Machine learning device, prediction device, and control device
The learning completion model generated by machine learning predicts the future movement distance of the moving parts of the machine tool, solving the problem of excessive alarms during manual feed and achieving accurate interference detection and efficient operation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- FANUC LTD
- Filing Date
- 2020-11-03
- Publication Date
- 2026-05-15
AI Technical Summary
During manual feed of machine tools, existing technology struggles to accurately predict the future position of the moving parts, leading to frequent and excessive alarms and impacting operational efficiency.
A machine learning device is used to generate a learning-complete model. Supervised learning is performed using manual feed status information to predict the future movement distance of the movable part. Interference checks are also performed in conjunction with the location of interfering objects to avoid excessive alarm generation.
It effectively prevents collisions of moving parts, reduces alarm frequency, and improves the efficiency of manual feeding processes.
Smart Images
Figure CN112784955B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to machine learning devices, prediction devices, and control devices. Background Technology
[0002] A known technique involves a machine tool where, to prevent collisions with moving parts such as tools, an interference check is performed based on the position of the moving part calculated from the executed machining program at a prior time, and a profile model pre-stored with the profiles of the moving part and the fixed part. If interference is detected, the moving part is decelerated and stopped, or an alarm is generated. For example, see Patent Document 1.
[0003] However, some machine tools have devices that allow manual feeding of the moving parts. In this case, the user moves the moving parts in real time without executing a machining program, making it difficult to predict the future position of the moving parts.
[0004] In this case, as a method for predicting the future position of the movable part, consider the following approach: using the pulse signal generated by the manual handle operated by the user during manual feed, calculate the future position of the movable part based on the premise that the current pulse signal state (i.e., the manual feed operation) can be maintained, and perform interference checks. However, even if the user does not continuously rotate the manual handle, or if the user carefully rotates the manual handle near a distracting object such as a workpiece or worktable, the above method sometimes generates an alarm because the calculated future position is considered to be interfering with a distracting object. As a result, excessive alarms cause the user's desired operation to be interrupted, leading to reduced work efficiency.
[0005] Patent Document 1: Japanese Patent No. 4221016 Summary of the Invention
[0006] Therefore, it is desirable to prevent collisions with the movable parts without generating excessive alarms during manual feeding.
[0007] (1) One aspect of the machine learning apparatus of this disclosure includes: a state observation unit that acquires manual feed state information as input data, the manual feed state information including a manual feed pulse waveform during any manual feed operation performed by a machine tool capable of manual feed; a tag acquisition unit that acquires tag data, the tag data representing the distance moved by a movable part of the machine tool within a predetermined time period after the manual feed pulse waveform of the manual feed state information included in the input data; and a learning unit that performs supervised learning using the input data acquired by the state observation unit and the tag acquisition unit to generate a learned model.
[0008] (2) One aspect of the prediction apparatus of this disclosure includes: a learning completion model generated by the machine learning apparatus of (1); an input unit that inputs the manual feed state information of the currently performed manual feed to a machine tool capable of manual feed; and a prediction unit that inputs the manual feed state information input by the input unit into the learning completion model, and predicts the movement distance of the movable part of the machine tool after a predetermined time from the present based on the manual feed state information.
[0009] (3) One aspect of the control device of this disclosure has the prediction device of (2).
[0010] One method can prevent collisions of the moving parts without generating excessive alarms during manual feeding. Attached Figure Description
[0011] Figure 1 This is a functional block diagram illustrating an example of the functional structure of a working system related to one implementation method.
[0012] Figure 2A This is a diagram illustrating an example of the prediction processing of a prediction device.
[0013] Figure 2B This is a diagram illustrating an example of the prediction processing of a prediction device.
[0014] Figure 3 This is a diagram illustrating an example of the manual feed pulse waveform during a manual feed operation, from the start to the end of the manual handle operation.
[0015] Figure 4 It means to provide Figure 1 The diagram shows an example of a predictive device learning a model.
[0016] Figure 5A This is a diagram illustrating an example of the forecasting process in the forecasting unit.
[0017] Figure 5B This is a diagram illustrating an example of the forecasting process in the forecasting unit.
[0018] Figure 6 This is a flowchart illustrating the prediction processing of the prediction device during the application phase.
[0019] Figure 7 This is a diagram illustrating an example of the structure of a working system.
[0020] Figure 8 This is a diagram illustrating an example of the structure of a working system.
[0021] Figure 9AThis diagram illustrates an example of a machine tool whose movable part performs actions periodically with a delay from the generation of a pulse signal.
[0022] Figure 9B This diagram illustrates an example of a machine tool whose movable part performs actions periodically with a delay from the generation of a pulse signal.
[0023] Figure 10 This is a diagram illustrating an example of a learned model that outputs a predicted pulse waveform.
[0024] Figure 11A This diagram illustrates an example of a machine tool whose movable part performs actions periodically with a delay from the generation of a pulse signal.
[0025] Figure 11B This diagram illustrates an example of a machine tool whose movable part performs actions periodically with a delay from the generation of a pulse signal.
[0026] Symbol Explanation
[0027] 10. Machine tools;
[0028] 15. Control device;
[0029] 20. Predictive devices;
[0030] 201 Input Section;
[0031] 202 Forecasting Department;
[0032] 203 Decision-Making Department;
[0033] 250. The learning process has been completed.
[0034] 30 Machine learning devices;
[0035] 301 Condition Observation Unit;
[0036] 302 Label Acquisition Department;
[0037] 303 Study Department;
[0038] 50 servers. Detailed Implementation
[0039] Hereinafter, an embodiment of the present disclosure will be described using the accompanying drawings.
[0040] <One Implementation Method>
[0041] Figure 1 This is a functional block diagram illustrating an example of the functional structure of a working system related to one implementation method. For example... Figure 1 As shown, the working system 1 includes: a machine tool 10, a prediction device 20, and a machine learning device 30.
[0042] Machine tool 10, prediction device 20, and machine learning device 30 can be directly connected to each other via a connection interface not shown. Alternatively, machine tool 10, prediction device 20, and machine learning device 30 can also be connected to each other via a network not shown, such as a LAN (Local Area Network) or the Internet. In this case, machine tool 10, prediction device 20, and machine learning device 30 have a communication unit not shown for communicating with each other via such a connection. Furthermore, as described later, prediction device 20 may include machine learning device 30. Additionally, machine tool 10 may include both prediction device 20 and machine learning device 30.
[0043] Machine tool 10 is a machine tool known to those skilled in the art, and includes a control device 15. Machine tool 10 operates according to the operation commands of the control device 15, and movable parts such as tools are moved by operating a manual handle (not shown) included in machine tool 10. Furthermore, as described later, when manually feeding by operating the manual handle (not shown), machine tool 10 can output the waveform of the pulse signal generated by the manual handle (not shown) during manual feeding as manual feed status information to prediction device 20 via a communication unit (not shown) of machine tool 10. In addition, the manual feed status information may include the distance to the workpiece or worktable (or other interfering objects), user identification information of the user performing the manual feeding, the date and time of the manual feeding, and the axis number. Furthermore, the manual feed status information may also include environmental conditions such as temperature or humidity of the location where machine tool 10 is located.
[0044] Furthermore, information related to the distance to the interfering object (i.e., the position of the interfering object) can be pre-stored in a storage unit (not shown) such as a ROM (Read Only Memory) included in the machine tool 10. The distance to the interfering object also affects, for example, the operation of the user's manual handle (not shown). For instance, when the distance between the movable part and the interfering object is large, the user rotates the manual handle (not shown) rapidly to move the movable part significantly, thus generating more pulse signals. Conversely, when the distance between the movable part and the interfering object is small, the user rotates the manual handle (not shown) slowly to move the movable part slightly, thus generating fewer pulse signals. Therefore, the distance to the interfering object is closely related to the manual feed pulse waveform, and thus, the distance to the interfering object is included in the manual feed status information.
[0045] Furthermore, the operation of the manual handle (not shown) varies from user to user, and there are often significant differences between large and small movements of the movable part. Therefore, in the machine learning device 30 described later, the pulse waveform is learned according to the user, thereby including user identification information in the manual feed state information in order to predict how the pulse waveform will change.
[0046] In addition, the axis number indicates the direction of movement of the movable part, i.e., the X-axis, Y-axis, Z-axis, etc.
[0047] The control device 15 is a numerical control device known to those skilled in the art. It generates motion commands based on control information and sends the generated motion commands to the machine tool 10. Thus, the control device 15 controls the motion of the machine tool 10. In addition, the control device 15 can output manual feed status information to the prediction device 20 in place of the machine tool 10 via a communication unit of the machine tool 10 (not shown).
[0048] Furthermore, the control device 15 can be a device independent of the machine tool 10.
[0049] During operation, when manual feeding is performed on the machine tool 10, the prediction device 20 obtains the current manual feed status information from the machine tool 10. The prediction device 20 inputs the obtained manual feed status information to the learning completion model provided by the machine learning device 30 (described later), thereby predicting the movement distance of the movable part of the machine tool 10 after a predetermined time from the present.
[0050] Specifically, for example Figure 2A As shown, when user A operates the movable part of machine tool 10 via a manual handle (not shown) from time hh:mm on weekday c, the prediction device 20 predicts the movement based on the manual feed pulse waveform from the current position to a[ms] ago and the distance s[mm] from the current position to the interference object. Figure 2B As shown, the predicted distance D of the movable part from the current position to b[ms] is calculated.
[0051] Therefore, in order for the prediction device 20 to predict the distance D, the machine learning device 30 described later takes at least a [ms] of manual feed pulse waveform as input data, and obtains the distance moved by the movable part of the machine tool 10 from the time the waveform is output until b [ms] as label data, which is used as training data.
[0052] Therefore, before describing the prediction device 20, the machine learning used to generate the learned model will be explained.
[0053] <Machine Learning Device 30>
[0054] The machine learning device 30 may, for example, acquire manual feed status information in advance as input data. The manual feed status information includes: the manual feed pulse waveform during any manual feed operation performed by the machine tool 10, the distance to the interference object during the manual feed operation, the user identification information of the user performing the manual feed operation, the date and time of the manual feed operation, and the axis number of the operation performed by the manual feed operation.
[0055] In addition, the machine learning device 30 obtains data representing distance as a label (correct answer), which is the distance the movable part of the machine tool 10 moves within a predetermined time after the manual feed pulse waveform in the obtained input data.
[0056] The machine learning device 30 performs supervised learning by acquiring input data and training data of the labeled group, and constructs the learning completion model described later.
[0057] In this way, the machine learning device 30 can provide the learned model to the prediction device 20.
[0058] The machine learning device 30 will be described in detail.
[0059] like Figure 1 As shown, the machine learning device 30 includes: a state observation unit 301, a label acquisition unit 302, a learning unit 303, and a storage unit 304.
[0060] During the learning phase, the status observation unit 301 obtains manual feed status information from the machine tool 10 via a communication unit (not shown) as input data. The manual feed status information includes: the manual feed pulse waveform during any manual feed operation performed by the machine tool 10, the distance to the interference object during the manual feed operation, the user identification information of the user performing the manual feed operation, the date and time of the manual feed operation, and the axis number that was operated by the manual feed operation.
[0061] Figure 3 This diagram illustrates an example of the manual feed pulse waveform during a manual feed operation, from the start to the end of the manual handle (not shown) operation. Additionally, Figure 3 The manual feed pulse waveforms shown in the upper and lower sections are the same.
[0062] like Figure 3 As shown in the previous paragraph, the state observation unit 301 specifies a time interval, for example, 500 ms (equivalent to...). Figure 2AThe manual feed pulse waveform (not shown) from the start to the end of the manual handle (manual feed pulse waveform) operation is segmented, generating segmented manual feed pulse waveforms 401-405. The status observation unit 301 obtains the generated manual feed pulse waveforms 401-405 and manual feed status information as input data. Figure 3 The distance to the interference point during manual feed operation, and the execution Figure 3 User identification information of users performing manual feed operations, and performing Figure 3 The date and time of the manual feed operation, and via Figure 3 The axis number is used for manual feed operations. The status observation unit 301 stores the acquired input data in the storage unit 304.
[0063] In addition, the specified time is not limited to 500ms; it can be set to any time.
[0064] The tag acquisition unit 302 acquires data representing distance, for example, as tag data (correct answer data). The distance is the distance that the movable part of the machine tool 10 moves within a predetermined time after each of the manual feed pulse waveforms 401-405 of the manual feed status information in the input data.
[0065] Specifically, for example Figure 3 As shown in the next paragraph, the tag acquisition unit 302 acquires the tag data (forward data) as a specified time (equivalent to) 200ms after each of the generated manual feed pulse waveforms 401-405. Figure 2B The distance the movable part of the machine tool 10 moves within the b[ms]) (for example, the time integral value of the pulse in the shaded area 411-415). The tag acquisition unit 302 stores the acquired tag data in the storage unit 304.
[0066] In addition, the specified time is not limited to 200ms, and can be set to any time shorter than the specified time of the manual feed pulse waveform 401-405 obtained by segmentation.
[0067] The learning unit 303 receives the aforementioned input data and labeled group as training data. The learning unit 303 uses the received training data to perform supervised learning, thereby constructing a learning completion model 250 that predicts the movement distance of the movable part of the machine tool 10.
[0068] Then, the learning unit 303 provides the constructed learning completion model 250 to the prediction device 20.
[0069] In addition, it is desirable to prepare a large amount of training data for supervised learning. For example, training data can be obtained from machine tools 10 in various actual operating locations such as customers' factories.
[0070] Figure 4 It means to provide Figure 1 A diagram illustrating an example of how the prediction device 20 has completed learning the model 250. Here, as shown... Figure 4 As shown, the learning completion model 250 is a multi-layer neural network with the following input layer: the manual feed state information of the current manual feed in the machine tool 10 is used as the input layer, and the estimated value of the movement distance of the movable part of the machine tool 10 after a predetermined time (e.g., 200ms) from the current time is estimated through the manual feed state information.
[0071] Here, the manual feed status information of the current manual feed operation includes: the manual feed pulse waveform during the manual feed operation, the distance to the interference object during the manual feed operation, the user identification information of the user performing the manual feed operation, the date and time of the manual feed operation, and the axis number of the operation performed through the manual feed operation.
[0072] In addition, the current manual feed status information may include environmental conditions such as temperature or humidity of the location where the machine tool 10 is set.
[0073] Furthermore, after constructing the learning completed model 250, the learning unit 303 further supervises learning the learning completed model 250 when new training data is obtained, thereby updating the learning completed model 250 constructed once.
[0074] In this way, training data can be automatically obtained from manual handle operations (not shown) performed by ordinary users, thus improving the accuracy of predictions on a daily basis.
[0075] The supervised learning described above can be conducted through online learning. Furthermore, supervised learning can also be conducted through batch learning. Additionally, supervised learning can be conducted through mini-batch learning.
[0076] Online learning refers to the following method: supervised learning is performed immediately whenever manual feed is performed on machine tool 10 and training data is generated. Batch learning, on the other hand, involves collecting multiple training data sets corresponding to each repetition during repeated manual feeds on machine tool 10, and using all collected training data for supervised learning. Mini-batch learning is an intermediate method between online and batch learning, where supervised learning is performed only when a certain amount of training data has accumulated.
[0077] The storage unit 304 is RAM (Random Access Memory) or the like, storing input data obtained by the state observation unit 301, tag data obtained by the tag acquisition unit 302, and the learning completion model 250 constructed by the learning unit 303.
[0078] The above describes the machine learning of the learning-completed model 250 in the generative prediction device 20.
[0079] Next, the prediction device 20 in the application phase will be described.
[0080] <Prediction device 20 during the application phase>
[0081] like Figure 1 As shown, the prediction device 20 in the application stage is configured to include: an input unit 201, a prediction unit 202, a decision unit 203, a notification unit 204, and a storage unit 205.
[0082] In addition, the prediction device 20 is designed to achieve Figure 1 The predictor device 20 operates the function blocks and has an arithmetic processing unit (not shown) such as a CPU (Central Processing Unit). Furthermore, the predictor device 20 has an auxiliary storage device (not shown) such as a ROM or HDD (Hard Disk Drive) for storing various control programs, or a main storage device (not shown) such as RAM for storing data temporarily needed when the arithmetic processing unit executes a program.
[0083] Furthermore, in the prediction device 20, the arithmetic processing unit reads an operating system or application software from the auxiliary storage device, expands the read OS or application software in the main storage device, and performs arithmetic processing based on the OS or application software. Based on the calculation results, the prediction device 20 controls each piece of hardware. Thus, [the following is achieved]... Figure 1 The functional blocks involve processing. In other words, the prediction device 20 can be implemented through a combination of hardware and software.
[0084] The input unit 201 inputs manual feed status information of the current manual feed being performed in the machine tool 10 from the machine tool 10. The input unit 201 outputs the input manual feed status information to the prediction unit 202.
[0085] The prediction unit 202 inputs the manual feed status information from the input unit 201 to... Figure 3 The learning model 250 is completed to predict the moving distance of the movable part of the machine tool 10 after a predetermined time from the present.
[0086] Figure 5A and Figure 5BThis diagram illustrates an example of the prediction processing performed by the prediction unit 202. Additionally, Figure 5A and Figure 5B The pulse waveform shown is an example of a manual feed pulse waveform obtained from machine tool 10 when the user of machine tool 10 operates the X-axis (e.g., axis number "1") via a manual handle (not shown) starting at 10:00 AM on Monday.
[0087] like Figure 5A As shown, the prediction unit 202 inputs the following data together from the acquired manual feed pulse waveform (from the current time up to 500ms ago), the distance to the interfering object during the current manual feed operation, the user identification information of the user performing the manual feed operation, the date and time of the manual feed operation, and the axis number "1" of the manual feed operation into the system. Figure 3 The learning process completed model 250. (For example...) Figure 5B As shown, the prediction unit 202 predicts the moving distance of the movable part of the machine tool 10 at a time 200ms from the current moment.
[0088] Furthermore, the length of the manual feed pulse waveform input to the learning completed model 250 can correspond to the time interval of the manual feed pulse waveform used to generate the input data of the learning completed model 250, i.e., 500ms. The estimated value of the moving distance output by the learning completed model 250 can correspond to the time taken for the movable part of the machine tool 10 to move the distance used to generate the tag data of the learning completed model 250, i.e., 200ms.
[0089] Furthermore, the prediction unit 202 can predict the estimated value of the travel distance with a prediction cycle of 10ms or 50ms. As a result, the machine tool 10 can avoid collisions during manual feed.
[0090] The decision unit 203 determines whether the movable part of the machine tool 10 collides with the interfering object based on the estimated value of the moving distance predicted by the prediction unit 202 according to the prediction cycle and the distance to the interfering object.
[0091] More specifically, when the estimated distance traveled is shorter than the distance to the interfering object, the decision unit 203 determines that there is no collision, and therefore decides to continue the manual feed operation without generating an alarm.
[0092] On the other hand, when the estimated distance of movement exceeds the distance to the interfering object, the decision unit 203 determines that a collision has occurred, thereby generating an alarm and deciding to stop the operation of manual feed.
[0093] In this way, the prediction device 20 performs interference checks based on the estimated value of the movement distance predicted by the learning completion model 250, thus enabling predictions that approximate the user's operation and suppressing the frequency of alarms.
[0094] When the decision unit 203 determines that a collision has occurred, the notification unit 204 can output an alarm and a shutdown signal to an output device (not shown) such as a liquid crystal display included in the machine tool 10 and / or the control device 15. Additionally, the notification unit 204 can also provide audio notification via a speaker (not shown).
[0095] The storage unit 205 can be a ROM or HDD, etc., storing various control programs and the learning completed model 250.
[0096] <Predictive Processing of Predictive Device 20 in the Application Phase>
[0097] Next, the operation related to the prediction processing of the prediction device 20 in this embodiment will be explained.
[0098] Figure 6 This is a flowchart illustrating the prediction processing of the prediction device 20 during the application phase. The process shown here is repeated according to the prediction cycle.
[0099] In step S11, the input unit 201 inputs manual feed status information of the manual feed currently being performed in the machine tool 10 from the machine tool 10.
[0100] In step S12, the prediction unit 202 inputs the manual feed status information of the current manual feed, which was input in step S11, into the learning completion model 250 to predict the estimated value of the moving distance of the movable part of the machine tool 10.
[0101] In step S13, the determination unit 203 determines whether the movable part of the machine tool 10 collides with the interference object by comparing the estimated value of the moving distance predicted in step S12 with the distance to the interference object. If a collision is determined, the process proceeds to step S14; if no collision is determined, the process ends.
[0102] In step S14, the notification unit 204 notifies the alarm and operation stop determined in step S13.
[0103] Through the above, the prediction device 20 of one embodiment inputs the manual feed state information of the current manual feed in the machine tool 10 into the learning completion model 250, and predicts the estimated value of the movement distance of the movable part of the machine tool 10. Then, the prediction device 20 can detect in advance whether the movable part will collide with the interference object by comparing the predicted estimated value of the movement distance with the distance to the interference object.
[0104] That is, the prediction device 20 performs interference checks based on the estimated value of the movement distance predicted by the learning completion model 250. Therefore, it can make predictions that approximate the user's operation, and will not generate excessive alarms during manual feeding, thus preventing collisions of the movable parts.
[0105] The above describes one embodiment, but the prediction device 20 and the machine learning device 30 are not limited to the above embodiment, and include variations and improvements within the scope of achieving the purpose.
[0106] <Variation Example 1>
[0107] In the above embodiments, the machine learning device 30 is shown to be a device different from the machine tool 10, the control device 15 and the prediction device 20, but the machine tool 10, the control device 15 or the prediction device 20 may have some or all of the functions of the machine learning device 30.
[0108] <Variation Example 2>
[0109] Furthermore, as in the above embodiments, the prediction device 20 is shown to be a device different from the machine tool 10 or the control device 15, but the machine tool 10 or the control device 15 may have some or all of the functions of the prediction device 20.
[0110] Alternatively, for example, the server may have some or all of the input unit 201, prediction unit 202, decision unit 203, notification unit 204, and storage unit 205 of the prediction device 20. Furthermore, the various functions of the prediction device 20 can be implemented in the cloud using virtual server functionality or the like.
[0111] Furthermore, the prediction device 20 can be used as a distributed processing system that appropriately distributes the functions of the prediction device 20 to multiple servers.
[0112] <Variation Example 3>
[0113] Furthermore, as in the above embodiment, the prediction device 20 uses the learning completion model 250 provided by the machine learning device 30 to predict an estimated value of the travel distance of the movable part of the machine tool 10 based on the manual feed state information of the currently performed manual feed obtained from the machine tool 10, but is not limited thereto. For example, such as Figure 7 As shown, server 50 can store the learned model 250 generated by machine learning device 30, and share the learned model 250 with m prediction devices 20A(1)-20A(m) connected to network 60 (m is an integer greater than 2). Thus, the learned model 250 can be applied even if new machine tools and prediction devices are configured.
[0114] In addition, each of the prediction devices 20A(1)-20A(m) is connected to each of the machine tools 10A(1)-10A(m).
[0115] Furthermore, each of machine tools 10A(1)–10A(m) corresponds to Figure 1 Machine tools 10 are all of the same model. Each of the prediction devices 20A(1)-20A(m) corresponds to Figure 1 The prediction device 20.
[0116] Or, such as Figure 8 As shown, server 50, for example, acts as prediction device 20, predicting the estimated travel distance of movable parts for each of the machine tools 10A(1)-10A(m) connected to network 60 based on the manual feed status information of the currently performed manual feed. Thus, the learning completion model 250 can be applied even when a new machine tool is configured.
[0117] In addition, when there are multiple different models of machine tool 10A(1)-10A(m) (more than 2), the machine learning device 30 generates a learning completed model 250 according to the model, and the server 50 can store the learning completed model 250 of each model generated.
[0118] <Variation Example 4>
[0119] Furthermore, as in the embodiment described above, the machine tool 10 manually feeds the movable part in response to the user's operation of a manual handle (not shown). Simultaneously, the prediction device 20 predicts the estimated travel distance of the movable part of the machine tool 10 at prediction periods of 50 ms, but is not limited to this. For example, even when the machine tool 10 performs an operation based on the user's manual handle (not shown), the movable part can be made to move with a delay of the prediction period from the generation of the pulse signal. That is, the prediction device 20 moves the movable part of the machine tool 10 after confirming the correctness of past predictions, thereby preventing collisions.
[0120] Figure 9A and Figure 9B This diagram illustrates an example of how the movable part of the machine tool 10 performs its operation periodically with a delay from the generation of a pulse signal. Additionally, Figure 9A and Figure 9B The pulse waveform shown is an example of a manual feed pulse waveform obtained from machine tool 10 when the user of machine tool 10 operates the X-axis (e.g., axis number "1") via a manual handle (not shown) starting at 10:00 AM on Monday.
[0121] Figure 9A The upper section, represented by a solid line, shows the manual feed pulse waveform obtained by the prediction device 20 from the machine tool 10 before the current moment. On the other hand, Figure 9AThe lower segment represents the manual feed pulse waveform of the movable part of the machine tool 10 before the current moment, i.e., it is related to... Figure 9A The pulse waveform is compared to the manual feed pulse waveform with a delay prediction period (50ms).
[0122] Specifically, the decision unit 203 of the prediction device 20 can, according to Figure 9A The time integral of the manual feed pulse waveform shown by the solid line in the upper section is used to calculate the actual movement distance D2 of the movable part from 200ms before the estimated value D1 of the movement distance at the current moment. The determination unit 203 compares the calculated actual movement distance D2 with the estimated value D1 of the movement distance. If the estimated value D1 of the movement distance is greater than or equal to the actual movement distance D2, the determination unit 203 determines that the estimated value D1 of the movement distance is correct, and can determine whether the movable part of the machine tool 10 has collided with the interference object based on the estimated value D1 of the movement distance and the distance to the interference object.
[0123] On the other hand, if the estimated travel distance D1 is smaller than the actual travel distance D2, the determination unit 203 determines that the estimated travel distance D1 is incorrect. In this case, if... Figure 9B As shown, the determination unit 203 can determine whether the movable part of the machine tool 10 collides with the interfering object based on the estimated value D3 of the movement distance shown in the shade, which is predicted by existing prediction methods such as setting the current pulse waveform to constant and continuous, and the distance to the interfering object.
[0124] <Variation Example 5>
[0125] Furthermore, as in the embodiments described above, the learned model 250 generates manual feed state information in advance as input data. However, it is not limited to this. The manual feed state information includes the manual feed pulse waveform during any manual feed operation performed by the machine tool 10, the distance to the interference object during the manual feed operation, the user identification information of the user performing the manual feed operation, the date and time of the manual feed operation, and the axis number of the operation performed by the manual feed operation. For example, the machine learning device 30 does not input the user identification information, the date and time of the manual feed operation, and the axis number as training data. Instead, it can generate the learned model 250 based on the user identification information, the week of the week in which the manual feed operation was performed, the time period in which the manual feed operation was performed, or the axis number. For example, if the learned model 250 is generated based on the user identification information, the prediction device 20 can take into account the operating habits of each user's manual handle (not shown) on the machine tool 10 and predict the estimated value of the movement distance of the movable part of the machine tool 10 with high accuracy.
[0126] <Variation Example 6>
[0127] Furthermore, for example, in the embodiments described above, such as Figure 4 As shown, the learning completion model 250, by inputting the manual feed status information of the currently performed manual feed, outputs an estimated value of the movement distance of the movable part of the machine tool 10 after a predetermined time (e.g., 200 ms) from the current time, but is not limited to this. For example, by inputting the manual feed status information of the currently performed manual feed, the learning completion model 250 can output an estimated pulse waveform generated by the manual handle (not shown) of the machine tool 10 after a predetermined time (e.g., 200 ms) from the current time.
[0128] Figure 10 This is a diagram illustrating an example of the learning completion model 250A, representing the output estimated pulse waveform.
[0129] exemplify Figure 10 The learning model 250A is a multi-layer neural network as follows: the manual feed state information of the current manual feed is used as the input layer, and the estimated pulse waveform output by the manual handle (not shown) of the machine tool 10 after a predetermined time (e.g., 200ms) from the current time is used as the output layer.
[0130] In addition, the prediction device 20 is in use Figure 10 When the learning model 250A predicts the estimated pulse waveform, even if the machine tool 10 performs operation based on the user's manual handle (not shown), the movable part moves with a delay from the generation of the pulse signal during the prediction period. After confirming the correctness of the past prediction, the movable part of the machine tool 10 can be moved. In this way, the prediction device 20 can prevent collisions of the movable part of the machine tool 10.
[0131] Figure 11A and Figure 11B This diagram illustrates an example of how the movable part of the machine tool 10 performs its operation periodically with a delay from the generation of a pulse signal. Additionally, Figure 11A and Figure 11B The pulse waveform shown is an example of a manual feed pulse waveform generated by a user of machine tool 10 operating the X-axis (e.g., axis number "1") via a manual handle (not shown) starting at 10:00 AM on Monday.
[0132] Figure 11A The upper section, represented by a solid line, shows the manual feed pulse waveform obtained by the prediction device 20 from the machine tool 10 before the current moment. On the other hand, Figure 11A The lower segment represents the manual feed pulse waveform of the movable part of the machine tool 10 before the current moment, i.e., it is related to... Figure 11A The pulse waveform is compared to the manual feed pulse waveform with a delay prediction period (50ms).
[0133] Specifically, the decision unit 203 of the prediction device 20, for example, Figure 11A As shown in the upper section, the estimated pulse waveform, predicted 200ms before the current time and represented by the dashed line, is compared with the actual manual feed pulse waveform, represented by the solid line. When the actual manual feed pulse waveform does not exceed the estimated pulse waveform, the determination unit 203 can determine whether the movable part of the machine tool 10 has collided with the interfering object based on the estimated value of the movable part's movement distance predicted from the estimated pulse waveform and the distance to the interfering object.
[0134] On the other hand, when the actual manual feed pulse waveform exceeds the estimated pulse waveform, the decision unit 203, such as Figure 11B As shown, the pulse waveform is estimated using existing prediction methods, such as setting the current pulse waveform to be constant and continuous. The determination unit 203 can determine whether the movable part of the machine tool 10 collides with the interfering object based on the estimated value of the movement distance calculated from the estimated pulse waveform and the distance to the interfering object.
[0135] Furthermore, the functions included in the prediction device 20 and the machine learning device 30 in one embodiment can be implemented separately by hardware, software, or a combination thereof. Here, "implemented by software" means that the computer implements them by reading and executing a program.
[0136] The various structural components included in the prediction device 20 and the machine learning device 30 can be implemented using hardware, software, or a combination thereof, including electronic circuits. When implemented in software, the program constituting the software is installed in a computer. Furthermore, these programs can be recorded on removable media and distributed to users, or downloaded to users' computers via a network for distribution. In addition, when implemented in hardware, for example, integrated circuits (ICs) such as ASICs (Application Specific Integrated Circuits), gate arrays, FPGAs (Field Programmable Gate Arrays), and CPLDs (Complex Programmable Logic Devices) can be used to constitute part or all of the functionality of the various structural components included in the aforementioned devices.
[0137] Various types of non-transitory computer-readable media can be used to store programs and provide them to a computer. Non-transitory computer-readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include: magnetic storage media (e.g., floppy disks, magnetic tapes, hard disks), magneto-optical storage media (e.g., optical disks), CD-ROMs (Read-Only Memory), CD-Rs, CD-R / Ws, and semiconductor memories (e.g., mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs), flash memory ROMs, and RAM). Conversely, programs can also be supplied to a computer via various types of transient computer-readable media. Examples of transient computer-readable media include electrical signals, optical signals, and electromagnetic waves. Transient computer-readable media can supply programs to a computer via wired communication paths such as wires and optical fibers, or wireless communication paths.
[0138] Furthermore, the steps describing the program recorded in the recording medium include not only processing performed in a time sequence, but also processing that is not necessarily performed in a time sequence, and processing performed in parallel or individually.
[0139] In other words, the machine learning apparatus, prediction apparatus, and control apparatus of this disclosure can be implemented in various ways having the following structures.
[0140] (1) The machine learning apparatus 30 of this disclosure includes: a state observation unit 301, which acquires manual feed state information as input data, the manual feed state information including a manual feed pulse waveform during a manual feed operation in any manual feed operation performed by a machine tool 10 capable of manual feed; a tag acquisition unit 302, which acquires tag data, the tag data representing the distance moved by the movable part of the machine tool within a predetermined time after the manual feed pulse waveform of the manual feed state information contained in the input data; and a learning unit 303, which uses the input data acquired by the state observation unit 301 and the tag data acquired by the tag acquisition unit 302 to perform supervised learning and generate a learning completed model 250.
[0141] According to the machine learning device 30, a learning completion model 250 can be generated, which is used to predict the estimated value of the movement distance of the movable part caused by manual feed to the machine tool 10.
[0142] (2) In the machine learning device 30 described in (1), the manual feed status information may include: the distance to the interference during the manual feed operation, the user identification information of the user performing the manual feed operation, the date and time of the manual feed operation, and one of the axis numbers of the operation performed by the manual feed operation.
[0143] In this way, the machine learning device 30 can generate a learning-completed model 250, which can more accurately predict the estimated value of the movement distance of the movable part caused by manual feeding.
[0144] (3) In the machine learning device 30 described in (1) or (2), the state observation unit 301 may obtain input data according to the model of the machine tool 10, the tag acquisition unit 302 may obtain tag data according to the model of the machine tool 10, and the learning unit 303 may use the input data and tag data of each model of the machine tool 10 to generate a learning completed model 250 according to the model of the machine tool 10.
[0145] In this way, the machine learning device 30 can generate a learning completed model 250, which predicts the estimated value of the moving distance of the movable part corresponding to the model of the machine tool 10.
[0146] (4) The prediction device 20 of this disclosure includes: a learning completion model 250, which is generated by the machine learning device 30 of any one of (1) to (3); an input unit 201, which inputs manual feed status information of the current manual feed to the machine tool 10 that is capable of manual feed; and a prediction unit 202, which inputs the manual feed status information input by the input unit 201 into the learning completion model 250, and predicts the movement distance of the movable part of the machine tool 10 after a predetermined time from the present using the manual feed status information.
[0147] According to the prediction device 20, collisions of the movable parts of the machine tool 10 can be prevented without generating excessive alarms during manual feed.
[0148] (5) In the prediction device 20 described in (4), the prediction unit 202 can periodically predict the moving distance at time intervals shorter than a predetermined time.
[0149] In this way, the prediction device 20 can prevent collisions of the movable parts of the machine tool 10 with high precision.
[0150] (6) In the prediction device 20 described in (4) or (5), a learning completed model 250 may be provided in a server 50 that can be accessed from the prediction device 20 via a network 60.
[0151] In this way, even if a new machine tool 10, control device 15 and prediction device 20 are configured, the prediction device 20 can still apply the learning to complete the model 250.
[0152] (7) In any of the prediction devices 20 described in (4) to (6), there may be a machine learning device 30 of any of (1) to (3).
[0153] In this way, the prediction device 20 can achieve the same effect as any of (1) to (6) above.
[0154] (8) The control device 15 of this disclosure has a prediction device 20 of (4) to (7).
[0155] According to the control device 15, the same effect as any of (4) to (7) above can be obtained.
Claims
1. A machine learning device, characterized in that, have: The status observation unit acquires manual feed status information as input data, which includes the manual feed pulse waveform during any manual feed operation performed by a machine tool capable of manual feed. A tag acquisition unit acquires tag data, which represents the distance the movable part of the machine tool has moved within a predetermined time period following the manual feed pulse waveform of the manual feed state information contained in the input data; and The learning unit uses the input data obtained by the state observation unit and the label data obtained by the label acquisition unit to perform supervised learning and generate a learned model.
2. The machine learning apparatus according to claim 1, characterized in that, The manual feed status information includes one of the following: the distance to the interference object during the manual feed operation, the user identification information of the user performing the manual feed operation, the date and time of the manual feed operation, and the axis number of the operation performed through the manual feed operation.
3. The machine learning apparatus according to claim 1, characterized in that, The status observation unit obtains the input data according to the model of the machine tool. The tag acquisition unit acquires the tag data according to the model of the machine tool. The learning unit uses the input data and the tag data for each model of the machine tool to generate a learning completion model according to the model of the machine tool.
4. A prediction device, characterized in that, have: The learning completes the model, which is generated by the machine learning apparatus according to any one of claims 1 to 3; The input unit, for machine tools capable of manual feed, inputs the manual feed status information of the currently performed manual feed; and The prediction unit inputs the manual feed status information from the input unit into the learning completion model, and uses the manual feed status information to predict the movement distance of the movable part of the machine tool after a predetermined time from the present.
5. The prediction device according to claim 4, characterized in that, The prediction unit periodically predicts the travel distance at time intervals shorter than the predetermined time.
6. The prediction device according to claim 4 or 5, characterized in that, The learned model is contained in a server that can be accessed from the prediction device via a network.
7. The prediction device according to claim 4, characterized in that, The prediction device has the machine learning device according to any one of claims 1 to 3.
8. A control device, characterized in that, have: The prediction device according to any one of claims 4 to 7.