LEARNING DEVICE, NUMERICAL CONTROL DEVICE, MACHINING RESULT PREDICTING METHOD AND MACHINING SYSTEM

Abstract

A learning device (1) includes, when a model is used in a second machining device (4), the model being learned by associating sensor information acquired by a detection unit (33) of a first machining device (3) and labels corresponding to the sensor information, the first machining device including the detection unit that detects any one of a sensor signal and a feature of the sensor signal as sensor information during machining, the sensor signal indicating a detection result of at least one of a state of a workpiece and a state of a machining device: a recording unit (11) that records, as first reference information, the sensor information when machining is performed with a reference machining parameter by the first machining device;An information acquisition unit (12) that acquires, as second reference information, the sensor information when machining with the reference machining parameter is performed by the second machining device; and a correction function learning unit (13) that learns a correction function (44) to be used when the sensor information is subjected to correction and then input into the model in the second machining device for correction, so that a feature of the second reference information after the correction is caused to approach a feature of the first reference information.

Description

Area

[0001] The present disclosure relates to a learning apparatus that generates a learning model to be applied to an industrial machine, a numerical control apparatus, a method for predicting a machining result, and a machining system. General state of the art

[0002] As an example of a conventional device implemented by applying a learning model, for example, Patent Literature 1 discloses an abnormality determination device including: an observation data acquisition unit that acquires observation data regarding an operation observed during the operation of an industrial machine; a correction unit that corrects the observation data depending on an operating condition of the industrial machine; a statistics extraction unit that extracts, from the observation data, partial time series data including a portion in which a feature of an operating condition appears at a predetermined time point, and calculates at least one statistic from the partial time series data;and a machine learning device that executes a machine learning process for determining an operation abnormality of the industrial machine based on the statistics calculated by the statistics extraction unit; List of citationsPatent literature

[0003] Patent Literature 1: Japanese Patent Application Laid-Open No. 2021-15573 Brief description of the inventionProblem to be solved by the invention

[0004] When generating a learning model, it is necessary to prepare a large amount of learning data to perform machine learning, which requires a lot of time and effort. Therefore, regarding a learning model obtained by performing machine learning using learning data collected by operating an industrial machine, its use in another industrial machine of the same type was investigated. By generating a learning model using learning data obtained from one industrial machine and using the learning model in a plurality of industrial machines of the same type, the need to perform machine learning for each industrial machine to generate a learning model is eliminated, and the time and effort required to create a learning model can be reduced.

[0005] On the other hand, it is known that when using a learning model, a small difference in the input to the learning model can greatly affect the output of the learning model. For example, in a case where a learning model is used in predicting an operating result of a machining device, even if the same machining parameter is set on machining devices of the same construction and machining is performed, a difference will occur between the acquired data due to a difference in the characteristics between tools used for machining, a difference in the characteristics between workpieces, an individual difference between sensors, each of which is a detection unit for inputting data to the learning model, and the like.Therefore, in a case where a learning model generated by performing machine learning using data collected from a certain machining device is installed on another machining device, errors in the output of the learning model increase, which may lead to inaccuracy.

[0006] In the anomaly determination device described in Patent Literature 1, when a specific industrial machine learns observation data, the industrial machine is operated to correct the acquired observation data depending on each of a variety of operating conditions, and a learning model is generated using the corrected observation data. Therefore, learning can be completed in a small number of times.However, in a case where a learning model generated by a certain industrial machine is applied to another industrial machine, a difference occurs between observation data to be acquired for a determination process of industrial machines, which is due to a difference in the characteristics of tools used in machining by the industrial machines, a difference in the characteristics of workpieces, an individual difference between sensors that observe processes, and the like. As a result, a problem arises in that errors in the output of the learning model increase, leading to an inaccurate determination result.

[0007] The present disclosure has been made in view of the foregoing, and one of its objects is to provide a learning apparatus capable of preventing the determination accuracy of the learning model from decreasing in the other industrial machine when a learning model created by learning an operation of a certain industrial machine is applied to another industrial machine. Ways to solve the problem

[0008] In order to solve the above problem and achieve an object, a learning device according to the present disclosure includes, when a model is used in a second processing device, the model is learned by associating sensor information acquired by an observation unit of a first processing device with labels corresponding to the sensor information, wherein the first processing device is a processing device that performs processing according to processing parameters and includes the observation unit that acquires any one of a sensor signal and a feature of the sensor signal as sensor information during processing, wherein the sensor signal indicates an observation result of at least one of a state of a workpiece and a state of a processing device, wherein the second processing device is the processing device,which is different from the first machining device: a recording unit for recording, as first reference information, the sensor information when machining with a reference machining parameter, which is one of the machining parameters, is performed by the first machining device; a reference information acquisition unit for acquiring, as second reference information, the sensor information when machining with the reference machining parameter is performed by the second machining device; and a correction function learning unit for learning a correction function to be used when the sensor information is subjected to correction and then input to the model in the second machining device for correction, so that a feature of the second reference information after the correction is caused to approach a feature of the first reference information. Effects of the invention

[0009] The present disclosure achieves an effect in that it is possible to realize a learning device that, when a learning model created by learning an operation of a certain industrial machine is applied to another industrial machine, is capable of preventing the determination accuracy of the learning model from decreasing in the other industrial machine. Short description of the drawings Fig. 1 is a diagram illustrating an exemplary configuration of a machining system according to a first embodiment. Fig. 2 is a diagram illustrating a method of creating a model used in the machining system according to the first embodiment. Fig. 3 is a diagram illustrating an exemplary configuration of an observation unit of a first processing apparatus. Fig. 4 is a diagram illustrating an exemplary configuration of an observation unit of a second processing apparatus. Fig. 5 is a flowchart illustrating an example of an operation of a learning device. Fig. 6 is a diagram for explaining an example of an operation in which a correction function learning unit of the learning device selects data to be used for learning a correction function. Fig. 7 is a diagram illustrating an example of hardware implementing the learning device. Fig. 8 is a diagram illustrating an exemplary configuration of an observation unit included in a first machining device of a machining system according to a second embodiment. Fig. 9 is a diagram illustrating an exemplary configuration of an observation unit included in a second machining device of the machining system according to the second embodiment. Fig. 10 is a diagram illustrating an exemplary configuration of a machining system according to a third embodiment. Description of embodiments

[0010] Hereinafter, a learning apparatus, a numerical control apparatus, a method for predicting a machining result, and a machining system according to each embodiment of the present disclosure will be described in detail with reference to the drawings. First embodiment.

[0011] Fig. 1 is a diagram illustrating an exemplary configuration of a machining system 100 according to a first embodiment. The machining system 100 includes a learning device 1, a first machining device 3, and a second machining device 4. The first machining device 3 and the second machining device 4 are machining devices of the same type. The term "same type" here means that the machining devices have similar functions and can perform similar machining on workpieces that are machining target objects. The content of the machining performed by the first machining device 3 and the second machining device 4 follows the setting of the machining parameter, which will be described later. Although details will be described later, in the Fig. In the machining system 100 illustrated in FIG. 1, when the learning device 1 performs learning, a reference machining parameter 2 is set as the same machining parameter for the first machining device 3 and the second machining device 4. When the learning device 1 does not perform learning, a machining parameter is set individually for each of the first machining device 3 and the second machining device 4.

[0012] The learning device 1 includes a recording unit 11, a reference information acquisition unit 12, and a correction function learning unit 13. The first machining device 3 includes a tool 31, a workpiece 32, and an observation unit 33. The second machining device 4 includes a tool 41, a workpiece 42, an observation unit 43, a correction function 44, and a model 45. It should be noted that each of the first machining device 3 and the second machining device 4 includes a numerical control device and a drive unit for changing the relative positions of the tool and the workpiece, which are not illustrated. Although the model 45 in the second machining device 4 is Fig. 1, a device different from the second machining device 4 may include the model 45. For example, a device having a function of predicting a result of machining of the workpiece 42 by the second machining device 4 may include the model 45, and this device may predict the machining result using the model 45.

[0013] Fig. Fig. 2 is a diagram illustrating a method for creating the model 45 used in the machining system 100 according to the first embodiment. The model 45 is created by a model creation device 50 based on sensor information obtained by the Fig. 1, and generates labels. The model creation device 50 includes a database 5 in which the sensor information and the labels are registered, and a model learning unit 6 that creates the model 45 by learning a relationship between the sensor information and the labels registered in the database 5. The first processing device 3 and the model creation device 50 may be directly connected by a communication cable or the like, or may be connected via a communication network.

[0014] The observation unit 33 of the first processing device 3 and the observation unit 43 of the second processing device 4 have configurations shown in Fig. 3 and 4 respectively. Fig. Fig. 3 is a diagram illustrating an exemplary configuration of the observation unit 33 of the first processing device 3, and Fig. 4 is a diagram illustrating an exemplary configuration of the observation unit 43 of the second processing device 4. As in Fig. 3, the observation unit 33 of the first processing device 3 includes a sensor 331 and observes the state of the first processing device 3 using the sensor 331. The sensor 331 outputs a sensor signal indicating an observation result. The observation unit 33 can observe the state of the workpiece 32 and use a result of the observation as an observation result obtained by the first processing device 3. That is, the sensor signal output from the sensor 331 can indicate the observation result for the workpiece 32. The observation unit 43 of the second processing device 4 includes a sensor 431 and observes the state of the second processing device 4 using the sensor 431. The sensor 431 outputs a sensor signal indicating an observation result.The observation unit 43 can observe the state of the workpiece 42 and use a result of the observation as an observation result obtained by the second processing device 4. That is, the sensor signal output from the sensor 431 can indicate the observation result for the workpiece 42. Although the sensor 331 and the sensor 431 belong to the same type of sensor, there is an individual difference between them, and even when the same object is observed under the same condition, there may be a difference between an observation result obtained by the sensor 331 and the observation result obtained by the sensor 431.

[0015] Here, concrete examples of a machining device used as the first machining device 3 and the second machining device 4 include a turning device, a milling device, a grinding device, a laser machining device, an electrical discharge machining device, a water jet machining device, and a press machining device.

[0016] The state of a machining device is the condition of the machining device expressed or observed as data, and examples include whether a drive unit is normal or abnormal, a current flowing through the drive unit, the magnitude of a voltage applied to the drive unit, sound or light generated at the time of machining, a temperature, a position, a speed, an acceleration, an angle, an angular velocity, an angular acceleration, a pressure and the amount of deformation of a concrete part, the degree of wear of a tool, and an image of the concrete part.

[0017] The state of the workpiece 32 is a value obtained by observing how the workpiece 32 is being machined. Examples include an acceleration that can be detected by an acceleration sensor attached to the workpiece 32, an angular velocity that can be detected by a gyro sensor, an observation value obtained by an optical sensor when light emission occurs during machining, and a temperature of the workpiece. The same applies to the state of the workpiece 42.

[0018] The elements to be observed by the observation unit 33 and the observation unit 43 and their values ​​are usually subject to mapping to elements to be predicted by outputs of the model 45. Therefore, it is necessary to determine the type of each sensor used in the observation unit 33 and the observation unit 43 depending on the elements desirably predicted by the model 45 and the type of machining device to which the model 45 is applied. In addition, depending on a purpose, a plurality of sensors may be used in combination as sensors used in the observation unit 33 and the observation unit 43.

[0019] A machining parameter is a set of variables associated with a machining operation set in a machining device. Examples of values ​​set as a machining parameter include a machining speed indicating the speed of the machining operation, a limit value for the acceleration of a moving section, a value specifying a relationship between the relative positions of a tool and a workpiece, and fixed values ​​for the amounts of cutting fluid and oil used to assist machining.

[0020] In one example, when the machining device is a rotary device, values ​​of a machining speed, cutting amount, a rotational speed of a workpiece, the amount of cutting fluid, and the like are set for the machining parameter as a set of variables. In another example, when the machining device is a laser machining device, a laser output, a focus position, a beam shape, a machining speed, a machining gas type, a machining gas pressure, a nozzle height, and the like are set for the machining parameter as a set of variables.

[0021] Part or all of the learning device 1 may be included in the first processing device 3 or provided outside the first processing device 3. Part or all of the learning device 1 may be provided in a numerical control device that constitutes the first processing device 3. Part or all of the learning device 1 may be provided in the second processing device 4. Alternatively, a configuration may be adopted in which the first processing device 3 is connected to a cloud via a network, and part or all of the learning device 1 is implemented by a processing circuit of a cloud server.

[0022] The first machining device 3, which is a first machining device, machines the workpiece 32 using the tool 31 according to a machining parameter and a machining program. The observation unit 33 detects the state of at least one of the workpiece 32 and the first machining device 3 as sensor information.

[0023] As in Fig. As illustrated in FIG. 2, the model 45 is generated by the model learning unit 6 of the model creation device 50 based on the sensor information collected by the first processing device 3 and the labels. Specifically, the first processing device 3 processes the workpiece 32 and records and accumulates the sensor information acquired by the observation unit 33 and the labels corresponding to the sensor information in the database 5 of the model creation device 50. The model learning unit 6 outputs the model 45 that has learned the relationship between the sensor information and the labels accumulated in the database 5. When the model learning unit 6 learns the relationship between the sensor information and the labels, features can be extracted from the sensor information, and a relationship between the features of the sensor information and the labels can be learned.When the relationship between the features of the sensor information and the labels has been learned, a feature is extracted with respect to sensor information given at the time of using the model 45 or corrected sensor information obtained by correcting the sensor information, and the extracted feature is input to the model 45 to obtain an output.

[0024] Model 45 can be any general machine learning model, an example of which is a neural network. Other examples include a branching tree, a support vector machine, and Gaussian process regression. Depending on the purpose of model 45, a regression model and a classification model can be used for general purposes.

[0025] Furthermore, an example of the output of Model 45 is the quality of machining performed by the machining device from which the sensor information to be input to Model 45 is acquired, and the quality is quantified. Other examples include the degree of abnormality of the machining device from which the sensor information to be input to Model 45 is acquired, abnormality determination information of the machining device, the degree of tool wear, and a value of the machining parameter.

[0026] Regarding the labels to be recorded in the database 5, it is necessary to record the labels corresponding to outputs of the model 45, and the sensor information acquired by the observation unit 33 must be information related to the labels.

[0027] Furthermore, the observation unit 33 generally samples a sensor signal acquired by the sensor 331 at regular time intervals and outputs the sensor signal as sensor information. The observation unit 33 may further determine whether the machining operation is being performed and output the sensor information only during the machining operation, or may output the sensor information when the start or end of a specific operation is detected. The same applies to the observation unit 43 of the second machining device 4 using the model 45.

[0028] As described above, Model 45 learns the relationship between the sensor information and the labels. Specific examples of the sensor information, the labels, and Model 45 are described.

[0029] As an example, a case will be described in which the first machining device 3 and the second machining device 4 are rotary devices.

[0030] In a rotary device, a workpiece is fixed to a rotating spindle, and a tool is pressed against the workpiece to thereby remove a chip from the workpiece. At this time, a vibration called "chatter" may occur depending on a material, a rotation speed, and the amount of cutting of the workpiece. To detect this "chatter," an acceleration sensor is attached to the tool to collect sensor information. When collecting the sensor information, it is desirable to acquire different types of sensor information at the time of "chatter" occurrence and at the time of normal machining by changing the machining parameters differently. The sensor information at the time of "chatter" occurrence is labeled "chatter," and the sensor information at the time of normal machining is labeled "normal."At this time, the acquired sensor information is time series data, and labels are assigned to the time series data, that is, the sensor information, based on the time of occurrence of chatter. Therefore, by sampling time series data having a constant time width, a data pair can be obtained from the sensor information in which the thus sampled time series data and a label correspond to each other. The model 45 learns a relationship of the data pair, that is, a relationship between the thus sampled time series data and the label, thereby being able to learn a relationship between time series data of the acceleration sensor attached to the tool and chatter.After the learning of the model 45 is completed, the time series data of the acceleration sensor mounted on the tool, i.e., the sensor information, is input into the model 45 in the rotary device, thereby obtaining an output indicating whether chatter has occurred, and chatter can be automatically detected. Furthermore, machining can be stopped or the machining parameter can be changed when chatter occurs.

[0031] In addition, as another example, a case will be described where the first processing device 3 and the second processing device 4 are laser cutting devices, which are one kind of laser processing devices.

[0032] A laser cutting device cuts a workpiece by irradiating the workpiece with a laser focused near a surface of the workpiece. Therefore, in a laser processing device such as a laser cutting device, the laser is equivalent to a tool. When the workpiece is irradiated with the laser, it is heated and emits light, so time-series data from an optical sensor that records the light emission from the workpiece can be used as sensor information to detect the machining state. Examples of machining results at the time of laser cutting include normal machining, slag (molten material adhering to a back surface of the workpiece), scratch (unevenness appearing on a cut surface), burn (not penetrating a cut portion), and melt blowout.At the time of collecting the sensor information, it is desirable to acquire sensor information pieces each corresponding to one of different machining results by changing the machining parameter in various ways. Then, the type of the machining result is associated as a label. In a case where a label is associated with the time series data of the optical sensor, a machining result and a time are associated with each other from time series data of a machining position and a machining result for the machining position, and then the time of the optical sensor and the time of the machining result are associated with each other, thereby associating the machining results, that is, the labels, and the sensor information, that is, the time series data of the optical sensor.By sampling time series data having a constant time width from the sensor information, a data pair can be obtained in which the thus sampled time series data and the annotation correspond to each other. The model 45 learns a relationship of the data pair, that is, a relationship between the thus sampled time series data and the annotation, thereby being able to learn a relationship between time series data of the optical sensor and the machining results. After the learning of the model 45 is completed, in the laser cutting device, the time series data of the optical sensor that detects the light emission of the workpiece, that is, the sensor information, is input to the model 45, and thereby a prediction of a machining result can be obtained as an output during machining.Furthermore, in a case where the predicted machining result is not favorable, the machining may be stopped or the machining parameter may be changed.

[0033] In addition, as another example, a case will be described in which the first machining device 3 and the second machining device 4 are sinking electric discharge machining devices, which are a kind of electric discharge machining devices.

[0034] In the die-cutting electrical discharge machining device, a workpiece is machined by an electrical discharge phenomenon generated by bringing a discharge electrode closer to the workpiece and applying a voltage between the workpiece and the discharge electrode. Therefore, in the die-cutting electrical discharge machining device, the discharge electrode and the electrical discharge phenomenon to be generated correspond to a tool.

[0035] The sinking electrical discharge machining device includes a drive unit for changing the relative distance between the discharge electrode and the workpiece. An example of the drive unit is a linear motor, and in this case, the position of the discharge electrode is detected by a linear encoder. In this example, the current flowing through the linear motor and the position of the discharge electrode are used as sensor information.

[0036] As an abnormality of the sinking electrical discharge machining device, for example, there may be a case where the position of the discharge electrode cannot be properly adjusted due to a lack of grease in a sliding portion of the linear motor, abnormal heat generation of the motor, or the like. To detect such an abnormality, sensor information is acquired in a state where the abnormality is reproduced and labeled as abnormal. In addition, sensor information is acquired at the time of normal operation and labeled as normal. Then, the model 45 is caused to learn a mapping relationship between the sensor information and the labels.After the learning of Model 45 is completed, it is possible to detect whether the linear motor is normal or abnormal in the electric discharge machining device by inputting the current of the linear motor and the position of the discharge electrode, that is, the sensor information, to Model 45.

[0037] By setting the observation unit and the labels depending on the characteristics of the processing device, collecting the sensor information indicating the observation result and the labels, and causing the Model 45 to learn the relationship between the sensor information and the labels, the Model 45 can be used in various applications as described above.

[0038] The second machining device 4 as a second machining device is a machining device that uses the model 45 created using the first machining device 3. The second machining device 4 is a machining device similar to and different from the first machining device 3, and machines the workpiece 42 using the tool 41 according to a machining parameter and a machining program. The observation unit 43 of the second machining device 4 observes sensor information similar to that of the observation unit 33 of the first machining device 3.

[0039] The second processing device 4 further includes the correction function 44 generated by the learning device 1 and the model 45 generated using the first processing device 3. In the second processing device 4, in a case where the correction function 44 is generated by the learning device 1, the correction function 44 is applied to the sensor information acquired by the observation unit 43 to obtain corrected sensor information. Then, the corrected sensor information is input to the model 45 to obtain an output of the model 45. Note that in a case where the correction function 44 is not provided, or in a case where a user selects not to use the correction function 44, the sensor information acquired by the observation unit 43 is directly input to the model 45 to obtain an output of the model 45.In a case where the user is allowed to select whether to use the correction function 44, the second processing device 4 further includes an input unit that receives a selection by the user.

[0040] By allowing the user to select whether to use the correction function 44, for example, in a case where the correction by the correction function 44 does not correspond to a purpose of using the model 45, it is possible to immediately return to the original state in which the correction function 44 is not used. At this time, it is desirable that the second processing device 4 includes a display unit that performs a comparative display of outputs, iean output obtained as a result of inputting corrected sensor information, which is sensor information after correction obtained by applying the correction function 44 to the sensor information, to the model 45, and an output obtained as a result of inputting the sensor information to the model 45 so that the user can determine whether the correction of the sensor information by the correction function 44 is appropriate.

[0041] Alternatively, a plurality of candidates for the correction function 44 are created, a candidate that provides a suitable output of the model 45 is selected by the user from among the candidates, and the selected candidate can be used as the correction function 44. Consequently, the user can select the correction function 44 that is more suitable for the purpose of using the model 45.

[0042] Furthermore, by providing a storage unit in the second machining device 4 and storing the sensor information acquired by the observation unit 43 therein, it is possible to check whether correction by the correction function 44 is appropriate using the stored sensor information. Providing the storage unit that stores the sensor information can skip machining to check whether correction by the correction function 44 is appropriate, and can reduce the consumption of the workpiece and wear of the tool. Furthermore, when a plurality of pieces of sensor information are stored in the storage unit, the output of the model 45 can be checked in a plurality of machining states, and the correction function 44 can be selected more appropriately.

[0043] In the Fig. In the machining system 100 illustrated in FIG. 1, the second machining device 4 includes the model 45, and the first machining device 3 does not include the model 45. However, the first machining device 3 may also include the model 45, and the model 45 may be used to perform, for example, a determination of the machining quality by the first machining device 3. Since the model 45 has learned the correspondence relationship between the sensor information output from the observation unit 33 of the first machining device 3 and the labels, the first machining device 3 does not include a correction function similar to the correction function 44 included in the second machining device 4, or the correction function included in the first machining device 3 outputs the input sensor information as it is without correcting the input sensor information.

[0044] The learning device 1 learns the correction function 44 used by the second processing device 4. Fig. 5 is a flowchart illustrating an example of an operation of the learning device 1. The learning device 1 repeatedly executes the Fig. 5, steps S1 to S3 are carried out to learn the correction function 44.

[0045] When learning the correction function 44, the learning device 1 first acquires an observation result of an operation of the first machining device 3 (step S1). Specifically, the learning device 1 acquires sensor information output from the observation unit 33 when the first machining device 3 is operated with the reference machining parameter 2 as a machining parameter, and records the acquired sensor information in the recording unit 11 as first reference information.

[0046] Next, the learning device 1 acquires an observation result of an operation of the second machining device 4 in which the same machining parameter as that of the first machining device 3 is set (step S2). Specifically, the reference information acquisition unit 12 in the learning device 1 acquires, as second reference information, sensor information output from the observation unit 43 when the second machining device 4 is operated with the reference machining parameter 2 as the machining parameter.

[0047] Next, the learning device 1 learns a correction function used in a process of correcting the observation result of the operation of the second processing device 4 and causing it to converge with the observation result of the operation of the first processing device 3 (step S3). Specifically, in the learning device 1, in a case where the correction function learning unit 13 applies the correction function 44 to the second reference information to acquire the second reference information after correction, the correction function 44 is learned so that features of the second reference information after correction are caused to converge with features of the first reference information.

[0048] Here, the sensor information recorded as the first reference information in the recording unit 11, among the sensor information items acquired by the model creation device 50 for creating the model 45, may be sensor information in a case where the machining parameter used at the time of machining is the same as the reference machining parameter 2. In addition, a machining parameter may be selected from machining parameters corresponding to the sensor information acquired for creating the model 45 and may be set as the reference machining parameter 2. Using the already acquired sensor information to create the model 45 allows the first machining device 3 to skip re-executing the machining in which the reference machining parameter 2 is set.

[0049] In addition, in a case where the type of sensor information changes due to a difference between machining programs and the change affects the output of the model 45, it is desirable to prepare a reference machining program that is a machining program similar to the reference machining parameter 2 used as a reference, and acquire the first reference information and the second reference information using the reference machining parameter 2 and the reference machining program.Using the reference machining program can eliminate a difference between the first reference information and the second reference information caused by a difference between the machining program used by the first machining device 3 when acquiring the first reference information and the machining program used by the second machining device 4 when acquiring the second reference information. Consequently, in a case of using the correction function 44 in the second machining device 4, it is possible to learn a more accurate correction function 44 and further reduce errors in the output of the model 45 than in a case where different machining programs are used.

[0050] The operation of the correction function learning unit 13 of the learning device 1 according to the first embodiment will be described in more detail. In the first embodiment, the first reference information and the second reference information are time series data of a sensor signal.

[0051] There may be various methods in which the correction function learning unit 13 learns the correction function 44 so as to cause the features of the second reference information after the correction to approach the features of the first reference information.

[0052] In one example, the correction function 44 may be a low-pass filter for the time series data. Regarding other examples of the correction function 44, there is a method in which the correction function 44 is a high-pass filter or a band-pass filter, and a method in which the correction function 44 is a function that performs multiplication processes by a constant, addition of an offset, and the like. It is also possible to combine these methods.

[0053] When learning the correction function 44, in one example, the correction function 44 is learned by adding an offset to the second reference information so that an average value of the second reference information after correction matches an average value of the first reference information. Other examples include a method of learning the correction function 44 that multiplies the second reference information so that the average value of the second reference information after correction matches the average value of the first reference information, and a method of learning parameters of the correction function to cause the feature extracted from the second reference information after correction and the feature extracted from the first reference information to converge.

[0054] Here, the feature is a quantity that indicates a characteristic of a sensor signal, and is, for example, a statistical measure such as an average value, a variance value, a standard deviation value, a median value, a maximum value, a minimum value, a crest factor, or the number of peaks of the sensor signal. Additionally, the feature may be a frequency analysis result, a filter bank analysis result (frequency analysis results summarized in a specific frequency range), a cepstrum analysis result, or the like. A general feature extraction method can be used to extract these features.In a case where a plurality of values ​​are used as the feature, it is only required that, for example, parameters of the correction function 44 are learned so that the sum of absolute values ​​of differences between respective features extracted from the second reference information after correction and respective features extracted from the first reference information is reduced, or parameters of the correction function 44 are learned so that the weighted sum of squares of the differences between the respective features extracted from the second reference information after correction and the respective features extracted from the first reference information is reduced.

[0055] A general machine learning method can be used to learn the parameters of the correction function 44. That is, the correction function learning unit 13 determines the parameters of the correction function 44 through the machine learning method. For example, the correction function learning unit 13 may learn the parameters of the correction function through Newton's method so that the sum of squares of differences between the features extracted from the second reference information after correction and the features extracted from the first reference information is reduced, or it may use a general optimization method such as a conjugate gradient method, a Bayesian optimization method, a stochastic gradient descent method, particle swarm optimization (PSO), a random search method, or the like.Regarding the optimization method, a suitable method is selected taking into account the shape of the correction function 44 and the number of parameters, the number of features of the first reference information and the number of features of the second reference information after the correction and their nature.

[0056] In a case where the first reference information and the second reference information have different labels, it is not possible to determine whether a difference between the features of the first reference information and the features of the second reference information occurred due to a difference between the labels or due to individual differences between tools, workpieces, sensors, and the like of the first processing device 3 and the second processing device 4. If the difference occurred due to the difference between the labels, there may be a case where the individual differences between tools, workpieces, sensors, and the like of the first processing device 3 and the second processing device 4 cannot be corrected by the correction function 44.

[0057] Therefore, it is desirable to record the first reference information and the label in the recording unit 11 in association with each other, and to learn the correction function 44 in a case where the label corresponding to the second reference information acquired by the reference information acquisition unit 12 matches the recorded label corresponding to the first reference information. Consequently, it is possible to determine that the difference between the features of the first reference information and the features of the second reference information occurred due to the individual differences between tools, workpieces, sensors, and the like of the first machining device 3 and the second machining device 4.Consequently, the correction function learning unit 13 can generate the correction function 44 capable of correcting the influence of the individual differences between tools, workpieces, sensors, and the like of the first machining device 3 and the second machining device 4, that is, the difference between the sensor information acquired by the first machining device 3 and the sensor information acquired by the second machining device 4, more accurately than in a case where it is not checked whether the labels agree with each other.

[0058] There may be a plurality of first reference information items and a plurality of second reference information items. Hereinafter, in this document, the plurality of first reference information items are set as a first group, and the plurality of second reference information items are set as a second group. The sensor information sampled at regular time intervals in the observation unit 33 and the observation unit 43, or given first reference information and second reference information sampled at specific time intervals, may be set as the first group and the second group, respectively.In addition, even in a case where machining is performed a plurality of times with the reference machining parameter 2, and in a case where a plurality of sets of machining parameters are set as the reference machining parameter 2, there are a plurality of items of the first reference information and a plurality of items of the second reference information set as the first group and the second group, respectively.

[0059] When a plurality of sets of machining parameters are set as the reference machining parameter 2, machining is performed with respective machining parameters set as the reference machining parameter 2 by the first machining device and the second machining device to acquire respective items of the first reference information and respective items of the second reference information, that is, the first group and the second group.

[0060] The correction function learning unit 13 uses the first group and the second group to learn parameters of the correction function to cause the distribution of features extracted from the second reference information after correction to approximate the distribution of features extracted from the first reference information. That is, the correction function learning unit 13 determines the parameters of the correction function 44 so that a distance between the distributions, that is, a distance between the distribution of features extracted from the second reference information after correction and the distribution of features extracted from the first reference information, is reduced.

[0061] By using the plurality of pieces of first reference information and the plurality of pieces of second reference information, it is possible to make the features of the second reference information after correction and the features of the first reference information more closely approximate than in a case where only one piece of first reference information and only one piece of second reference information are used. Therefore, the correction function learning unit 13 can learn the accurate correction function 44, and errors in the output of the model 45 when the correction function 44 is used in the second processing device 4 can be further reduced.

[0062] In addition, since the correction function is learned based on machining in a plurality of different states by setting a plurality of sets of machining parameters as the reference machining parameter 2, the correction function learning unit 13 can learn the correction function 44 with high versatility compared to a case where one machining parameter is set as the reference machining parameter 2, and errors in the output of the model 45 when the correction function 44 is used in the second machining device 4 can be reduced.

[0063] Here, a general method can be used as the distance between distributions described above. Examples include Kullback-Leibler divergence (KL divergence), Jensen-Shannon divergence (JS divergence), histogram cut, L1 distance, L2 distance, Pearson distance, and relative Pearson distance. Furthermore, a general method for approximating the distance between distributions can be used when calculating the distance.

[0064] The distribution of features can also be considered as a probability density function. For example, a general probability density function estimation method, such as a histogram method, a kernel density estimation method, a maximum likelihood method for a parametric model, a Bayesian estimation method for a parametric model, or a method using a mixture distribution, can be used. A probability density function estimation method and a method for calculating a distance between distributions are selected together, and each of the selected methods is used by the learning unit 13 for the correction function.

[0065] Even if there are a plurality of pieces of first reference information and a plurality of pieces of second reference information, it is effective to check whether the labels match in the same way as described above. The correction function learning unit 13 extracts, as a first reference group and a second reference group, the first reference information and the second reference information associated with the same machining parameter and the same label from the first group and the second group, respectively.Then, the correction function learning unit 13 sets the distribution of the features of the first reference information included in the first reference group as the features of a first reference information group, sets the distribution of the features of the second reference information included in the second reference group as the features of a second reference information group, and learns the correction function 44 so that the features of the second reference information group after the correction are caused to approach the features of the first reference information group.

[0066] Consequently, it is possible to determine that the difference between the features of the first reference information and the features of the second reference information occurred due to the individual differences between tools, workpieces, sensors, etc. of the first machining device 3 and the second machining device 4. Consequently, the correction function learning unit 13 can generate the correction function 44 capable of correcting the influence of the individual differences between tools, workpieces, sensors, etc. of the first machining device 3 and the second machining device 4, that is, the difference between the sensor information acquired by the first machining device 3 and the sensor information acquired by the second machining device 4, more accurately than when not checking whether the labels match each other.

[0067] At this time, if there is an unevenness in the number of data items between the first reference group and the second reference group with respect to the sets of respective machining parameters as the reference machining parameter 2 and the labels, the unevenness in the number of data items will result in an unevenness in the distribution (probability density function) in a case where the distribution is obtained (probability density function), which may affect the learning of the correction function 44. That is, the learned correction function may not be able to accurately correct the individual differences between tools, workpieces, sensors, and the like of the first machining device 3 and the second machining device 4.

[0068] Therefore, the correction function learning unit 13 extracts the first reference information and the second reference information corresponding to the same machining parameter and the same label from the first group and the second group such that the first reference information and the second reference information have the same relationship, and sets the first reference information and the second reference information as the first reference group and the second reference group, respectively. For example, assuming that the machining parameters A and B are included in the reference machining parameter and a label assigned to the sensor information is any one of L1 and L2, in a case where the correction function learning unit 13 extracts the first reference information of the machining parameter A and the label L1, the first reference information of the machining parameter A and the label L2,the first reference information of the machining parameter B and the label L1 and the first reference information of the machining parameter B and the label L2 in a ratio of 1: 2: 1: 2, the correction function learning unit 13 further extracts the second reference information of the machining parameter A and the label L1, the second reference information of the machining parameter A and the label L2, the second reference information of the machining parameter B and the label L1 and the second reference information of the machining parameter B and the label L2 in a ratio of 1: 2: 1: 2. At this time, it is also possible to extract the first reference information and the second reference information corresponding to the same machining parameter and the same label,such that the number of elements of the first reference information and the number of elements of the second reference information are equal to each other, and to set the first reference information and the second reference information as the first reference group and the second reference group, respectively. Then, the correction function learning unit 13 sets the distribution of the features of the first reference information included in the first reference group as the features of a first reference information group, sets the distribution of the features of the second reference information included in the second reference group as the features of a second reference information group, and learns the correction function 44 so that the features of the second reference information group, after correction, are caused to approximate the features of the first reference information group.

[0069] Consequently, it is possible to eliminate the distribution unevenness due to the unevenness in the number of data items in the sets of corresponding machining parameters and labels between the first reference group and the second reference group. Therefore, compared to when the first reference information and the second reference information corresponding to the same machining parameter and label are extracted as the first reference group and the second reference group, respectively, the individual differences between tools, workpieces, sensors, and the like of the first machining device 3 and the second machining device 4 can be corrected more accurately than in a case where the labels are not checked to match each other.

[0070] With reference to Fig. 6, a concrete example of an operation is described in which the correction function learning unit 13 extracts the same number of pieces of the first reference information and the second reference information corresponding to the same machining parameter and the same label from the first group and the second group, respectively, and sets the first reference information and the second reference information as the first reference information group and the second reference information group, respectively. It should be noted that Fig. 6 is a diagram for explaining an example of an operation in which the correction function learning unit 13 of the learning device 1 selects data to be used for learning the correction function 44.

[0071] In the Fig. In the example illustrated in Figure 6, in the first group, six items of first reference information #1 to #6 are assigned to machining parameters A to C and labels 1 to 2. In the second group, four items of second reference information #1 to #4 are assigned to machining parameters A to B and labels 1 to 2.

[0072] In this case, when the first group and the second group are compared with each other, the first reference information and the second reference information associated with the same machining parameter and the same label in the first group and the second group are those where the machining parameter is A and the label is 1, and those where the machining parameter is B and the label is 2.

[0073] Since there are two pieces of reference information in each of the first group and the second group, in which the machining parameter is A and the label is 1, respectively, the correction function learning unit 13 adds them to the first reference information group and the second reference information group. Specifically, the correction function learning unit 13 extracts first reference information #1 and first reference information #3 from the first group, adds first reference information #1 and first reference information #3 to the first reference information group, extracts second reference information #1 and second reference information #2 from the second group, and adds second reference information #1 and second reference information #2 to the second reference information group.

[0074] Furthermore, regarding a piece of reference information where the machining parameter is B and the label is 2, one piece of the reference information is present in the first group and two pieces thereof are present in the second group, and therefore, in a case of extracting the same number of pieces thereof, the correction function learning unit 13 appropriately selects one of the pieces thereof from the second group. Note that the selection may be performed randomly or arbitrarily. Specifically, the correction function learning unit 13 extracts first reference information #5 from the first group, adds first reference information #5 to the first reference information group, and extracts second reference information #3 from the second group and adds second reference information #3 to the second reference information.

[0075] Furthermore, in a case where the data acquisition for learning the model 45 is performed by a plurality of first processing devices 3, that is, in a case where there are a plurality of first processing devices 3, with respect to the first reference information of each first processing device 3, it is also possible to extract the features from respective pieces of the first reference information, and the average value or the median of the features of respective pieces of the first reference information can be treated as the features extracted from the first reference information.In an example with features as the average value, the correction function learning unit 13 learns the parameters of the correction function 44 so that the average value extracted from the second reference information after the correction is caused to approach the average value of respective items of the first reference information.

[0076] When learning the correction function 44, in the present embodiment, it is not necessary to store the entire database 5 used at the time of creating the model 45, and only the initial reference information needs to be recorded in the recording unit 11 of the learning device 1. Therefore, learning can be performed even in a case where the storage capacity of the recording unit 11 is small.

[0077] As other methods for correcting an individual difference between the machining devices, there may be a method in which sensor information is corrected for learning depending on a machining device to which the model 45 is applied and the learning of the model 45 is performed again, and a method in which sensor information is acquired for each machining device and the learning of the model 45 is performed. However, in these methods, in order to perform the learning of the model 45, a large number of data items are recorded and learning processes are performed using the large number of data items. Therefore, a large memory capacity is required, a large computation load is incurred, and the processes take a long time.

[0078] In contrast, in the present embodiment, the learning of the correction function 44 is performed using the first reference information and the second reference information obtained by processing with the reference processing parameter 2, and can therefore be implemented with a small memory capacity. Furthermore, since the number of data items to be learned is small, the computation load is low, and the processes can be performed at high speed.

[0079] In addition, the correction function 44 can be learned so that outputs obtained as a result of inputting the first reference information and the second reference information after correction to the model 45 are consistent with each other. However, a nonlinear model such as a neural network or a branching tree is often used as the model 45. Therefore, when learning the correction function 44, in order to obtain the relationship between the output of the correction function 44 and the output of the model 45, it is necessary to obtain a large number of outputs of the model 45 with respect to outputs obtained as a result of inputting the second reference information to candidates for the correction function 44, which makes it difficult to learn the correction function 44.

[0080] On the other hand, since the learning of the correction function 44 in the present embodiment is performed with the first reference information and the second reference information obtained by machining using the reference machining parameter 2, it is not necessary to calculate the output of the model 45 for learning the correction function 44, and the first reference information and the second reference information to which the correction function 44 is applied directly correspond to each other. Therefore, compared with a case where the correction function 44 is learned so that outputs obtained as a result of inputting the first reference information and the second reference information after correction to the model 45 coincide with each other, the amount of calculation is reduced, and the correction function 44 can be easily learned.

[0081] Next, a hardware configuration of the learning device 1 will be described. Fig. Fig. 7 is a diagram illustrating an example of hardware implementing the learning device 1. The learning device 1 may be implemented by a processor 91, a memory 92, and a Fig. 7 illustrated interface circuit 93 can be implemented.

[0082] The processor 91 is a central processing unit (CPU), also referred to as a processing device, an arithmetic device, a microprocessor, a microcomputer, or a digital signal processor (DSP) or a system large-scale integration (LSI). The memory 92 is a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM, registered trademark), a hard disk drive, or the like. The interface circuit 93 is a circuit for the learning device 1 to exchange data with an external device, such as the first processing device 3 or the second processing device 4.The interface circuit 93 may include circuitry connected to a network and, for example, performing the transmission and reception of data to and from other devices over the network.

[0083] The reference information acquisition unit 12 and the correction function learning unit 13 of the learning device 1 are implemented by the processor 91, which executes programs for acting as these components. The programs for acting as the reference information acquisition unit 12 and the correction function learning unit 13 are stored in advance in the memory 92. The processor 91 reads the programs from the memory 92 and executes the programs, thereby acting as the reference information acquisition unit 12 and the correction function learning unit 13. The recording unit 11 is implemented by the memory 92.

[0084] It is assumed that the programs for acting as the reference information acquiring unit 12 and the correction function learning unit 13 are stored in advance in the memory 92, but this is not limited to this. The above-mentioned programs may be provided to a user of the learning device 1 in a state where they are written onto a recording medium such as a compact disc-ROM (CD-ROM) or a digital versatile disc-ROM (DVD-ROM) and installed onto the memory 92 by the user. In this case, the hardware implementing the learning device 1 further includes a reading device for reading the programs from the recording medium. The programs can be installed by connecting the reading device to the interface circuit 93. Furthermore, the above-mentioned programs may be provided from a server via the network.

[0085] As described above, the learning device 1 according to the present embodiment learns the relationship between the first reference information and the second reference information by using the first reference information, which is the sensor information obtained by observing the machining operation of the first machining device 3 in which the reference machining parameter 2 is set, and the second reference information, which is the sensor information obtained by observing the machining operation of the second machining device 4 in which the reference machining parameter 2 is set, and the learning device 1 generates the correction function 44 that corrects the sensor information obtained by the second machining device 4, so that the sensor information obtainedWhen the second machining device 4 performs the machining operation using the same machining parameter as that of the first machining device 3, it approximates the sensor information obtained by the first machining device 3. Consequently, the learning model that determines the operation result of the first machining device 3 or the like based on the sensor information can also be used to determine the operation result of the second machining device 4 or the like. That is, it is possible to prevent the determination accuracy of the learning model from deteriorating when the learning model that has performed learning is applied to the second machining device 4 by observing the operation of the first machining device 3. Second embodiment.

[0086] In a second embodiment, differences from the first embodiment will be described. It should be noted that in the figures used in the description of the present embodiment, components denoted by the same reference numerals and not described in detail are similar to those described in the first embodiment.

[0087] The observation unit 33 included in the first processing device 3 and the observation unit 43 included in the second processing device 4 of the processing system 100, which are described in the first embodiment, differ in the second embodiment. Specifically, the observation unit 33 is replaced by an observation unit 33a, which in Fig. 8, and the observation unit 43 is replaced by an observation unit 43a, which in Fig. 9 is illustrated.

[0088] In the following description, the machining system, the first machining device, the second machining device, and the learning device according to the second embodiment are referred to as a machining system 100a, a first machining device 3a, a second machining device 4a, and a learning device 1a, respectively, to be distinguished from those in the first embodiment.

[0089] Fig. 8 is a diagram illustrating an exemplary configuration of the observation unit 33a included in the first processing device 3a of the processing system 100a according to the second embodiment. Fig. 9 is a diagram illustrating an exemplary configuration of the observation unit 43a included in the second processing device 4a of the processing system 100a according to the second embodiment.

[0090] The observation unit 33a of the first processing device 3a includes the sensor 331 and a feature extraction unit 332, and the observation unit 43a of the second processing device 4a includes the sensor 431 and the feature extraction unit 332. Note that the content of the process is intended by the feature extraction unit 332, and the feature extraction unit 332 is denoted by the same reference numeral in the sense that a process to extract the feature is executed in both the observation unit 33a and the observation unit 43a. The feature extraction unit 332 may also be provided outside the observation units 33a and 43a, as long as the order of the processes performed on the sensor signals output from the sensors 331 and 431 does not change.

[0091] The feature extraction unit 332 of the observation unit 33a receives a sensor signal output from the sensor 331, extracts a feature from the sensor signal, and outputs the feature of the sensor signal as sensor information. Similarly, the feature extraction unit 332 of the observation unit 43a receives a sensor signal output from the sensor 431, extracts a feature from the sensor signal, and outputs the feature of the sensor signal as sensor information.

[0092] Furthermore, the observation units 33a and 43a generally sample sensor signals acquired by the sensors 331 and 431 at regular time intervals and output features of the thus sampled sensor signals as sensor information. The observation units 33a and 43a may further determine whether the machining operation is being performed and output the sensor information only during the machining operation, or may sample sensor signals and extract features when the start or end of a specific operation is detected and output the features as sensor information.

[0093] In the second embodiment, the sensor information is a feature of a sensor signal. Therefore, as described in the first embodiment, the model 45 that has learned the relationship between the sensor information and the labels is a model that has learned the relationship between the features of the sensor signals and the labels. Similarly, the correction function 44 is a function that corrects the features of the sensor signals. Therefore, the correction function 44 performs processes on the features of the sensor signals such as adding an offset with respect to each feature, multiplying by a constant, parallel movement in a feature space, rotating in the feature space, and a combination of these processes. The correction function 44 is desirably a function that is bijective to the feature space.

[0094] A method for creating the correction function 44 may be similar to that described in the first embodiment by interpreting “the features of the first reference information” and “the features of the second reference information” described in the first embodiment as “first reference information” and “second reference information,” respectively.

[0095] In the learning device 1a according to the second embodiment, the recording unit 11 records the first reference information which is the feature of the sensor signal measured using the reference machining parameter 2 by the first machining device 3a, the reference information acquiring unit 12 acquires the second reference information which is the feature of the sensor signal when machining is performed using the reference machining parameter 2 by the second machining device 4a, and the correction function learning unit 13 learns the correction function 44 so that the second reference information after the correction is caused to approach the first reference information.

[0096] As described above, in the processing system 100a according to the present embodiment, the observation unit 33a of the first processing device 3a and the observation unit 43a of the second processing device 4a each include the feature extraction unit 332 that extracts features from sensor signals. According to the learning device 1a of the second embodiment, it is satisfactory as long as sensor information, which is a feature of a sensor signal, is recorded in the recording unit 11. Since the data capacity of a feature of a sensor signal is generally smaller than the data capacity of the sensor signal, storage thereof can be performed with a storage capacity smaller than that in the case of storing the sensor signal.In addition, for a similar reason, in learning the correction function 44, the amount of computation required for learning is reduced and the processes can be performed at high speed compared to the case of learning the correction for the sensor signal.

[0097] Furthermore, when learning a correction function that corrects a sensor signal as in the first embodiment, learning is performed by applying a temporary correction function to the sensor signal, performing feature extraction, and then comparing the features. However, when the feature is directly corrected as in the second embodiment, it is not necessary to extract the feature for comparison. Therefore, in the second embodiment, learning of the correction function 44 can be performed easily compared to the case where the sensor signal is corrected as in the first embodiment. Third embodiment.

[0098] In a third embodiment, differences from the first and second embodiments will be mainly described. Note that in the figures used in the description of the present embodiment, components denoted by the same reference numerals and not described in detail are similar to those described in the first embodiment.

[0099] Fig. 10 is a diagram illustrating an exemplary configuration of a machining system 100b according to the third embodiment. The machining system 100b includes a learning device 1b, the first machining device 3, N (N is an integer of 1 or more) second machining devices 4-1 to 4-N, the database 5, and a generalized model 8. The second machining devices 4-1 to 4-N are each a machining device similar to the second machining device 4 described in the first embodiment (see Fig. 1). In Fig. 10, no components other than the observation unit 43 are shown with respect to the second processing devices 4-1 to 4-N. In the following description, the second processing devices 4-1 to 4-N may be collectively referred to as second processing devices 4.

[0100] The learning device 1b includes the recording unit 11, the reference information acquisition unit 12, the correction function learning unit 13, an inverse correction function generation unit 14, an individual simulation data generation unit 15, and a generalized model learning unit 16. The learning device 11b has a configuration obtained by adding the inverse correction function generation unit 14, the individual simulation data generation unit 15, and the generalized model learning unit 16 to the learning device 1 according to the first embodiment. In addition, the learning device 1b uses the database 5 described in the first embodiment.

[0101] Here, individual differences between tools, workpieces, sensors, and the like are considered to exist in respective second processing devices 4. Even if these individual differences exist, the learning device 1b according to the third embodiment creates the generalized model 8, which is a model with a small influence on the output and high generalization performance. Note that the generalized model 8 is applied to the second processing devices 4 and predicts an operation result of the second processing devices 4. An example of the operation result is a machining result.

[0102] Similar to the first embodiment, the recording unit 11 records the first reference information, which is sensor information acquired by the observation unit 33 of the first processing device 3. Although in Fig. 10, the first reference information is, similarly to the first embodiment, sensor information obtained when the first machining device 3 performs the machining operation with the reference machining parameter 2.

[0103] The reference information acquisition unit 12 acquires the sensor information output from the observation unit 43 of each of the second machining devices 4-1 to 4-N as the second reference information. The sensor information acquired as the second reference information is sensor information obtained when the second machining devices 4-1 to 4-N each perform the machining operation using the reference machining parameter 2.

[0104] The correction function learning unit 13 uses the first reference information recorded in the recording unit 11 and the second reference information acquired by the reference information acquiring unit 12 to learn a correction function by a learning method similar to that in the first embodiment.

[0105] The inverse correction function generation unit 14 generates an inverse correction function, which is an inverse function of the correction function learned by the correction function learning unit 13. Note that the correction function learned by the correction function learning unit 13 is an invertible function.

[0106] The individual simulation data creation unit 15 creates the individual simulation data using the inverse correction function generated by the inverse correction function creation unit 14. Specifically, the individual simulation data creation unit 15 extracts the sensor information and a corresponding label from the database 5 in which the sensor information acquired by the first processing device 3 and the labels are recorded, applies each inverse correction function to the sensor information to create individual simulation information, and creates the individual simulation data by associating it with the label corresponding to the sensor information. The individual simulation data created by the individual simulation data creation unit 15 is used by the unit 16 to learn a generalized model.

[0107] The generalized model learning unit 16 creates the generalized model 8 based on the sensor information registered in the database 5 and the individual simulation data created by the individual simulation data creation unit 15. Specifically, the generalized model learning unit 16 extracts the sensor information from the database 5, extracts the individual simulation information from the individual simulation data, and learns a relationship between the extracted sensor information and the extracted individual simulation information and the corresponding label, thereby creating the generalized model 8.

[0108] The generalized model 8 can be any general machine learning model, and one example is a neural network. Other examples include a branching tree and a support vector machine. Depending on the purpose of the model, a regression model and a classification model can be used for general purposes. Furthermore, the output of the generalized model 8 in one example is the quantified machining quality. Other examples include the degree of machining device abnormality, machining device abnormality determination information, the degree of tool wear, and a machining parameter value.

[0109] Part or all of the learning device 1b may be included in each second processing device 4 or may be provided outside each second processing device 4. Alternatively, a configuration may be adopted in which each second processing device 4 is connected to a cloud via a network, and all or part of the learning device 1b is implemented by a processing circuit of a cloud server. For example, when part of the learning device 1b is configured by the cloud server via a network connection, each of the recording unit 11, the reference information acquisition unit 12, the correction function learning unit 13, and the inverse correction function generation unit 14 may be provided in the cloud server connected to the network, or may be provided in each second processing device 4.The individual simulation data creation unit 15 and the generalized model learning unit 16 can also be freely arranged as long as the data of the database 5 can be used. Alternatively, a configuration may be adopted in which, with respect to each second processing device 4, the second reference information acquired by the reference information acquisition unit 12, the correction function learned by the correction function learning unit 13, or the inverse correction function created by the inverse correction function creation unit 14 are manually acquired and collected in an electronic computer connected to the database 5, and all or part of the processes of the learning device 1b are performed by the electronic computer.

[0110] The learning device 1b may be configured with at least one second processing device 4, but it is desirable to prepare more than two second processing devices 4 to configure the learning device 1b. This is to configure the learning device 1b to be able to learn individual differences between different second processing devices 4, and the more second processing devices 4 there are, the higher the generality is achieved for the generalized model 8 that can be created by the generalized model learning unit 16.

[0111] The generalized model 8 created by the learning device 1b is applied to the second processing devices 4 via a network or manually. The second processing devices 4 to which the generalized model 8 is applied may include the second processing device 4 from which the learning device 1b has not acquired the second reference information used to create the generalized model 8. That is, the generalized model 8 may be installed on a new second processing device 4.When the second processing device 4 includes the generalized model 8 instead of the model 45 described in the first embodiment, the correction function 44 of the second processing device 4 including the generalized model 8 outputs the sensor information input from the observation unit 43 as it is and inputs the sensor information into the generalized model 8.

[0112] As described above, in the learning device 1b according to the present embodiment, the inverse correction function generation unit 14 generates the inverse correction function, which is an inverse function of the correction function generated by the correction function learning unit 13 described in the first embodiment. The individual simulation data generation unit 15 generates the individual simulation data using the inverse correction function, the sensor information obtained by observing the state of the first processing device 3, and the labels associated with the sensor information. The generalized model learning unit 16 generates the generalized model 8 based on the sensor information and the individual simulation data.According to the present embodiment, even when a plurality of second processing devices 4 are provided and have individual differences, it is possible to reduce the influence of the individual differences to obtain a highly accurate output by applying the generalized model 8. That is, the generalized model 8 can be created with high generality.

[0113] When a part of the components of the learning device 1b, for example, the correction function learning unit 13, is included in each of the second processing devices 4-1 to 4-N, the learning device 1b acquires the correction function created by each of the second processing devices 4-1 to 4-N instead of the second reference information, and the inverse correction function creation unit 14 generates the inverse correction function using the acquired correction functions.When the correction function learning unit 13 and the inverse correction function learning unit 14 are included in each of the second machining devices 4-1 to 4-N, the learning device 1b acquires the inverse correction function created by each of the second machining devices 4-1 to 4-N instead of the second reference information, and the individual simulation data creating unit 15 creates the individual simulation data using the acquired inverse correction functions.

[0114] In the learning device 1b according to the present embodiment, only the second reference information, the correction function, or the inverse correction function is collected from at least one second processing device 4, and the number of data collected from each second processing device 4 is smaller than that when different types of sensor information are collected from each second processing device 4. That is, the capacity for data collected for learning can be reduced. Therefore, the time required for data collection is short, and the need for a large storage area is eliminated. In addition, since there is no need to form a high-speed network, it is easy to acquire data from each second processing device 4.

[0115] The configurations described in the above embodiments are merely examples and may be combined with other known techniques, the embodiments may be combined with each other, and part of the configurations may be omitted or modified without departing from the gist thereof. List of reference symbols

[0116] 1, 1b learning device; 2 reference machining parameters; 3 first machining device; 4, 4-1, 4-N second machining device; 5 database; 6 model learning unit; 8 generalized model; 11 recording unit; 12 reference information acquisition unit; 13 correction function learning unit; 14 inverse correction function creation unit; 15 individual simulation data creation unit; 16 generalized model learning unit; 31, 41 tool; 32, 42 workpiece; 33, 33a, 43, 43a observation unit; 44 correction function; 45 model; 50 model creation device; 100, 100b machining system; 331, 431 sensor; 332 feature extraction unit. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] JP 2021-15573

[0003]

Claims

[1] A learning device comprising, when a model is used in a second machining device, the model being learned by associating sensor information acquired by an observation unit of a first machining device with labels corresponding to the sensor information, wherein the first machining device is a machining device that performs machining according to machining parameters and includes the observation unit that acquires any one of a sensor signal and a feature of the sensor signal as sensor information during machining, the sensor signal indicating an observation result of at least one of a state of a workpiece and a state of a machining device, wherein the second machining device is the machining device different from the first machining device: a recording unit for recording, as first reference information, the sensor information when machining with a reference machining parameter, which is one of the machining parameters, is performed by the first machining device; a reference information acquisition unit for acquiring, as second reference information, the sensor information when machining with the reference machining parameter is performed by the second machining device; and a correction function learning unit for learning a correction function to be used when the sensor information is subjected to correction and then input to the model in the second processing device for the correction, so as to cause a feature of the second reference information after the correction to approach a feature of the first reference information. [2] The learning device according to claim 1, wherein the sensor information is a feature of the sensor signal. [3] Learning device according to claim 1, wherein the unit for learning a correction function performing the learning using a distribution of features of the first reference information as the feature of the first reference information and using a distribution of features of the second reference information as the feature of the second reference information. [4] Learning device according to one of claims 1 to 3, wherein the recording unit further records the labels corresponding to the first reference information, and the correction function learning unit selects from the recording unit the first reference information corresponding to a same label as a label associated with the second reference information acquired by the reference information acquiring unit, and performs the learning using the selected first reference information and the second reference information acquired by the reference information acquiring unit. [5] Learning device according to one of claims 1 to 3, wherein the reference machining parameter is a plurality of sets of machining parameters, the recording unit records, as a first group, the first reference information which is each piece of the sensor information when the first machining device performs machining with each machining parameter included in the reference machining parameter, and the labels each corresponding to each of the pieces of the first reference information, the reference information acquisition unit acquires, as a second group, the second reference information which is each piece of the sensor information when the second machining device performs machining with each machining parameter included in the reference machining parameter, and the labels corresponding to each of the pieces of the second reference information, and the unit for learning a correction function extracting from the first group and the second group the first reference information and the second reference information which are created when machining is performed with the same machining parameter and which are assigned to the same label as a first reference group and a second reference group, respectively, a distribution of features of the first reference information included in the first reference group is used as the feature of the first reference information, and performing the learning using a distribution of features of the second reference information included in the second reference group as the feature of the second reference information. [6] The learning device according to claim 5, wherein the correction function learning unit extracts, from the first group and the second group, the first reference information and the second reference information, respectively, which are created when machining is performed with the same machining parameter and which are associated with the same label so that the first reference information and the second reference information have an equal relationship, and sets the first reference information and the second reference information as the first reference group and the second reference group. [7] The learning device according to claim 5, wherein the correction function learning unit extracts from the first group and the second group an equal number of pieces of the first reference information and the second reference information, respectively, which are created when machining is performed with the same machining parameter and which are associated with the same label, and sets the first reference information and the second reference information as the first reference group and the second reference group. [8] Learning device according to claim 1, wherein the reference information acquisition unit acquires, as the second reference information, the sensor information when machining with the reference machining parameter is performed by the second machining device, wherein a number of the second machining device is at least one, the correction function learning unit learns the correction function for each of the second processing devices individually, the learning device includes: an inverse correction function generating unit for generating an inverse correction function which is an inverse function of the correction function; an individual simulation data creation unit for creating individual simulation data by creating individual simulation information by applying the inverse correction function to the sensor information acquired by the observation unit of the first processing device, and assigning the individual simulation information to labels corresponding to the sensor information used to create the individual simulation information; and a generalized model learning unit for learning a relationship between the sensor information acquired by the observation unit of the first processing device and the individual simulation information and the labels corresponding thereto to create a generalized model, and sets the generalized model as the model to be used by the second processing device. [9] A numerical control device comprising the learning device according to claim 1 or 8 and controlling a machining device acting as the second machining device, the numerical control device comprising: an input unit for receiving, from a user, a selection of whether to input into the model to be used in the second machining device the sensor information before the correction performed using the correction function or the sensor information after the correction performed using the correction function. [10] A numerical control device according to claim 9, comprising: a display unit for displaying, when the input unit receives the selection from the user, an output of the model when the sensor information is input before the correction is performed, and an output of the model when the sensor information is input after the correction is performed. [11] A numerical control device comprising the learning device according to claim 1 and controlling a machining device acting as the second machining device, wherein the numerical control device corrects the sensor information acquired by the second machining device in operation with the correction function and inputs the sensor information after the correction to the model to obtain an output. [12] A numerical control device comprising the generalized model created by the learning device according to claim 8 and controlling a machining device operated as the second machining device, wherein the numerical control device inputs the sensor information acquired by the second machining device in operation to the generalized model to obtain an output. [13] A method for predicting a machining result when a model is used in a second machining device, wherein the model is learned by associating sensor information acquired by an observation unit of a first machining device with labels corresponding to the sensor information, wherein the first machining device is a machining device that performs machining according to machining parameters and includes an observation unit that acquires any one of a sensor signal and a feature of the sensor signal as sensor information during machining, wherein the sensor signal indicates an observation result of at least one of a state of a workpiece and a state of a machining device, wherein the second machining device is the machining device that is different from the first machining device,wherein the method for predicting a machining result comprises: a step performed by a recording unit of recording, as first reference information, the sensor information when machining with a reference machining parameter, which is one of the machining parameters, is performed by the first machining device; a step performed by a reference information acquiring unit for acquiring, as second reference information, the sensor information when machining with the reference machining parameter is performed by the second machining device; a step, performed by a correction function learning unit, of learning a correction function to be used when the sensor information is subjected to correction and then input to the model in the second processing device for the correction, so as to cause a feature of the second reference information after the correction to approach a feature of the first reference information; a step performed by the correction function for correcting the sensor information acquired by the second processing device during processing; and a step performed by the model of predicting a result of processing by the second processing device based on the sensor information corrected by the correction function. [14] Processing system comprising: a first machining device for performing machining according to machining parameters, the first machining device including an observation unit that acquires any one of a sensor signal and a feature of the sensor signal as sensor information during machining, the sensor signal indicating an observation result of at least one of a state of a workpiece and a state of a machining device; a second processing device including an observation unit similar to the observation unit of the first processing device and a model that performs a prediction based on the sensor information acquired by the observation unit; and a learning device for, when a model is used in the second processing device, the model being learned by assigning the sensor information acquired by the observation unit of the first processing device to labels corresponding to the sensor information, and the sensor information acquired by the observation unit of the second processing device being subjected to correction and then inputted into the model, performing a learning of a correction function to be used for the correction, wherein the learning device includes: a recording unit for recording, as first reference information, the sensor information when machining with a reference machining parameter, which is one of the machining parameters, is performed by the first machining device; a reference information acquisition unit for acquiring, as second reference information, the sensor information when machining with the reference machining parameter is performed by the second machining device; and a correction function learning unit for learning the correction function so as to cause a feature of the second reference information after the correction to approach a feature of the first reference information.

Images