (3) One aspect of a control device (15) according to this disclosure includes the prediction device (20) described in (2).

[0008] According to one aspect, it is possible to prevent a collision of the moving part without generating an excessive alarm during manual feeding. BRIEF DESCRIPTION OF THE DRAWINGS Fig. Figure 1 is a functional block diagram showing a functional configuration example of a working system according to an embodiment; Fig. 2A is a diagram showing an example that explains the prediction processing of a prediction device; Fig. 2B is a diagram showing an example that explains the prediction processing of a prediction device; Fig. Figure 3 is a diagram showing an example of a manual feed impulse waveform during manual feed operation from the beginning to the end of a handle operation; Fig. Figure 4 is a diagram showing an example of a learned model used by the prediction device of Fig. 1 is provided; Fig. 5A is a diagram showing an example of prediction processing by a prediction unit; Fig. 5B is a diagram showing an example of the prediction processing of the prediction unit; Fig. 6 is a flowchart that explains the prediction processing of the prediction device during an operational phase; Fig. 7 is a diagram showing an example of the configuration of a work system; Fig. Figure 8 is a diagram showing an example of the configuration of the working system; Fig. Figure 9A is a diagram showing an example of a case of operating the moving part of the machine tool by delaying by one prediction cycle from the generation of the pulse signal; Fig. 9B is a diagram showing an example of a case of the operation of the moving part of the machine tool by delaying by one prediction cycle from the generation of the pulse signal; Fig. Figure 10 is a diagram showing an example of a learned model that outputs an estimated impulse waveform; Fig. Figure 11A is a diagram showing an example of a case of operating the moving part of the machine tool by delaying by one prediction cycle from the generation of the pulse signal; and Fig. Figure 11B is a diagram showing an example of a case of operating the moving part of the machine tool by delaying by one prediction cycle from the generation of the pulse signal. DETAILED DESCRIPTION OF THE INVENTION

[0009] An embodiment of the present disclosure is described below with reference to the drawings. <Ausführungsform>

[0010] Fig. Figure 1 is a functional block diagram showing a functional configuration example of a working system according to one embodiment. As in Fig. As shown in Figure 1, the work system 1 comprises a machine tool 10, a prediction device 20 and a machine learning device 30.

[0011] The machine tool 10, the predictive device 20, and the machine learning device 30 can be directly connected to each other via a connection interface (not shown). Alternatively, the machine tool 10, the predictive device 20, and the machine learning device 30 can be connected to each other via a network (not shown), such as a local area network (LAN) and the internet. In this case, the machine tool 10, the predictive device 20, and the machine learning device 30 are equipped with a communication unit (not shown) to communicate with each other over such a connection. It should be noted that, as described later, the predictive device 20 can include the machine learning device 30. Furthermore, the machine tool 10 can include both the predictive device 20 and the machine learning device 30.

[0012] The machine tool 10 is a machine tool known to those skilled in the art and includes a control device 15. The machine tool 10 operates on the basis of an operating command from the control device 15 and actuates the moving parts, such as a tool, by means of a handle (not shown) which is contained in the machine tool 10 to be actuated. As will be described later, when the handle (not shown) is actuated and manual feed is performed, the machine tool 10 can also output the waveform of the pulse signal generated by the handle (not shown) to the prediction device 20 as a manual feed pulse waveform of manual feed state information, while the manual feed is executed via a communication unit (not shown) of the machine tool 10.Furthermore, the manual feed status information can include the distance to a disturbance object such as a workpiece or table, user identification information of the user performing the manual feed, the date and time of the manual feed, and the number of axes. Additionally, the manual feed status information can include environmental conditions such as the temperature and humidity at the location where the machine tool 10 is installed.

[0013] It is noted that information regarding the distance to the disturbance object (i.e., the position of the disturbance object) can be stored in advance in a memory unit (not shown), such as a ROM (resettable memory), which may be included in the machine tool 10. Furthermore, the distance to the disturbance object influences, for example, the user's operation of the handle (not shown). In a case where the distance between the moving part and the disturbance object is large, the user, for instance, turns the handle (not shown) sharply (faster) to move the moving part a significant amount. Therefore, the handle (not shown) generates many pulse signals. On the other hand, in a case where the distance between the moving part and the disturbance object is small, the user turns the handle (not shown) slightly (more slowly) to move the moving part slightly.Thus, the handle (not shown) generates a small number of pulse signals. Since the distance to the interfering object is closely related to the hand feed pulse waveform, the distance to the interfering object is included in the hand feed state information.

[0014] Furthermore, the operation of the handle (not shown) varies from user to user and often differs significantly between forceful and gentle movement of the moving part. Therefore, user identification information is included in the hand feed state information to predict how the pulse waveform will change by learning the pulse waveform for each user in the machine learning device 30 described later.

[0015] Furthermore, the number of axes indicates, for example, the X-axis, the Y-axis and the Z-axis, which are the directions of movement of the moving part.

[0016] The control device 15 is a numerical control device known to those skilled in the art and generates an operating command based on the control information and sends the generated operating command to the machine tool 10. Thus, the control device 15 controls the operation of the machine tool 10. It should be noted that the control device 15 can output the manual feed status information to the prediction device 20 via the communication unit of the machine tool 10 (not shown) instead of to the machine tool 10.

[0017] The control device 15 can also be a device independent of the machine tool 10.

[0018] When manual feed is performed in the machine tool 10 during the operating phase, the predictive device 20 determines the current manual feed status information of the machine tool 10. The predictive device 20 inputs the determined manual feed status information into the learned model provided by the machine learning device 30, which will be described later, whereby the movement distance of the moving part of the machine tool 10 can be predicted after a predetermined time from the current time.

[0019] Especially if, as in Fig. As shown in Figure 2A, if user A, for example, operates the moving part in the X-axis (e.g., axis number "1" or the like) of the machine tool 10 by means of the handle (not shown) from time hh: mm of weekday c, the predictive device 20 indicates the distance D in which the moving part moves from the present time until b [ms], as shown in Figure 2A. Fig. 2B shown, based on the manual feed impulse waveform from the present time to the time before a [ms] and the distance s [mm] from the present position to the disturbance object ahead.

[0020] In order for the prediction device 20 to be able to predict the distance D, the machine learning device 30, which will be described later, uses at least the manual feed impulse waveform of time a [ms] as input data and determines as marker data the distance by which the moving part of the machine tool 10 has been moved in the time of b [ms] since the time immediately after the output of the waveform, and adopts the data as training data.

[0021] Before describing the prediction device 20, machine learning for generating a learned model is described. <Maschinenlernvorrichtung 30>

[0022] For example, the machine learning device 30 determines in advance as input data a manual feed pulse waveform at the time of a manual feed operation in each manual feed operation performed by the machine tool 10, a distance to a disturbance object at the time of the manual feed operation, user identification information of a user who performed the manual feed operation, a date and time at which the manual feed operation was performed, and manual feed state information, which includes a number of axes actuated in the manual feed operation.

[0023] Furthermore, the machine learning device 30 determines data as a marker (correct answer) that indicates the distance by which the moving part of the machine tool 10 has moved within a predetermined time immediately after the manual feed pulse waveform in the recorded input data.

[0024] The machine learning device 30 performs supervised learning using training data, which is a set of labels and acquired input data, and creates a learned model, which will be described later.

[0025] In this way, the machine learning device 30 is able to provide the learned model it has created to the prediction device 20. The machine learning device 30 is described in detail.

[0026] As in Fig. As shown in Figure 1, the machine learning device 30 comprises a condition monitoring unit 301, a marker detection unit 302, a learning unit 303 and a storage unit 304.

[0027] During the learning phase, the condition monitoring unit 301 determines the following input data from the machine tool 10 via the communication unit (not shown): the manual feed state information, including the manual feed pulse waveform at the time of the manual feed operation in any manual feed operation performed by the machine tool 10; the distance to the disturbance object at the time of the manual feed operation; the user identification information of the user who performed the manual feed operation; the date and time at which the manual feed operation was performed; and the number of axes actuated in the manual feed operation.

[0028] Fig. Figure 3 is a diagram illustrating an example of a manual feed impulse waveform in a manual feed operation from the beginning to the end of a handle actuation (not shown). It should be noted that the upper and lower sections of Fig. The 3 shown manual feed impulse waveforms are the same.

[0029] As in the section above of Fig. As shown in Figure 3, the state monitoring unit 301, for example, divides the hand feed impulse waveform from the beginning to the end of the actuation of the hand handle (not shown) at each specified time point (corresponding to a [ms] in Fig. 2A), such as 500 ms, to generate segmented manual feed pulse waveforms 401 to 405. The condition monitoring unit 301 acquires as input data the generated manual feed pulse waveforms 401 to 405 together with the distance to the disturbance object at the time of manual feed operation in Fig. 3. Manual feed status information, the user identification information of the user who is operating the manual feed system in Fig. 3, the date and time at which the manual feed operation was carried out in Fig. 3 was executed, and the number of axes used in the manual feed operation in Fig. 3 was activated. The condition monitoring unit 301 stores the acquired input data in the storage unit 304.

[0030] Please note that the specified time is not limited to 500 ms and can be adjusted at any time.

[0031] The marking determination unit 302 determines, for example, as marking data (correct answer), data that indicate the distance by which the moving part of the machine tool 10 has moved within a predetermined time immediately after each of the manual feed impulse waveforms 401 to 405 of the manual feed state information in the input data.

[0032] In particular, as recorded in the lower section of Fig. Figure 3 shows the marking detection unit 302, for example, as marking data (correct answer) the distance by which the moving part of the machine tool 10 moved within a certain time, such as 200 ms (corresponding to b [ms] in Fig. 2B), from times t1 to t5 immediately following the respective generated manual feed pulse waveforms 401 to 405 (e.g., the value of the time integration of the pulses in ranges 411 to 415), as shown by hatching. The marking detection unit 302 stores the recorded marking data in the storage unit 304.

[0033] It should be noted that the specified time is not limited to 200 ms and may be set to a time shorter than at least the specified time of the subdivided manual feed pulse waveforms 401 to 405.

[0034] The learning unit 303 receives the aforementioned set of input data and the marker as training data. The learning unit 303 performs supervised learning using the received training data, thereby creating a learned model 250 for predicting the movement distance of the moving part of the machine tool 10.

[0035] Learning unit 303 provides the created learned model 250 of the prediction device 20.

[0036] It is noted that it is preferable to prepare a number of training data sets for supervised learning. For example, training data from machine tools can be collected at various locations, such as a customer's factory or similar.

[0037] Fig. Figure 4 is a diagram showing an example of the learned model 250, which is used for the prediction device 20 of Fig. 1 is provided. Here, the learned model 250 illustrates how in Fig. Figure 4 shows a multi-layer neural network in which the manual feed state information of the manual feed currently being performed in the machine tool 10 is used as input layers, wherein the estimated value of the movement distance of the moving part of the machine tool 10 after a predetermined time (for example, 200 ms etc.) from the present time is used as the output layer by the manual feed state information.

[0038] Here, the manual feed status information of a currently executed manual feed includes the manual feed impulse waveform during manual feed operation, the distance to the interfering object at the time of the manual feed operation, the user identification information of the user performing the manual feed operation, the date and time of the execution of the manual feed operation, and the number of axes actuated during the manual feed operation.

[0039] It should be noted that the manual feed status information of a currently running manual feed may include environmental conditions such as temperature and humidity in which the machine tool 10 is installed.

[0040] Furthermore, in a case where new training data is obtained after the learned model 250 has been created, the learning unit 303 can continue to perform supervised learning on the learned model 250 to update the learned model 250 that has been created.

[0041] In this way, training data can be automatically obtained from the operation of a handle (not shown) by a normal user, making it possible to routinely increase the accuracy of the prediction.

[0042] Supervised learning can be implemented through online learning. Furthermore, supervised learning can be implemented through batch learning. Additionally, supervised learning can be implemented through mini-batch learning.

[0043] Online learning refers to a learning method where manual feed is performed in machine tool 10, with supervised learning being executed immediately each time training data is generated. Batch learning refers to a learning method where, while manual feed is performed in machine tool 10 and training data is repeatedly generated, multiple training data sets corresponding to each repetition are collected, and supervised learning is performed using all of the collected training data. Furthermore, mini-batch learning refers to a learning method that is an intermediate method between online learning and batch learning, in which supervised learning is performed when a certain amount of training data has been collected.

[0044] The storage unit 304 is a RAM (random access memory) or the like and stores input data determined by the state monitoring unit 301, the marking data determined by the marking detection unit 302, the learned model 250 created by the learning unit 303, and the like.

[0045] The machine learning process for generating the learned model 250, which is contained in the prediction device 20, was explained above. Next, the prediction device 20 will be explained in its operational phase. <Vorhersagegerät 20 in Betriebsphase>

[0046] As in Fig. As shown in Figure 1, the prediction device 20 in the operating phase comprises an input unit 201, a prediction unit 202, a determination unit 203, a notification unit 204 and a storage unit 205.

[0047] It is noted that the prediction device 20 includes an arithmetic processing unit (not shown), such as a central processing unit (CPU), to enable the operation of the in Fig. to implement the functional blocks shown in Figure 1. Furthermore, the prediction device 20 includes an auxiliary storage device (not shown), such as a ROM or a hard disk drive (HDD), on which various control programs are stored, and a main storage device (not shown), such as a RAM, for storing data that is temporarily required for the execution of programs by the arithmetic processing unit.

[0048] Furthermore, in the predictive device 20, the arithmetic processing unit reads the operating system and application software from the auxiliary storage device and performs arithmetic processing based on this operating system and application software, while the read operating system and application software are deployed to the main storage device. The predictive device 20 controls each hardware component based on the arithmetic processing result. In this way, the processing of the in Fig. The functional blocks shown in section 1 are implemented. That is, the prediction device 20 can be implemented through the interaction of hardware and software.

[0049] Input unit 201 receives the manual feed status information from machine tool 10, indicating the current manual feed operation performed in machine tool 10. Input unit 201 then outputs this manual feed status information to prediction unit 202.

[0050] The prediction unit 202 feeds the manual feed status information entered by the input unit 201 into the learned model 250. Fig. 3 and predicts the distance of movement of the moving part of the machine tool 10 after a predetermined time from the present time.

[0051] Fig. 5A and Fig. Figure 5B contains diagrams, each showing an example of the prediction processing of the prediction unit 202. It should be noted that the diagrams in Fig. 1 and Fig. 2 shown impulse waveforms. 5A and 5B are examples of the manual feed impulse waveforms that are captured by the machine tool 10 in a case where the user of the machine tool 10 operates the X-axis (e.g., axis number "1" or the like) by means of the handle (not shown) from Monday, 10:00 a.m.

[0052] As in Fig. As shown in 5A, the prediction unit 202 feeds into the learned model 250 of Fig. 3. The manual feed pulse waveform from the present time to 500 ms prior, together with the distance to the disturbance object at the time of the currently executed manual feed operation, the user identification information of the user performing the manual feed operation, the date and time at which the manual feed operation is performed, and the axis number “1” being actuated in the manual feed operation, are entered below the determined manual feed pulse waveform. As in Fig. As shown in Figure 5B, the prediction unit 202 predicts the distance of movement of the moving part of the machine tool 10 at the time 200 ms from the present time.

[0053] It is noted that the duration of the manual feed pulse waveform input into the learned model 250 can correspond to the time interval of the manual feed pulse waveform of the input data used to generate the learned model 250, i.e., 500 ms. The estimated value of the movement distance output by the learned model 250 can correspond to the time required by the moving part of the machine tool 10 to move the distance of the marking data used to generate the learned model 250, i.e., 200 ms.

[0054] Furthermore, the prediction unit 202 can predict the estimated value of the movement distance within the prediction cycle of time intervals such as 10 ms and 50 ms. This enables the machine tool 10 to avoid collisions during manual feed.

[0055] The determination unit 203 determines whether the moving part of the machine tool 10 collides with the disturbance object or not, based on the estimated value of the movement distance predicted by the prediction unit 202 for each prediction cycle and the distance to the disturbance object.

[0056] In particular, in a case where the estimated value of the movement distance is shorter than the distance to the disturbance object, the determining unit 203 determines that no collision occurs, thereby determining that the operation of the manual feed continues without generating an alarm.

[0057] On the other hand, in a case where the estimated value of the movement distance is equal to or greater than the distance to the interfering object, the determination unit 203 determines that the collision will occur, thereby generating an alarm and stopping the operation of the manual feed.

[0058] In this way, the prediction device 20 performs the fault check based on the estimated value of the movement distance predicted using the learned model 250, thereby enabling a prediction to be made close to the user's activity and reducing the frequency of the alarm.

[0059] In a case where the detection unit 203 determines that a collision will occur, the notification unit 204 can issue an alarm and a shutdown to an output device (not shown), such as a liquid crystal display, contained in the machine tool 10 and / or the control device 15. It should be noted that the notification unit 204 can also be notified by voice via a loudspeaker (not shown).

[0060] The storage unit 205, for example, is a ROM, a hard drive or the like, and can store the learned model 250 along with various control programs. <Vorhersageverarbeitung der Vorhersagevorrichtung 20 in der Betriebsphase>

[0061] Next, an operation will be explained which relates to the prediction processing of the prediction device 20 according to the present embodiment.

[0062] Fig. Figure 6 is a flowchart illustrating the prediction processing of the prediction device 20 during the operational phase. The sequence shown here is executed repeatedly for each prediction cycle.

[0063] In step S11, the input unit 201 from the machine tool 10 enters the manual feed status information of the manual feed currently being performed in the machine tool 10.

[0064] In step S12, the prediction unit 202 inputs the manual feed status information of the currently executed manual feed, which was entered in step S11, into the learned model 250 in order to predict an estimated value of the movement distance of the moving part of the machine tool 10.

[0065] In step S13, the determination unit 203 determines whether the moving part of the machine tool 10 will collide with the interfering object or not, based on a comparison between the estimated value of the movement distance predicted in step S12 and the distance to the interfering object. If a collision is determined to occur, processing continues with step S14, and if no collision is determined to occur, processing ends.

[0066] In step S14, the notification unit 204 notifies the alarm and the operational stop determined in step S13.

[0067] Thus, according to the embodiment, the prediction device 20 can input the manual feed status information of the manual feed currently being performed in the machine tool 10 into the learned model 250 and predict an estimated value for the movement distance of the moving part of the machine tool 10. Furthermore, the prediction device 20 can detect in advance whether the moving part will collide with the obstruction or not, based on a comparison between the estimated value of the predicted movement distance and the distance to the obstruction.

[0068] This means that since the prediction device 20 performs the fault check based on the estimated value of the movement distance predicted using the learned model 250, the prediction can be made close to the user's activity, and thus it is possible to prevent the collision of the moving part without generating an excessive alarm during manual feed.

[0069] Although one embodiment has been described above, the prediction device 20 and the machine learning device 30 are not limited to the embodiment described above and include modifications, improvements and the like of a scope that can achieve an objective of the present invention. <Change example 1>

[0070] In the embodiment described above, the machine learning device 30 is shown by way of example as a device that differs from the machine tool 10, the control device 15, and the prediction device 20. However, it can be configured such that the machine tool 10, the control device 15, or the prediction device 20 includes some or all of the functions of the machine learning device 30. <Change example 2>

[0071] Furthermore, in the embodiment described above, for example, the prediction device 20 is represented as a device that differs from the machine tool 10 and the control device 15. However, it can be configured such that the machine tool 10 or the control device 15 can contain some or all of the functions of the prediction device 20.

[0072] Alternatively, a server can, for example, comprise one or all of the input unit 201, prediction unit 202, determination unit 203, notification unit 204, and storage unit 205 of the prediction device 20. The prediction device 20 can be implemented in the cloud using a virtual server function or the like.

[0073] Furthermore, the prediction device 20 can be a distributed processing system in which the functions of the prediction device 20 are appropriately distributed across multiple servers. <Change example 3>

[0074] Furthermore, in the embodiment described above, for example, the prediction device 20 predicts an estimated value of the movement distance of the moving part of the machine tool 10 from the manual feed state information of the currently executed manual feed prediction, which was determined by a machine tool 10, using the learned model 250 provided by the machine learning device 30. However, the present invention is not limited to this. For example, as shown in Fig. Figure 7 shows that server 50 stores the learned model 250 generated by machine learning device 30 and shares the learned model 250 with number m of prediction devices 20A (1) to 20A (m) connected to network 60 (m being an integer equal to or greater than 2). In this way, it is possible to adopt the learned model 250 even if a new machine tool and a new prediction device are installed.

[0075] It is noted that the prediction devices 20A (1) to 20A (m) are each connected to the machine tools 10A (1) to 10A (m).

[0076] Furthermore, each of the machine tools 10A (1) to 10A (m) corresponds to the machine tool 10 of Fig. 1 where they are machine tools of the same model to each other. Each of the prediction devices 20A (1) to 20A (m) corresponds to the prediction device 20 of Fig. 1.

[0077] Alternatively, as in Fig. As shown in Figure 8, server 50, for example, can operate as a predictive device 20 and predict an estimated value for the movement distance of the moving part from the manual feed state information of the manual feed currently being performed for each of the machines 10A (1) to 10A (m) connected to the network 60. This allows the learned model 250 to be adopted even when a new machine tool is installed.

[0078] It is pointed out that in a case where the machine tools 10A (1) to 10A (m) are at least two different models, the machine learning device 30 can generate the learned model 250 for each model and the server 50 can store the generated learned model 250 for each model. <Change example 4>

[0079] For example, in the embodiment described above, the machine tool 10 manually feeds the moving part according to the user's (not shown) actuation of the handle, while the prediction device 20 forecasts an estimated value for the movement distance of the moving part of the machine tool 10 for each prediction cycle, such as 50 ms. However, the present invention is not limited to this. For example, in the machine tool 10, even if the user (not shown) actuates the handle, the moving part can be moved by delaying the generation of the pulse signal by the prediction cycle. That is, it is possible for the prediction device 20 to prevent a collision by allowing the machine tool 10 to operate the moving part after the accuracy of the previous prediction has been confirmed.

[0080] Fig. 9A and Fig. Figure 9B is a diagram illustrating an example of a case where the operation of the moving part of the machine tool 10 is delayed by the prediction cycle from the generation of the pulse signal. It should be noted that the diagram in Fig. 9A and Fig. The impulse waveforms shown in Figure 9B are examples of the manual feed impulse waveforms that were captured by the machine tool 10 in a case where the user of the machine tool 10 operated the X-axis (e.g., axis number "1" or the like) from 10:00 a.m. on Monday using the handle (not shown).

[0081] The upper section of Fig. Figure 9A shows, in a solid line, the manual feed impulse waveform that was detected by the prediction device 20 of the machine tool 10 at the present time. On the other hand, the lower section of Fig. 9A the manual feed pulse waveform in which the moving part of the machine tool 10 has been operated up to the present time, i.e. the manual feed pulse waveform, which is modified by the prediction cycle (50 ms) compared to the pulse waveform of Fig. 9A is delayed.

[0082] In particular, the determining unit 203 of the predicting device 20 can calculate the actual movement distance D2 of the moving part in the time interval from 200 ms before the present time until the present time, during which the estimated value D1 of the movement distance of the moving part at the present time is predicted, based on the time integration of the manual feed impulse waveform, which is defined by the solid line in the upper section of Fig. 9A is displayed. The determination unit 203 compares the calculated actual movement distance D2 with the estimated value D1 of the movement distance. In a case where the estimated value D1 of the movement distance is equal to or greater than the actual movement distance D2, the determination unit 203 can determine that the estimated value D1 of the movement distance is correct and can determine whether the moving part of the machine tool 10 collides with the disturbance object, based on the estimated value D1 of the movement distance and the distance to the disturbance object.

[0083] On the other hand, in a case where the estimated value D1 of the movement distance is less than the actual movement distance D2, the determination unit 203 determines that the estimated value D1 of the movement distance is incorrect. In this case, as in Fig. Figure 9B shows the determination unit 203, which determines whether the moving part of the machine tool 10 collides with the disturbing object, based on the estimated value D3 of the movement distance shown by hatching and the distance to the disturbing object, which are predicted by a conventional prediction method that assumes that the existing impulse waveform will continue to remain constant or the like. <Change example 5>

[0084] For example, in the embodiment described above, the learned model 250 is generated in advance with the manual feed impulse waveform at the time of the manual feed operation in each manual feed operation performed by the machine tool 10, the distance to the interfering object at the time of the manual feed operation, the user identification information of the user who performed the manual feed operation, the date and time at which the manual feed operation was performed, and the number of axes actuated by the manual feed operation as input data. However, the present invention is not limited to this.For example, the machine learning device 30 can generate the learned model 250 for all user identification information for each day of the week on which the manual feed operation was performed, for each time zone in which the manual feed operation was performed, or for each number of axes, instead of inputting the training data for the user identification information, the date and time at which the manual feed was performed, and the number of axes. For example, the learned model 250, generated for each user identification information, enables the predictive device 20 to accurately predict the estimated value of the movement distance of the moving part of the machine tool 10, taking into account the operating habit of the handle (not shown) of the machine tool 10 for each user. <Change example 6>

[0085] Furthermore, the learned model 250 in the embodiment described above, as shown in Fig. As shown in Figure 4, for example, by inputting the manual feed status information of the currently executed manual feed, the estimated value of the movement distance of the moving part of the machine tool 10 after a predetermined time from the present time (e.g., 200 ms, etc.) can be generated. However, the present invention is not limited to this. For example, the learned model 250 can, using the manual feed status information of the currently executed manual feed that is being entered, output an estimated pulse waveform that will be generated by a handle (not shown) of the machine tool 10 after a predetermined time (e.g., 200 ms, etc.) from the present time.

[0086] Fig. Figure 10 is a diagram showing an example of a learned model 250A that outputs an estimated impulse waveform.

[0087] The learned model 250A from Fig. Figure 10 illustrates a multi-layered neural network in which the manual feed state information of the currently executed manual feed is used as input layers and the estimated pulse waveform output by the handle (not shown) of the machine tool 10 after a predetermined time (e.g. 200 ms) from the present time is used as the output layer.

[0088] In a case of predicting the estimated impulse waveform using the learned model 250A from Fig. 10 and operation of the moving part after a delay of one prediction cycle from the generation of the pulse signal, even if the machine tool 10 is operated by the user with a handle (not shown), the prediction device 20 can cause the machine tool 10 to operate the moving part after checking whether the previous prediction was correct or incorrect. In this way, the prediction device 20 is able to prevent the collision of the moving part of the machine tool 10.

[0089] Fig. 11A and Fig. Figure 11B contains diagrams, each showing an example of a case of operating the moving part of the machine tool 10 by delaying it by one prediction cycle from the generation of the pulse signal. It should be noted that the diagrams in Fig. 11A and Fig. The pulse waveforms shown in 11B are examples of manual feed pulse waveforms generated by the user of machine tool 10, who operates the X-axis (e.g., axis number "1" or the like) from Monday at 10:00 a.m. by a manual movement (not shown).

[0090] The upper section of Fig. Figure 11A shows the manual feed impulse waveform, detected by the prediction device 20 of the machine tool 10 at the present time, in a solid line. On the other hand, the lower section of Fig. 11A the manual feed pulse waveform in which the moving part of the machine tool 10 has been operated up to the present time, i.e. the manual feed pulse waveform generated by the prediction cycle (50 ms) compared to the pulse waveform of Fig. 11A is delayed.

[0091] In particular, the determining unit 203 of the predicting device 20 compares, for example, the estimated impulse waveform, indicated by a dashed line, predicted at time 200 ms before the present time, with the actual manual feed impulse waveform, indicated by a solid line, as shown in the upper section of Fig. 11A shown. In a case where the actual manual feed pulse waveform does not exceed the estimated pulse waveform, the determination unit 203 can determine whether the moving part of the machine tool 10 collides with the disturbing object or not, based on the estimated value of the movement distance of the moving part, which is predicted from the estimated pulse waveform and the distance to the disturbing object.

[0092] On the other hand, in a case where the actual manual feed impulse waveform may exceed the estimated impulse waveform, the determining unit 203 estimates the impulse waveform based on a conventional prediction method in which the existing impulse waveform remains constant or the like, as in Fig. Figure 11B shows that the determination unit 203 can determine whether the moving part of the machine tool 10 collides with the disturbing object, based on the estimated value of the movement distance, which is calculated from the estimated impulse waveform and the distance to the disturbing object.

[0093] It should be noted that in one embodiment, each function included in the prediction device 20 and the machine learning device 30 may be implemented by hardware, software, or a combination thereof. Here, implementation by software means that it is carried out by a computer that reads and executes a program.

[0094] Each component contained in the prediction device 20 and the machine learning device 30 may be implemented by hardware, including electronic circuits or the like, software, or a combination thereof. If implemented by software, the programs comprising that software are installed on the computer. Alternatively, these programs may be recorded on removable media and distributed to the user or downloaded to the user's computer via a network. If configured by hardware, some or all of the functions of each component contained in the devices described above may be implemented by an integrated circuit (IC), such as an ASIC (Application Specific Integrated Circuit), a gate array, an FPGA (Field Programmable Gate Array), a CPLD (Complex Programmable Logic Device), or the like.

[0095] Programs can be stored on and provided to a computer using various types of non-volatile, computer-readable media. Non-volatile media include various types of physical storage media. Examples of non-volatile, computer-readable media include magnetic recording media (for example, a flexible floppy disk, magnetic tape, and a hard disk drive), magneto-optical recording media (for example, a magneto-optical disk), CD-ROM (solid-state storage), CD-R, CD-R / W, semiconductor memory (e.g., mask ROM, PROM (programmable ROM), EPROM (erasable PROM), flash ROM, and RAM (random access memory). Programs can also be provided to a computer using any of the various types of volatile, computer-readable media. Examples of volatile, computer-readable media include electrical signals, optical signals, and electromagnetic waves.A volatile, computer-readable medium can provide programs to a computer via a wired communication path such as an electrical cable, an optical fiber, or the like, or via a wireless communication path.

[0096] It is pointed out that the step of writing programs to be recorded on a recording medium involves processing carried out in a sequential manner according to the order and processing, which is carried out in parallel or independently, even if the processing is not necessarily carried out in a sequential manner.

[0097] In other words, the machine learning device, the prediction device and the control device of the present disclosure can assume various embodiments with the following configurations. (1) A machine learning device 30 according to the present disclosure comprises: a state monitoring unit 301, which as input data determines manual feed state information comprising a manual feed pulse waveform at a time point of a manual feed operation during each manual feed operation performed in a machine tool 10 suitable for manual feed; a mark detection unit 302, which determines mark data indicating a distance by which a moving part of the machine tool 10 has moved within a predetermined time immediately after the manual feed pulse waveform of the manual feed state information contained in the input data; and a learning unit 303, which performs supervised learning using the input data determined by the state monitoring unit 301 and the mark data determined by the mark detection unit 302 and generates a learned model 250.

[0098] According to the machine learning device 30, it is possible to generate a learned model 250 that predicts an estimated value of the movement distance of the moving part by the manual feed, which is performed in relation to the machine tool 10.

[0099] (2) In the machine learning device 30 according to (1), the manual feed state information at a time during manual feed operation can include any distance to a disturbance object and / or user identification information of a user who performed the manual feed operation and / or a date and time at which the manual feed operation was performed and / or a number of axes that were actuated during the manual feed operation. In this way, the machine learning device 30 is able to generate the learned model 250, which can more accurately predict the estimated value of the movement distance of the moving part by the manual feed.

[0100] (3) In the machine learning device 30 according to (1) or (2), the condition monitoring unit 301 can determine the input data for each model of the machine tool 10, the marking determination unit 302 can record the marking data for each model of the machine tool 10, and the learning unit 303 can generate a learned model 250 for each model of the machine tool 10 using the input data and the marking data for each model of the machine tool 10.

[0101] Thus, it is possible for the machine learning device 30 to generate the learned model 250, which predicts an estimated value of the movement distance of the moving part according to the model of the machine tool 10.

[0102] (4) A prediction device 20 according to the present disclosure comprises: a learned model 250 generated by the machine learning device 30 according to one of (1) to (3); an input unit 201 which inputs the manual feed state information of the manual feed currently being performed with respect to a machine tool 10 suitable for manual feed; and a prediction unit 202 which inputs the manual feed state information received from the input unit 201 into the learned model 250 and predicts a movement distance of a moving part of the machine tool 10 after a predetermined time from a present time based on the manual feed state information.

[0103] According to this prediction device 20, it is possible to prevent the collision of the moving part of the machine tool 10 without generating an excessive alarm during manual feed.

[0104] (5) In the prediction device (20) according to (4) the prediction unit 202 can cyclically predict the movement distance in a shorter time interval than the predetermined time.

[0105] In this way, the prediction device 20 is able to prevent the collision of the moving part of the machine tool 10 with high accuracy.

[0106] (6) In the prediction device 20 according to (4) or (5), the learned model 250 may be contained in a server 50 which is connected to the prediction device 20 via a network 60.

[0107] In this way, it is possible for the prediction device 20 to adopt the learned model 250, even if a new machine tool 10, a new control device 15 and a new prediction device 20 are installed.

[0108] (7) The prediction device 20 according to paragraphs (4) to (6) may also include the machine learning device 30 according to paragraphs (1) to (3).

[0109] In this way, it is possible for the prediction device 20 to achieve the same effects as those described above (1) to (6).

[0110] (8) A control device 15 according to the present disclosure comprises a prediction device 20 according to any of (4) to (7). According to this control device 15, it is possible to achieve the same effects as with any of the above described (4) to (7). EXPLANATION OF THE REFERENCE SYMBOLS 10 machine tool 15 Control device 20 Predictive Device 201 Input unit 202 Forecast Unit 203 Unit of determination 250 learned model 30 machine learning device 301 Condition monitoring unit 302 Marking Determination Unit 303 Learning Unit 50 servers

Claims

[1] Machine learning device (30), comprising: a state monitoring unit (301) which takes as input data manual feed state information comprising a manual feed impulse waveform at a time point of a manual feed operation during each manual feed operation performed in a machine tool (10) suitable for manual feed; a marking detection unit (302) that determines marking data indicating a distance by which a moving part of the machine tool (10) has moved within a predetermined time immediately after the manual feed impulse waveform of the manual feed state information contained in the input data; and a learning unit (303) that performs supervised learning using the input data determined by the state observation unit (301) and the mark data determined by the mark detection unit (302) and generates a learned model (250). [2] Machine learning device (30) according to claim 1, wherein the manual feed state information includes a distance to a disturbance object and / or user identification information of a user who performed the manual feed operation, and / or a date and time at which the manual feed operation was performed, and / or a number of axes that were actuated in the manual feed operation. [3] Machine learning device (30) according to claim 1 or 2, wherein the condition monitoring unit (301) determines the input data for each model of the machine tool (10), the marking determination unit (302) determines the marking data for each model of the machine tool (10); and the learning unit (303) generates a learned model (250) for each model of the machine tool (10) using the input data and the labeling data for each model of the machine tool (10). [4] Predictive device (20) comprising: a learned model (250) generated by the machine learning device (30) according to one of claims 1 to 3; an input unit (201) that inputs the manual feed status information of the manual feed currently being performed with respect to a machine tool (10) that is suitable for manual feed; and a prediction unit (202) which inputs the manual feed state information from the input unit (201) into the learned model (250) and predicts a movement distance of a moving part of the machine tool (10) after a predetermined time from a present time based on the manual feed state information. [5] Predictive device (20) according to claim 4, wherein the predictive unit (202) can cyclically predict the movement distance in a shorter time interval than the predetermined time. [6] Predictive device (20) according to claim 4 or 5, wherein the learned model (250) is contained in a server (50) which is accessible from the predictive device (20) via a network (60). [7] Predictive device (20) according to any one of claims 4 to 6, further comprising the machine learning device (30) according to any one of claims 1 to 3. [8] Control device (15) comprising a prediction device (20) according to any one of claims 4 to 7.

Images