Learning device, numerical control device, machining result prediction method, and machining system

Abstract

A learning device (1) uses a model obtained by associating sensor information acquired by an observation unit of a first processing device (3) having an observation unit (33) with a tag corresponding to the sensor information and learning the sensor information in a second processing device (4). The observation unit acquires, as sensor information during machining, either a sensor signal indicating an observation result of at least one of a state of a workpiece and a state of a machining device, or a characteristic quantity of the sensor signal. The first reference information recording unit records, as first reference information, sensor information when machining performed by reference machining parameters is performed in a first machining device. A reference information acquisition unit (12) that acquires, as second reference information, sensor information when machining performed by the reference machining parameters is performed in the second machining device; and a correction function learning unit (13) that learns a correction function (44) used for correction when the sensor information is corrected in the second processing device and then input to the model such that the feature of the corrected second reference information approaches the feature of the first reference information.

Description

Technical Field

[0001] The present invention relates to a learning device, a numerical control device, a processing result prediction method and a processing system for generating a learning model applied to industrial machinery. Background Art

[0002] As an example of an existing device implemented by applying a learning model, for example, Patent Document 1 discloses an abnormality determination device, which comprises: an observation data acquisition unit, which acquires observation data related to the action observed during the action of an industrial machine; a correction unit, which corrects the observation data in accordance with the action conditions of the industrial machine; a statistic extraction unit, which extracts partial time series data including a portion showing characteristics of an action state under a predetermined specified timing from the observation data, and calculates at least one statistic based on the partial time series data; and a machine learning device, which performs machine learning processing related to the determination of abnormal action of the industrial machine based on the statistic calculated by the statistic extraction unit.

[0003] Patent Document 1: Japanese Patent Application Publication No. 2021-15573 Summary of the invention

[0004] When generating a learning model, it is necessary to prepare a large amount of learning data and perform machine learning, which requires a lot of time and work. Therefore, research is being conducted on using learning data collected by operating one industrial machine to perform machine learning, and using the obtained learning model in other industrial machines of the same type. By generating a learning model using learning data obtained from one industrial machine and using the learning model in multiple industrial machines of the same type, it is not necessary to perform machine learning for each industrial machine to generate a learning model, which can reduce the time and work for preparing the learning model.

[0005] On the other hand, when using a learning model, it is known that sometimes small differences in the input to the learning model can have a large impact on the output of the learning model. For example, when using a learning model to predict the action results of a processing device, even if the same processing parameters are set in a processing device of the same design and processing is performed, differences in the characteristics of the tools used in the processing, differences in the characteristics of the workpiece, individual differences in the sensors that constitute the data acquisition unit that inputs the learning model, etc., will cause differences in the obtained data. Therefore, when a learning model generated by machine learning using data collected by a certain processing device is installed on another processing device, it is possible that the error of the output of the learning model will become larger and inaccurate.

[0006] In the abnormality determination device described in Patent Document 1, when learning observation data through a certain industrial machine, the observation data obtained by causing the industrial machine to move is corrected in accordance with each of a plurality of action conditions, and a learning model is generated using the corrected observation data. Therefore, it is possible to end the learning with a small number of times. However, when the learning model generated by a certain industrial machine is applied to other industrial machines, differences will arise between the observation data obtained from each industrial machine for determination processing due to differences in the characteristics of the tools used in processing by each industrial machine, differences in the characteristics of the workpiece, individual differences in the sensors that observe the action, etc. As a result, the error in the output of the learning model becomes larger, and the determination result becomes inaccurate.

[0007] The present invention is proposed in view of the above situation, and its purpose is to obtain a learning device that can prevent the judgment accuracy of the learning model in other industrial machines from being reduced when a learning model generated by learning the action of a certain industrial machine is applied to other industrial machines.

[0008] In order to solve the above-mentioned problems and achieve the purpose, the learning device involved in the present invention uses a model learned by associating sensor information obtained by an observation part of a processing device having an observation part, i.e., a first processing device, and a label corresponding to the sensor information in a case where the model is used in a processing device different from the first processing device, i.e., a second processing device, wherein processing is performed according to processing parameters, and the observation part obtains a sensor signal representing an observation result of at least any one of a state of a workpiece and a state of the processing device during processing, and any one of a feature quantity of the sensor signal as sensor information, and comprises: a recording part, which records sensor information when processing is performed in the first processing device using one of the processing parameters, i.e., a reference processing parameter, as first reference information; a reference information acquisition part, which acquires sensor information when processing is performed in the second processing device using the reference processing parameter, as second reference information; and a correction function learning part, which learns a correction function used for correction when the sensor information is corrected in the second processing device and then input into the model in such a way that the characteristics of the corrected second reference information are close to the characteristics of the first reference information.

[0009] Effects of the Invention

[0010] According to the present invention, there is an effect of realizing a learning device that can prevent a decrease in the determination accuracy of the learning model in other industrial machines when a learning model generated by learning the operation of a certain industrial machine is applied to other industrial machines. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1This is a diagram showing a configuration example of a processing system according to the first embodiment.

[0012] Figure 2 This is a diagram showing a method of creating a model used by the processing system according to the first embodiment.

[0013] Figure 3 It is a diagram showing a configuration example of the observation section of the first processing device.

[0014] Figure 4 It is a diagram showing a configuration example of the observation section of the second processing device.

[0015] Figure 5 This is a flowchart showing an example of the operation of the learning device.

[0016] Figure 6 This is a diagram for explaining an example of an operation of selecting data used for learning a correction function by a correction function learning unit of a learning device.

[0017] Figure 7 This is a diagram showing an example of hardware for realizing the learning device.

[0018] Figure 8 This is a diagram showing a configuration example of an observation unit included in a first processing device of a processing system according to a second embodiment.

[0019] Fig. 9 This is a diagram showing a configuration example of an observation unit included in the second processing device of the processing system involved in the second embodiment.

[0020] Fig.10 This is a diagram showing a configuration example of a processing system according to the third embodiment. DETAILED DESCRIPTION

[0021] Hereinafter, a learning device, a numerical control device, a machining result prediction method, and a machining system according to embodiments of the present invention will be described in detail with reference to the accompanying drawings.

[0022] Implementation method 1.

[0023] Figure 1 1 is a diagram showing a configuration example of a processing system 100 according to Embodiment 1. The processing system 100 is composed of a learning device 1, a first processing device 3, and a second processing device 4. The first processing device 3 and the second processing device 4 are processing devices of the same type. The same type here refers to processing devices that have the same functions and can perform the same processing on a workpiece as a processing object. The content of the processing performed by the first processing device 3 and the second processing device 4 is in accordance with the setting of the processing parameters described later. In addition, the details will be described later. Figure 1In the machining system 100 shown, when the learning device 1 performs learning, the 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, the machining parameters are set individually for the first machining device 3 and the second machining device 4.

[0024] The learning device 1 includes a recording unit 11, a reference information acquisition unit 12, and a correction function learning unit 13. The first processing device 3 includes a tool 31, a workpiece 32, and an observation unit 33. The second processing device 4 includes a tool 41, a workpiece 42, an observation unit 43, a correction function 44, and a model 45. In addition, the first processing device 3 and the second processing device 4 each have a numerical control device (not shown) and a drive unit for changing the relative position of the tool and the workpiece. Figure 1 Although the model 45 is included in the second processing device 4, it may be configured so that another device different from the second processing device 4 includes the model 45. For example, a device having a function of predicting the processing result of the workpiece 42 obtained by the second processing device 4 may include the model 45, and the device may use the model 45 to predict the processing result.

[0025] Figure 2 The diagram shows a method of creating the model 45 used by the machining system 100 according to the first embodiment. The model 45 is based on Figure 1 The sensor information and labels collected by the first processing device 3 shown are created by the model creation device 50. The model creation device 50 has: a database 5, which registers the sensor information and labels; and a model learning unit 6, which learns the relationship between the sensor information and labels registered in the database 5 to create a model 45. 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.

[0026] The observation unit 33 of the first processing device 3 and the observation unit 43 of the second processing device 4 are Figure 3 and Figure 4 The structure shown. Figure 3 2 is a diagram showing a configuration example of the observation unit 33 of the first processing device 3. Figure 4 4 is a diagram showing a configuration example of the observation unit 43 of the second processing device 4. Figure 3As shown, the observation unit 33 of the first processing device 3 has a sensor 331, and the state of the first processing device 3 is observed using the sensor 331. The sensor 331 outputs a sensor signal indicating the observation result. The observation unit 33 can observe the state of the workpiece 32 and use the observation result as the observation result of the first processing device 3. That is, the sensor signal output by the sensor 331 can indicate the observation result of the workpiece 32. The observation unit 43 of the second processing device 4 has a sensor 431, and the state of the second processing device 4 is observed using the sensor 431. The sensor 431 outputs a sensor signal indicating the observation result. The observation unit 43 can observe the state of the workpiece 42 and use the observation result as the observation result of the second processing device 4. That is, the sensor signal output by the sensor 431 can indicate the observation result of the workpiece 42. The sensors 331 and 431 are the same type of sensors, but there are individual differences. When the same object is observed under the same conditions, there is a possibility that the observation result obtained by the sensor 331 and the observation result obtained by the sensor 431 are different.

[0027] Here, specific examples of processing devices used as the first processing device 3 and the second processing device 4 are turning devices, milling devices, grinding devices, laser processing devices, electrical discharge processing devices, water jet processing devices, and punching devices.

[0028] The status of the processing device is obtained by expressing the status of the processing device through data or observation, such as whether the driving part is normal or abnormal, the current flowing in the driving part, the voltage borne by the driving part, the sound, light emitted during processing, the temperature of a specific part, position, speed, acceleration, angle, angular velocity, angular acceleration, pressure, deformation, degree of tool consumption, image of a specific part, etc.

[0029] The state of the workpiece 32 is a value obtained by observing the condition of the workpiece 32 to be processed, and includes, for example, the acceleration that can be obtained by an acceleration sensor installed on the workpiece 32, the angular velocity that can be obtained by a gyro sensor, the observation value of a light sensor if light is emitted during processing, the temperature of the workpiece, etc. The state of the workpiece 42 is the same.

[0030] The items and values ​​observed by the observation unit 33 and the observation unit 43 are usually correlated with the items predicted by the output of the model 45. Therefore, the types of sensors used by the observation unit 33 and the observation unit 43 need to be designed according to the items to be predicted by the model 45 and the type of processing equipment to which the model 45 is applied. In addition, a plurality of sensors used by the observation unit 33 and the observation unit 43 may be combined according to the purpose.

[0031] Machining parameters are a group of variables associated with machining actions set in a machining device. The values ​​set in the machining parameters include machining speed indicating the speed of the machining action, a limit value for the acceleration of a movable part, a value for specifying the relative position relationship between a tool and a workpiece, and a specified value for the amount of cutting fluid and oil used to assist machining.

[0032] In one example, if the processing device is a turning processing device, the processing parameters include processing speed, feed amount, rotation speed of the workpiece, amount of cutting fluid, etc., which are set as a group of variables. In another example, if the processing device is a laser processing device, the processing parameters include laser output, focus position, beam shape, processing speed, processing gas type, processing gas pressure, nozzle height, etc., which are set as a group of variables.

[0033] A part or all of the learning device 1 may be provided in the first processing device 3, or may be provided outside the first processing device 3. A part or all of the learning device 1 may be provided in the numerical control device constituting the first processing device 3. A part or all of the learning device 1 may be provided in the second processing device 4. In addition, it is also possible to configure that the first processing device 3 is cloud-connected via a network, and all or a part of the learning device 1 is implemented by the processing circuit of the cloud server.

[0034] The first processing device 3 as the first processing device processes the workpiece 32 using the tool 31 according to the processing parameters and the processing program. The observation unit 33 acquires the state of at least one of the workpiece 32 and the first processing device 3 as sensor information.

[0035] like Figure 2 As shown, the model 45 is created by the model learning unit 6 of the model creation device 50 based on the sensor information and labels collected by the first processing device 3. Specifically, the workpiece 32 is processed by the first processing device 3, and the sensor information obtained from the observation unit 33 and the labels corresponding to the sensor information are recorded in the database 5 of the model creation device 50 and accumulated. The model learning unit 6 outputs the model 45 obtained by learning 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, feature quantities can be extracted from the sensor information, and the relationship between the feature quantities of the sensor information and the labels can be learned. When the relationship between the feature quantities of the sensor information and the labels is learned, feature quantities are extracted for the sensor information given when the model 45 is used or the corrected sensor information obtained by correcting the sensor information, and the extracted feature quantities are input to the model 45 to obtain an output.

[0036] The model 45 may be a general machine learning model, an example of which is a neural network. Other examples include a branching tree, a support vector machine, a Gaussian process regression, etc. A general regression model or classification model can be used according to the purpose of the model 45.

[0037] In addition, an example of the output of the model 45 is a value obtained by digitizing the pass / fail of processing performed by the processing device that is the source of the sensor information input to the model 45. Other examples are the degree of abnormality of the processing device that is the source of the sensor information input to the model 45, abnormality determination information of the processing device, the degree of wear of the tool, the value of the processing parameter, etc.

[0038] The tags recorded in the database 5 need to record information corresponding to the output of the model 45 , and the sensor information acquired by the observation unit 33 needs to be information related to the tags.

[0039] In addition, the observation unit 33 usually cuts the sensor signal obtained from the sensor 331 at regular intervals and outputs it as sensor information. The observation unit 33 can also further determine whether it is in a processing operation, and output sensor information only during the processing operation, or output sensor information when the start or end of a specific operation is detected. The observation unit 43 of the second processing device 4 using the model 45 is the same.

[0040] As described above, the model 45 learns the relationship between the sensor information and the label. A specific example of the sensor information, the label, and the model 45 will be described.

[0041] As an example, a case where the first processing device 3 and the second processing device 4 are turning processing devices will be described.

[0042] In a turning device, a workpiece is mounted on a rotating spindle, and a tool is pressed against the workpiece to perform cutting. At this time, a vibration called "chatter" may occur depending on the material, rotation speed, and feed amount of the workpiece. In order to detect the "chatter", it is preferred to install an acceleration sensor on the tool to collect sensor information. When collecting sensor information, various sensor information during normal processing is obtained by making various changes to the processing parameters when "chatter" occurs. Moreover, the sensor information when "chatter" occurs is accompanied by a label of "chatter", and the sensor information during normal processing is accompanied by a label of "normal". At this time, the obtained sensor information is time series data, and the label is associated with the time series data, i.e., the sensor information, in accordance with the time when "chatter" occurs. Therefore, by cutting the time series data with a certain time width from the sensor information, a data pair in which the cut time series data and the label are associated is obtained. The model 45 can learn the relationship between the time series data of the acceleration sensor mounted on the tool and "chatter" by learning the relationship between the data pair, i.e., the cut time series data and the label. After the learning of the model 45 is completed, in the turning processing device, the time series data of the acceleration sensor installed on the tool, that is, the sensor information, is input to the model 45, thereby obtaining an output of whether "chatter" occurs, and "chatter" can be automatically detected. In addition, when "chatter" occurs, the processing can be stopped or the processing parameters can be changed.

[0043] Furthermore, as another example, a case where the first processing device 3 and the second processing device 4 are laser cutting processing devices which are a type of laser processing devices will be described.

[0044] In a laser cutting processing device, a laser focused near the surface of the workpiece is irradiated to the workpiece, thereby cutting the workpiece. Therefore, in a laser processing device such as a laser cutting processing device, the laser is equivalent to a tool. If the laser is irradiated, the workpiece will heat up and emit light, so in order to obtain the processing state, the time series data of the optical sensor that captures the light emission of the workpiece can be set as sensor information. As the processing results during laser cutting processing, there are normal processing, slag attached to the back of the workpiece by the molten material, damage with unevenness on the cut section, and burning in which the molten material blows up without penetrating the cut section. It is preferred that when collecting sensor information, various changes are made to the processing parameters to obtain sensor information corresponding to various processing results. Moreover, the type of processing result is attached as a label. In the case of attaching a label to the time series data of the optical sensor, according to the time series data of the processing position and the processing result for the processing position, after the processing result and the time are associated, the time of the optical sensor and the time of the processing result are associated, thereby the processing result, i.e., the label, and the sensor information, i.e., the time series data of the optical sensor, are associated. By cutting the time series data with a certain time width from the sensor information, a data pair associated with the cut time series data and a label is obtained. The model 45 can learn the relationship between the time series data of the optical sensor and the processing result by learning the relationship between the data pair, i.e., the cut time series data and the label. After the learning of the model 45 is completed, in the laser cutting processing device, the time series data of the optical sensor that captures the light emission of the workpiece, i.e., the sensor information, is input to the model 45, thereby obtaining a prediction of the processing result as an output during the processing. In addition, when the predicted processing result is not good, the processing can be stopped or the processing parameters can be changed.

[0045] Furthermore, as another example, a case where the first processing device 3 and the second processing device 4 are a type of electric discharge processing device, that is, a die-cast electric discharge processing device will be described.

[0046] In a die-casting electrical discharge machining device, a discharge electrode is brought close to a workpiece, and the workpiece is machined using a discharge phenomenon that occurs when a voltage is applied between the workpiece and the discharge electrode. Therefore, in the electrical discharge machining device, the discharge electrode and the generated discharge phenomenon are equivalent to a tool.

[0047] In a die-casting electrical discharge machining device, there is a driving unit for changing the relative distance between a discharge electrode and a workpiece. An example of the driving unit is a linear motor, in which case the position of the discharge electrode is acquired by a linear encoder. In this example, the current flowing in the linear motor and the position of the discharge electrode are used as sensor information.

[0048] As an abnormality of the die-casting EDM device, for example, a situation where the position of the discharge electrode cannot be smoothly adjusted due to the exhaustion of grease in the sliding part of the linear motor, abnormal heating of the motor, etc. is considered. In order to detect the abnormality as described above, sensor information is obtained in a state where the abnormality is reproduced, and an abnormal label is attached. In addition, sensor information during normal operation is obtained, and a normal label is attached. Moreover, the model 45 is made to learn the correspondence between the sensor information and the label. After the learning of the model 45 is completed, in the die-casting EDM device, by inputting the current of the linear motor and the position of the discharge electrode, that is, the sensor information, into the model 45, it is possible to detect whether the linear motor is normal or abnormal.

[0049] As described above, the observation unit and labels are set to match the characteristics of the processing device, sensor information and labels representing the observation results are collected, and the model 45 learns the relationship between the sensor information and the labels, thereby enabling the model 45 to be effectively used in various ways.

[0050] The second processing device 4, which is the second processing device, is a processing device that effectively utilizes the model 45 created using the first processing device 3. The second processing device 4 is a different entity from the first processing device 3 and is the same processing device, and processes the workpiece 42 using the tool 41 according to the processing parameters and the processing program. The observation unit 43 of the second processing device 4 observes the same sensor information as the observation unit 33 of the first processing device 3.

[0051] The second processing device 4 also includes a correction function 44 created by the learning device 1 and a model 45 created using the first processing device 3. In the second processing device 4, when the correction function 44 is created by the learning device 1, the correction function 44 is applied to the sensor information obtained by the observation unit 43 to obtain the corrected sensor information. The corrected sensor information is input to the model 45 to obtain the output of the model 45. In addition, when the correction function 44 is not assigned or when the user chooses not to use the correction function 44, the sensor information obtained by the observation unit 43 is directly input to the model 45 to obtain the output of the model 45. When the user can choose whether to use the correction function 44, the second processing device 4 also includes an input unit for receiving the selection made by the user.

[0052] The user can choose to use the correction function 44 or not to use the correction function 44, so that, for example, when the correction performed by the correction function 44 does not meet the effective use purpose of the model 45, it is possible to immediately return to the original state without using the correction function 44. At this time, in order to enable the user to judge whether the correction of the sensor information performed by the correction function 44 is appropriate, the second processing device 4 preferably has a display unit, which compares and displays the output obtained by inputting the corrected sensor information obtained by applying the correction function 44 to the sensor information into the model 45 and the output obtained by inputting the sensor information into the model 45.

[0053] Alternatively, a plurality of correction function 44 candidates may be created, and the user may select a candidate suitable for the output of the model 45 from the candidates, and the selected candidate may be used as the correction function 44. Thus, the user can select a correction function 44 more suitable for the purpose of effectively using the model 45.

[0054] Furthermore, by providing a storage unit in the second processing device 4 to store the sensor information obtained by the observation unit 43, it is possible to use the stored sensor information to confirm whether the correction performed by the correction function 44 is appropriate. By having a storage unit for storing the sensor information, the processing for confirming whether the correction performed by the correction function 44 is appropriate can be omitted, and the consumption of workpieces and tools can be reduced. Furthermore, if a plurality of sensor information is stored in the storage unit, the output of the model 45 can be confirmed in a plurality of processing states, and the correction function 44 can be selected more appropriately.

[0055] exist Figure 1 In the processing system 100 shown, the second processing device 4 is configured to have the model 45, and the first processing device 3 is not configured to have the model 45. However, the first processing device 3 may also have the model 45, and the model 45 may be effectively used to determine whether the processing performed by the first processing device 3 is acceptable or not. In addition, the model 45 learns the correspondence between the sensor information output by the observation unit 33 of the first processing device 3 and the label, so that the first processing device 3 does not have the same correction function as the correction function 44 of the second processing device 4, or the correction function of the first processing device 3 does not correct the input sensor information and directly outputs it.

[0056] The learning device 1 learns the correction function 44 used by the second processing device 4 . Figure 5 is a flowchart showing an example of the operation of the learning device 1. The learning device 1 repeatedly executes Figure 5 The correction function 44 is learned through steps S1 to S3 shown in FIG.

[0057] In learning the correction function 44, the learning device 1 first obtains the observation result of the operation of the first processing device 3 (step S1). In detail, the learning device 1 obtains the sensor information output from the observation unit 33 when the first processing device 3 is operated using the reference processing parameter 2 as the processing parameter, and records the obtained sensor information in the recording unit 11 as the first reference information.

[0058] The learning device 1 then obtains the observation result of the operation of the second processing device 4 set with the same processing parameters as the first processing device 3 (step S2). In detail, in the learning device 1, the reference information acquisition unit 12 obtains the sensor information output from the observation unit 43 when the second processing device 4 is operated using the reference processing parameter 2 as the processing parameter as the second reference information.

[0059] Next, the learning device 1 corrects the observation result of the action of the second processing device 4 and learns a correction function used for processing close to the observation result of the action of the first processing device 3 (step S3). In detail, in the learning device 1, the correction function learning unit 13 causes the correction function 44 to act on the second reference information, and when the corrected second reference information is obtained, the correction function 44 is learned in a manner that makes the characteristics of the corrected second reference information close to the characteristics of the first reference information.

[0060] Here, the sensor information recorded as the first reference information in the recording unit 11 may be sensor information obtained by the model creation device 50 to create the model 45 when the processing parameters used in the processing are the same as the reference processing parameters 2. Alternatively, the processing parameters may be selected from the processing parameters corresponding to the sensor information obtained to create the model 45 and used as the reference processing parameters 2. By using the sensor information obtained to create the model 45, it is possible to omit the need to perform the processing for which the reference processing parameters 2 are set again in the first processing device 3.

[0061] In addition, when the nature of the sensor information changes due to the difference in the processing program and affects the output of the model 45, it is preferable to prepare a processing program that serves as a reference, that is, a reference processing program, similarly to the reference processing parameter 2, and obtain the first reference information and the second reference information using the reference processing parameter 2 and the reference processing program. By using the reference processing program, the difference between the first reference information and the second reference information caused by the difference between the processing program used by the first processing device 3 when obtaining the first reference information and the processing program used by the second processing device 4 when obtaining the second reference information can be eliminated. As a result, compared with the case where different processing programs are used, a more accurate correction function 44 can be learned, and the error 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] The operation of the correction function learning unit 13 of the learning device 1 according to Embodiment 1 will be described in more detail. In Embodiment 1, the first reference information and the second reference information are time series data of sensor signals.

[0063] There are various possible methods for the correction function learning unit 13 to learn the correction function 44 so that the characteristics of the second reference information after correction are close to the characteristics of the first reference information.

[0064] In one example, the correction function 44 can be a low-pass filter for time series data. Other examples of the correction function 44 include a method of using a high-pass filter or a band-pass filter, and a method of using a function that adds a constant multiple of the deviation, and these can also be combined.

[0065] In one example, the correction function 44 is learned by adding a deviation to the second reference information so that the average value of the second reference information after correction coincides with the average value of the first reference information. In other examples, there is a method of learning the correction function 44 that multiplies the second reference information by a factor so that the average value of the second reference information after correction coincides with the average value of the first reference information, a method of learning the parameters of the correction function so that the feature quantity extracted from the second reference information after correction is close to the feature quantity extracted from the first reference information, and the like.

[0066] Here, the feature quantity is a quantity that represents the characteristics of the sensor signal, for example, a statistical quantity such as the average value, variance value, standard deviation value, central value, maximum value, minimum value, crest factor, and number of peaks of the sensor signal. In addition, the result of frequency analysis, the result of filter bank analysis (the result obtained by summing the result of frequency analysis to a certain frequency range), the result of cepstrum analysis, etc. can also be used as a feature quantity. When extracting these feature quantities, a general feature quantity extraction method can be used. In the case of using multiple values ​​as feature quantities, for example, the parameters of the correction function 44 are learned in a manner that the sum of the absolute values ​​of the differences between each feature quantity extracted from the corrected second reference information and each feature quantity extracted from the first reference information becomes smaller, and the parameters of the correction function 44 are learned in a manner that the weighted sum of the differences between each feature quantity extracted from the corrected second reference information and each feature quantity extracted from the first reference information becomes smaller.

[0067] When learning the parameters of the correction function 44, a general machine learning method can be used. That is, the correction function learning unit 13 determines the parameters of the correction function 44 by a machine learning method. For example, the correction function learning unit 13 can also learn the parameters of the correction function by Newton's method in a manner in which the sum of the squares of the differences between the feature quantity extracted from the corrected second reference information and the feature quantity extracted from the first reference information becomes smaller, and can use general optimization methods such as the conjugate gradient method, the Bayesian optimization method, the probabilistic gradient descent method, the particle swarm optimization (PSO: Particle Swarm Optimization), and the random exploration method. The optimization method selects an appropriate method according to the shape of the correction function 44 and the number of parameters, the feature quantity of the first reference information and the number of feature quantities of the corrected second reference information, and their properties.

[0068] When the labels in the first reference information and the second reference information are different, it is impossible to determine whether the difference between the features of the first reference information and the features of the second reference information is caused by the difference in labels or caused by individual differences in the tools, workpieces, sensors, etc. in the first processing device 3 and the second processing device 4. In the case where the difference is caused by the difference in labels, it is sometimes impossible to correct the individual differences in the tools, workpieces, sensors, etc. in the first processing device 3 and the second processing device 4 by the correction function 44.

[0069] Therefore, the first reference information and the label are recorded in association in the recording unit 11, and preferably, when the label corresponding to the second reference information acquired by the reference information acquisition unit 12 is consistent with the label corresponding to the recorded first reference information, the correction function 44 is learned. Thus, it can be determined that the difference between the feature of the first reference information and the feature of the second reference information is caused by the individual difference of the tool, workpiece, sensor, etc. in the first processing device 3 and the second processing device 4. As described above, the correction function learning unit 13 can create a correction function 44 that can more accurately correct the influence of the individual difference of the tool, workpiece, sensor, etc. in the first processing device 3 and the second processing device 4, that is, the difference between the sensor information obtained by the first processing device 3 and the sensor information obtained by the second processing device 4, compared with the case where the consistency of the label is not confirmed.

[0070] Consider that there are multiple first reference information and multiple second reference information. Hereinafter, multiple first reference information is set as the first group, and multiple second reference information is set as the second group. It is also possible to cut the given first reference information and second reference information at regular intervals when the sensor information is cut by the observation unit 33 and the observation unit 43, and set them as the first group and the second group respectively. In addition, when multiple processing is performed by the reference processing parameter 2 and multiple sets of processing parameters are specified in the reference processing parameter 2, there are multiple first reference information and multiple second reference information, which are the first group and the second group respectively.

[0071] In addition, when multiple sets of processing parameters are specified in the reference processing parameter 2, the first processing device and the second processing device perform processing realized by each processing parameter set in the reference processing parameter 2 to obtain each first reference information and each second reference information, namely the first group and the second group.

[0072] The correction function learning unit 13 uses the first group and the second group to learn the parameters of the correction function in such a way that the distribution of the feature quantity extracted from the corrected second reference information is close to the distribution of the feature quantity extracted from the first reference information. That is, the correction function learning unit 13 determines the parameters of the correction function 44 in such a way that the distance between the distribution of the feature quantity extracted from the corrected second reference information and the distribution of the feature quantity extracted from the first reference information becomes small.

[0073] By using a plurality of first reference information and a plurality of second reference information, the characteristics of the corrected second reference information and the characteristics of the first reference information can be made closer to each other more accurately than when only one first reference information and one second reference information are used. Therefore, the correction function learning unit 13 can learn an accurate correction function 44, and can further reduce the error of the output of the model 45 when the correction function 44 is used in the second processing device 4.

[0074] In addition, by specifying multiple sets of processing parameters as the reference processing parameters 2, the correction function is learned based on the processing performed in multiple different states. Therefore, compared with the case where one processing parameter is specified as the reference processing parameter 2, the correction function learning unit 13 can learn the correction function 44 with high versatility, and can reduce the error of the output of the model 45 when the correction function 44 is used in the second processing device 4.

[0075] Here, the distance between the above-mentioned distributions can utilize general distances, such as KL (Kullback-Leibler) divergence, JS (Jensen-Shannon) divergence, histogram intersection method, L1 distance, L2 distance, Pearson distance, relative Pearson distance, etc., and when calculating the distance, the general approximation method of the distance between distributions can be used.

[0076] In addition, the distribution of the feature quantity 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 most likely method for a parameterized model, a Bayesian estimation method for a parameterized model, and a method using a mixed distribution can be used. The estimation method of the probability density function and the calculation method of the distance between the distributions are selected to match, and the selected methods are used by the correction function learning unit 13.

[0077] In the case where there are multiple first reference information and multiple second reference information, it is also effective to confirm the consistency of the labels in the same manner as described above. The correction function learning unit 13 extracts the first reference information and the second reference information with the same processing parameters and associated with the same labels from the first group and the second group as the first reference group and the second reference group, respectively. Moreover, the correction function learning unit 13 uses the distribution of the feature quantity of the first reference information included in the first reference group as the feature of the first reference information group, and uses the distribution of the feature quantity of the second reference information included in the second reference group as the feature of the second reference information group, so as to learn the correction function 44 in such a way that the feature of the corrected second reference information group is close to the feature of the first reference information group.

[0078] Thus, the difference between the feature of the first reference information and the feature of the second reference information can be judged to be caused by the individual difference of the tool, workpiece, sensor, etc. in the first processing device 3 and the second processing device 4. As described above, the correction function learning unit 13 can create a correction function 44 that can more accurately correct the influence of the individual difference of the tool, workpiece, sensor, etc. in the first processing device 3 and the second processing device 4, that is, the difference between the sensor information obtained by the first processing device 3 and the sensor information obtained by the second processing device 4, compared with the case where the consistency of the labels is not confirmed.

[0079] At this time, for each processing parameter and label set as the reference processing parameter 2, if there is an imbalance in the number of data between the first reference group and the second reference group, when the distribution (probability density function) is obtained, the distribution (probability density function) becomes unbalanced due to the imbalance in the number of data, which sometimes affects the learning of the correction function 44. In other words, the correction function obtained by learning sometimes cannot accurately correct the individual differences of the tools, workpieces, sensors, etc. in the first processing device 3 and the second processing device 4.

[0080] Therefore, the correction function learning unit 13 extracts the first reference information and the second reference information associated with the same processing parameter and the same label from the first group and the second group in the same proportion, respectively, and sets them as the first reference group and the second reference group. For example, when the reference processing parameters include processing parameters A and B, and the label assigned to the sensor information is any of L1 and L2, the correction function learning unit 13 extracts the first reference information of processing parameter A and label L1, the first reference information of processing parameter A and label L2, the first reference information of processing parameter B and label L1, and the first reference information of processing parameter B and label L2 in a proportion of 1:2:1:2, and also extracts the second reference information of processing parameter A and label L1, the second reference information of processing parameter A and label L2, the second reference information of processing parameter B and label L1, and the second reference information of processing parameter B and label L2 in a proportion of 1:2:1:2. At this time, the first reference information and the second reference information associated with the same processing parameter and the same label can also be extracted in the same number and set as the first reference group and the second reference group respectively. In addition, the correction function learning unit 13 uses the distribution of the feature quantity of the first reference information included in the first reference group as the feature of the first reference information group, and uses the distribution of the feature quantity of the second reference information included in the second reference group as the feature of the second reference information group, so as to make the feature of the corrected second reference information group close to the feature of the first reference information group, and learns the correction function 44.

[0081] Thus, the first reference group and the second reference group can eliminate the distribution imbalance caused by the imbalance of the number of data in each processing parameter and label group. Therefore, compared with the case where the first reference information and the second reference information associated with the same processing parameter and the same label are extracted as the first reference group and the second reference group, respectively, compared with the case where the consistency of the labels is not confirmed, the individual differences of the tools, workpieces, sensors, etc. in the first processing device 3 and the second processing device 4 can be corrected more accurately.

[0082] use Figure 6 A specific example of the operation of the correction function learning unit 13 extracting the same number of first reference information and second reference information associated with the same label for the same processing parameter from the first group and the second group, respectively, and setting them as the first reference information group and the second reference information group is described. Figure 6 This is a diagram for explaining an example of the operation of the correction function learning unit 13 of the learning device 1 selecting data used for learning the correction function 44 .

[0083] exist Figure 6 In the example shown, the six first reference information #1 to #6 of the first group are associated with process parameters A to C and labels 1 to 2. The four second reference information #1 to #4 of the second group are associated with process parameters A to B and labels 1 to 2.

[0084] In this case, if the first group and the second group are compared, the first reference information and the second reference information associated with the same processing parameters and the same label in the first group and the second group are the case where the processing parameter is A and the label is 1 and the case where the processing parameter is B and the label is 2.

[0085] There are two cases in which the processing parameter is A and the label is 1 in the first group and the second group, respectively, so 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 the first reference information #1 and the first reference information #3 from the first group and adds them to the first reference information group, and extracts the second reference information #1 and the second reference information #2 from the second group and adds them to the second reference information group.

[0086] In addition, there is one case where the processing parameter is B and the label is 2 in the first group and two cases in the second group. Therefore, when the same number is extracted, the correction function learning unit 13 appropriately selects one from the second group. In addition, the selection method may be random or a preferred method may be selected. Specifically, the correction function learning unit 13 extracts the first reference information #5 from the first group and adds it to the first reference information group, and extracts the second reference information #3 from the second group and adds it to the second reference information.

[0087] In addition, when data for learning the model 45 is acquired by a plurality of first processing devices 3, that is, when there are a plurality of first processing devices 3, it is also possible to extract feature quantities from each first reference information for each first processing device 3, and process the average value and median value of the feature quantities of each first reference information as the feature quantities extracted from the first reference information. In one example, the correction function learning unit 13 uses the feature quantity as the average value, and learns the parameters of the correction function 44 in such a way that the average value extracted from the corrected second reference information is close to the average value of each first reference information.

[0088] When learning the correction function 44 in the present embodiment, it is not necessary to store the entire database 5 used when creating the model 45. It is sufficient to record only the first reference information in the recording unit 11 of the learning device 1. Therefore, it can be implemented even when the storage capacity of the recording unit 11 is small.

[0089] As other methods for correcting individual differences in processing devices, there are conceivable methods of correcting sensor information for learning in accordance with the processing device to which the model 45 is applied and re-learning the model 45, or methods of acquiring sensor information for each processing device and learning the model 45. However, in these methods, a lot of data is recorded for learning the model 45, and since learning processing using a lot of data is performed, a large storage capacity is required, the calculation load is also large, and the processing takes time.

[0090] On the other hand, the learning of the correction function 44 in the present embodiment can be realized with a small storage capacity because it is performed using the first reference information and the second reference information obtained in the processing based on the reference processing parameter 2. In addition, since the number of data during learning is small, the calculation load is also reduced, and processing can be performed at a high speed.

[0091] In addition, it is also considered that the first reference information and the corrected second reference information are input to the model 45, and the correction function 44 is learned in a manner that the output is consistent. However, a nonlinear model such as a neural network or a branching tree is often used in 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, the second reference information needs to be input to the candidate of the correction function 44, and the output of multiple models 45 needs to be obtained for the obtained output, which makes it difficult to learn the correction function 44.

[0092] On the other hand, since the learning of the correction function 44 in the present embodiment is performed using the first reference information and the second reference information obtained in the processing based on the reference processing parameters 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 the case where the first reference information and the corrected second reference information are input to the model 45 and the correction function 44 is learned in such a manner that the outputs are consistent, the amount of calculation is reduced, and the learning of the correction function 44 can be performed more simply.

[0093] Next, the hardware configuration of the learning device 1 will be described. Figure 7 is a diagram showing an example of hardware for realizing the learning device 1. The learning device 1 can Figure 7 The processor 91, memory 92 and interface circuit 93 shown are implemented.

[0094] The processor 91 is a CPU (also called a Central Processing Unit, a central processing device, a processing device, an arithmetic device, a microprocessor, a microcomputer, a DSP (Digital Signal Processor)), a system LSI (Large Scale Integration), etc. The memory 92 is a RAM (Random Access Memory), a ROM (Read Only Memory), an EPROM (Erasable Programmable Read Only Memory), an EEPROM (registered trademark) (Electrically Erasable Programmable Read Only Memory), a hard disk drive, etc. The interface circuit 93 is a circuit used by the learning device 1 to transfer data with external devices such as the first processing device 3 and the second processing device 4. The interface circuit 93 may include a circuit that is connected to a network and transmits and receives data with other devices via the network.

[0095] The reference information acquisition unit 12 and the correction function learning unit 13 of the learning device 1 are implemented by executing a program for operating as these units by the processor 91. The program for operating as the reference information acquisition unit 12 and the correction function learning unit 13 is stored in advance in the memory 92. The processor 91 reads and executes the above program from the memory 92, thereby operating as the reference information acquisition unit 12 and the correction function learning unit 13. The recording unit 11 is implemented by the memory 92.

[0096] In addition, it is assumed that the program for operating as the reference information acquisition unit 12 and the correction function learning unit 13 is stored in the memory 92 in advance, but it is not limited to this. The above-mentioned program can also be provided to the user of the learning device 1 in the state of being written to a recording medium such as a CD (Compact Disc)-ROM, DVD (Digital Versatile Disc)-ROM, and the user installs the above-mentioned program in the memory 92. In this case, the hardware that realizes the learning device 1 also includes a reading device for reading the program from the recording medium. It can also be a method of connecting the reading device to the interface circuit 93 to install the program. In addition, it can also be a method in which the above-mentioned program is provided from a server via a network.

[0097] As described above, the learning device 1 according to the present embodiment uses the first reference information obtained by observing the processing action of the first processing device 3 set with the reference processing parameter 2 and the second reference information obtained by observing the processing action of the second processing device 4 set with the reference processing parameter 2, to learn the relationship between the first reference information and the second reference information, and generates a correction function 44 for correcting the sensor information obtained by the second processing device 4 so that the sensor information obtained when the second processing device 4 performs the processing action with the same processing parameters as the first processing device 3 is close to the sensor information obtained by the first processing device 3. As a result, the learning model for judging the action result of the first processing device 3 based on the sensor information can also be used for judging the action result of the second processing device 4. That is, it is possible to prevent the judgment accuracy of the learning model from being reduced when the learning model learned by observing the action of the first processing device 3 is applied to the second processing device 4.

[0098] Implementation method 2.

[0099] In Embodiment 2, the differences from Embodiment 1 will be mainly described. In the drawings used in the description of this embodiment, the same reference numerals are given, and components not specifically described are the same as those described in Embodiment 1.

[0100] In the second embodiment, 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 described in the first embodiment are different. Specifically, the observation unit 33 is set to Figure 8 The observation unit 33a shown in FIG. 4 is set as the observation unit 43 Fig. 9 The observation portion 43a is shown.

[0101] In the following description, in order to distinguish from Embodiment 1, the processing system, first processing device, second processing device and learning device involved in Embodiment 2 are respectively referred to as processing system 100a, first processing device 3a, second processing device 4a and learning device 1a.

[0102] Figure 8 This is a diagram showing a configuration example of an observation unit 33a included in the first processing device 3a of the processing system 100a according to the second embodiment. Fig. 9 This is a diagram showing a configuration example of an observation unit 43a included in the second processing device 4a of the processing system 100a according to the second embodiment.

[0103] The observation unit 33a of the first processing device 3a includes a sensor 331 and a feature extraction unit 332, and the observation unit 43a of the second processing device 4a includes a sensor 431 and a feature extraction unit 332. In addition, the feature extraction unit 332 is given the same reference numeral to indicate the content of processing by both the observation unit 33a and the observation unit 43a to extract the feature. The feature extraction unit 332 can also be provided outside the observation units 33a and 43a within the range in which the order of processing the sensor signals output from the sensors 331 and 431 is not changed.

[0104] The feature quantity extraction unit 332 of the observation unit 33a receives the sensor signal output from the sensor 331, extracts the feature quantity from the sensor signal, and outputs the feature quantity of the sensor signal as sensor information. Similarly, the feature quantity extraction unit 332 of the observation unit 43a receives the sensor signal output from the sensor 431, extracts the feature quantity from the sensor signal, and outputs the feature quantity of the sensor signal as sensor information.

[0105] In addition, the observation units 33a and 43a usually cut the sensor signals obtained from the sensors 331 and 431 at regular intervals, and output the feature quantities of the cut sensor signals as sensor information. The observation units 33a and 43a also judge whether they are in a processing action, and can also output sensor information only during a processing action, or cut the sensor signals when the start or end of a specific action is detected to extract the feature quantities and output them as sensor information.

[0106] In the second embodiment, the sensor information is the characteristic quantity of the sensor signal. Therefore, as described in the first embodiment, the model 45 obtained by learning the relationship between the sensor information and the label is a model obtained by learning the relationship between the characteristic quantity of the sensor signal and the label. In addition, similarly, the correction function 44 is a function for correcting the characteristic quantity of the sensor signal. Therefore, the correction function 44 is, with respect to the characteristic quantity of the sensor signal, for example, the addition of deviations of each characteristic quantity, constant multiplication, parallel translation of the characteristic quantity space, rotation of the characteristic quantity space, and the combination of these processes. The correction function 44 is preferably a function that is a bijective function with respect to the characteristic quantity space.

[0107] The method of creating the correction function 44 can be interpreted as the “feature of the first reference information” described in the first embodiment being the “first reference information” and the “feature of the second reference information” being the “second reference information”, and thus can be the same as the description in the first embodiment.

[0108] In the learning device 1a involved in embodiment 2, the recording unit 11 records the characteristic quantity of the sensor signal measured by the benchmark processing parameter 2, that is, the first benchmark information, in the first processing device 3a, the benchmark information acquisition unit 12 acquires the characteristic quantity of the sensor signal when the processing is performed by the benchmark processing parameter 2, that is, the second benchmark information, in the second processing device 4a, and the correction function learning unit 13 learns the correction function 44 in such a way that the corrected second benchmark information is close to the first benchmark information.

[0109] As described above, in the processing system 100a involved in the present embodiment, the observation unit 33a of the first processing device 3a and the observation unit 43a of the second processing device 4a have a feature extraction unit 332 that extracts the feature of the sensor signal. According to the learning device 1a involved in the second embodiment, the feature of the sensor signal, that is, the sensor information, can be recorded in the recording unit 11. Generally speaking, the data capacity of the feature of the sensor signal is smaller than the data capacity of the sensor signal, so it can be implemented with a small storage capacity compared to the case of saving the sensor signal. In addition, for the same reason, in the learning of the correction function 44, the amount of calculation required for learning is also reduced compared to the case of learning the correction for the sensor signal, and it can be processed at a high speed.

[0110] Furthermore, when learning the correction function for correcting the sensor signal as in Embodiment 1, a provisional correction function is applied to the sensor signal, and learning is performed by comparing the features after extracting the feature quantity, but when the feature quantity is directly corrected as in Embodiment 2, it is not necessary to extract the feature quantity for comparison. Therefore, Embodiment 2 can learn the correction function 44 more simply than when the sensor signal is corrected as in Embodiment 1.

[0111] Implementation method 3.

[0112] In the third embodiment, the differences from the first and second embodiments will be mainly described. In the drawings used in the description of this embodiment, the same reference numerals are given, and components not specifically described are the same as those described in the first embodiment.

[0113] Fig.10 1 is a diagram showing a configuration example of a processing system 100b according to Embodiment 3. The processing system 100b includes a learning device 1b, a first processing device 3, N (N is an integer greater than or equal to 1) second processing devices 4-1 to 4-N, a database 5, and a generalized model 8. The second processing devices 4-1 to 4-N are the same processing devices as the second processing device 4 described in Embodiment 1 (see Figure 1 ).exist Fig.10 In the description, description of the components other than the observation unit 43 of the second processing devices 4 - 1 to 4 -N is omitted. In the following description, the second processing devices 4 - 1 to 4 -N may be collectively referred to as the second processing device 4 .

[0114] The learning device 1b includes a recording unit 11, a reference information acquisition unit 12, a correction function learning unit 13, an inverse correction function creation unit 14, an individual simulation data creation unit 15, and a generalized model learning unit 16. The learning device 11b is a structure in which the inverse correction function creation unit 14, the individual simulation data creation unit 15, and the generalized model learning unit 16 are added to the learning device 1 involved in the first embodiment. In addition, the learning device 1b uses the database 5 described in the first embodiment.

[0115] Here, it is considered that there are individual differences in tools, workpieces, sensors, etc. in each second processing device 4. Even if there are these individual differences, the learning device 1b involved in the third embodiment creates a generalized model 8 that has a small impact on the output and a high generalization performance. In addition, the generalized model 8 is applied to each second processing device 4 to predict the action result of each second processing device 4. An example of an action result is a processing result.

[0116] The recording unit 11 records the first reference information, which is the sensor information obtained from the observation unit 33 of the first processing device 3, similarly to the first embodiment. Fig.10 Although description is omitted in the figure, the first reference information is sensor information obtained when the first processing device 3 executes a processing operation according to the reference processing parameter 2, similarly to the first embodiment.

[0117] The reference information acquisition unit 12 acquires sensor information output from the observation unit 43 of each second processing device 4-1 to 4-N as second reference information. The sensor information acquired as the second reference information is sensor information obtained when the second processing devices 4-1 to 4-N perform processing operations according to the reference processing parameters 2.

[0118] The correction function learning unit 13 learns the correction function by the same learning method as that of the first embodiment, using the first reference information recorded in the recording unit 11 and the second reference information acquired by the reference information acquisition unit 12 .

[0119] The inverse correction function creation unit 14 creates an inverse correction function which is the inverse function of the correction function learned by the correction function learning unit 13. The correction function learned by the correction function learning unit 13 is assumed to be a reversible function.

[0120] The individual simulation data creation unit 15 creates individual simulation data using the inverse correction function created by the inverse correction function creation unit 14. Specifically, the individual simulation data creation unit 15 extracts the label corresponding to the sensor information from the database 5 recording the sensor information and labels obtained by the first processing device 3, applies each inverse correction function to the sensor information to create individual simulation information, and creates individual simulation data by associating the label corresponding to the sensor information. The individual simulation data created by the individual simulation data creation unit 15 is used in the generalization model learning unit 16.

[0121] 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 creating unit 15. Specifically, the generalized model learning unit 16 extracts the sensor information from the database 5 and extracts the individual simulation information from the individual simulation data, and learns the relationship between the extracted sensor information and the label corresponding to the individual simulation information, thereby creating the generalized model 8.

[0122] The generalization model 8 may be a common machine learning model, and in one example, it is a neural network. In other examples, it is a branch tree, a support vector machine, etc., and a common regression model or classification model can be used according to the purpose of the model. In addition, in one example, the output of the generalization model 8 is obtained by digitizing the pass or fail of the processing. In other examples, it is the abnormality degree of the processing device, the abnormality judgment information of the processing device, the wear degree of the tool, the value of the processing parameter, etc.

[0123] A part or all of the learning device 1b may be provided in each second processing device 4, or may be provided outside each second processing device 4. In addition, it may also be configured that each second processing device 4 is cloud-connected via a network, and all or part of the learning device 1b is implemented by the processing circuit of the cloud server. For example, in the case where a part of the learning device 1b is configured by a cloud server via a network connection, the recording unit 11, the reference information acquisition unit 12, the correction function learning unit 13, and the inverse correction function creation unit 14 may be respectively provided in the cloud server connected to the network, or may be provided in each second processing device 4. If the individual simulation data creation unit 15 and the generalized model learning unit 16 are also within the range that the data of the database 5 can be used, they can be freely configured. In addition, it may be configured that for each second processing device 4, the second reference information obtained 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 collected manually, and gathered in an electronic computer connected to the database 5, and all or part of the processing of the learning device 1b is implemented by the electronic computer.

[0124] If there is at least one second processing device 4, the learning device 1b can be configured, but it is preferable to prepare more than two second processing devices 4 to configure the learning device 1b. The reason is that the learning device 1b is configured so as to be able to learn the individual differences of various second processing devices 4, and the more second processing devices 4 there are, the more generalized generalization model 8 with higher generalization can be created in the generalization model learning unit 16.

[0125] The generalized model 8 created by the learning device 1b is applied to the second processing device 4 via a network or a human hand. The second processing device 4 to which the generalized model 8 is applied may include a second processing device 4 from which the learning device 1b has not obtained the second reference information used for creating the generalized model 8. That is, the generalized model 8 can also be installed in a new second processing device 4. In the case where the second processing device 4 has the generalized model 8 instead of the model 45 described in the first embodiment, the correction function 44 in the second processing device 4 having the generalized model 8 directly outputs the sensor information input from the observation unit 43 and inputs it to the generalized model 8.

[0126] As described above, in the learning device 1b involved in this embodiment, the inverse correction function creation unit 14 creates the inverse correction function, which is the inverse function of the correction function created by the correction function learning unit 13 described in the first embodiment, the individual simulation data creation unit 15 creates individual simulation data using the inverse correction function, the sensor information obtained by observing the state of the first processing device 3, and the label attached to the sensor information, and the generalization model learning unit 16 creates the generalization model 8 based on the sensor information and the individual simulation data. According to this embodiment, there are a plurality of second processing devices 4, and even if they have individual differences, by applying the generalization model 8, the influence of the individual differences can be suppressed and a high-precision output can be obtained. That is, a generalization model 8 with high generalization can be created.

[0127] Furthermore, in the case where the second processing devices 4-1 to 4-N each have a part of the components of the learning device 1b, such as the correction function learning unit 13, the learning device 1b obtains the correction function created by each of the second processing devices 4-1 to 4-N instead of the second reference information, and uses the obtained correction function to create an inverse correction function by the inverse correction function creating unit 14. Furthermore, in the case where the correction function learning unit 13 and the inverse correction function creating unit 14 are configured as the second processing devices 4-1 to 4-N each, the learning device 1b obtains the inverse correction function created by each of the second processing devices 4-1 to 4-N instead of the second reference information, and uses the obtained inverse correction function to create individual simulation data by the individual simulation data creating unit 15.

[0128] In the learning device 1b involved in this 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 less than the case where various sensor information is collected from each second processing device 4. That is, the capacity of the data collected for learning can be reduced. Therefore, the time required for data collection is short, and a large storage area is not required. In addition, it is not necessary to form a high-speed network, so it becomes easy to obtain data from each second processing device 4.

[0129] The configuration described in the above embodiment is merely an example, and may be combined with other known technologies, the embodiments may be combined with each other, and part of the configuration may be omitted or changed without departing from the gist.

[0130] Description of the label

[0131] 1, 1b learning device, 2 reference processing parameters, 3 first processing device, 4, 4-1, 4-N second processing 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 processing system, 331, 431 sensor, 332 feature quantity extraction unit.

Claims

1. A learning device, It is characterized in that In a case where a model obtained by associating sensor information obtained by an observation unit of a processing device having an observation unit, namely a first processing device, and a label corresponding to the sensor information and learning the model is used in a processing device different from the first processing device, namely a second processing device, wherein the observation unit obtains a sensor signal representing an observation result of at least any one of a state of a workpiece and a state of the processing device during processing according to processing parameters, and any one of feature quantities of the sensor signal as the sensor information, have: a recording unit that records, as first reference information, the sensor information when processing is performed in the first processing device using a reference processing parameter that is one of the processing parameters; a reference information acquisition unit that acquires the sensor information when processing is performed by the reference processing parameter in the second processing device as second reference information; and A correction function learning unit learns a correction function used for the correction when the sensor information is corrected in the second processing device and then input into the model so that the characteristics of the corrected second reference information approach the characteristics of the first reference information.

2. The learning device according to claim 1, It is characterized in that The sensor information is used as a feature quantity of the sensor signal.

3. The learning device according to claim 1, It is characterized in that The correction function learning unit performs the learning by using the distribution of the feature amount of the first reference information as the feature of the first reference information and the distribution of the feature amount of the second reference information as the feature of the second reference information.

4. The learning device according to any one of claims 1 to 3, It is characterized in that The recording unit further records the tag corresponding to the first reference information. The correction function learning unit selects the first reference information corresponding to the same label as the label associated with the second reference information acquired by the reference information acquisition unit from the recording unit, and performs the learning using the selected first reference information and the second reference information acquired by the reference information acquisition unit.

5. The learning device according to any one of claims 1 to 3, It is characterized in that The reference processing parameters are multiple groups of processing parameters. The recording unit records, as a first group, each of the sensor information when the first processing device performs processing by each processing parameter included in the reference processing parameter, i.e., the first reference information, and each of the tags corresponding to each of the first reference information. The reference information acquisition unit acquires, as a second group, each of the sensor information when the second processing device performs processing implemented by each processing parameter included in the reference processing parameter, that is, the second reference information and the label corresponding to each of the second reference information, The correction function learning unit extracts the first reference information and the second reference information which are created when machining is performed using the same machining parameter and are associated with the same label from the first group and the second group, respectively, as the first reference group and the second reference group. using a distribution of a feature amount of the first reference information included in the first reference group as the feature of the first reference information, The learning is performed using a distribution of feature amounts 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, It is characterized in that The correction function learning unit extracts the first reference information and the second reference information created when processing is performed using the same processing parameters and associated with the same label from the first group and the second group in the same proportion, respectively, and sets them as the first reference group and the second reference group.

7. The learning device according to claim 5, It is characterized in that The correction function learning unit extracts the same number of the first reference information and the second reference information that are created when machining is performed using the same machining parameters and are associated with the same label from the first group and the second group, respectively, and sets them as the first reference group and the second reference group.

8. The learning device according to claim 1, It is characterized in that The reference information acquisition unit acquires, in at least one of the second processing devices, the sensor information when processing is performed using the reference processing parameters as the second reference information. The correction function learning unit learns the correction function individually for each of the second processing devices. have: An inverse correction function creation unit, which creates an inverse function of the correction function, namely, an inverse correction function; an individual simulation data creation unit that creates individual simulation information by applying the inverse correction function to the sensor information acquired by the observation unit of the first processing device, and creates individual simulation data by associating the individual simulation information with a tag corresponding to the sensor information used to create the individual simulation information; as well as a generalization model learning unit that creates a generalization model by learning the relationship between the sensor information and the individual simulation information obtained by the observation unit of the first processing device and the corresponding labels, The generalized model is used as the model used by the second processing device.

9. A numerical control device comprising the learning device according to claim 1 or 8, which controls a processing device that operates as the second processing device, The numerical control device is characterized in that An input unit is provided for receiving a selection from a user as to whether to input the sensor information before the correction using the correction function or the sensor information after the correction using the correction function into the model used in the second processing device.

10. The numerical control device according to claim 9, It is characterized in that A display unit that, when the input unit receives the selection from the user, displays the output of the model when the sensor information before the correction is input and the output of the model when the sensor information after the correction is input.

11. A numerical control device having the learning device described in claim 1, controlling a processing device that operates as the second processing device, The numerical control device is characterized in that the sensor information obtained by the second processing device during operation is corrected by the correction function, and the corrected sensor information is input to the model to obtain an output.

12. A numerical control device having the generalization model created by the learning device described in claim 8, controlling a processing device that operates as the second processing device, The numerical control device is characterized in that the sensor information obtained by the second processing device during operation is input to the generalization model to obtain an output.

13. A machining result prediction method that is a machining result prediction method in the case of using a model learned by associating sensor information obtained by an observation unit of a machining device, i.e., a first machining device, with a label corresponding to the sensor information for a machining device different from the first machining device, i.e., a second machining device, wherein the observation unit obtains, as the sensor information, at least any one of a sensor signal representing an observation result of at least any one of the state of the workpiece and the state of the machining device and a feature amount of the sensor signal during machining while machining according to machining parameters, The machining result prediction method is characterized by including the following steps: A recording unit records the sensor information during machining when machining is performed by a reference machining parameter, which is one of the machining parameters, in the first machining device as first reference information; A reference information acquisition unit acquires the sensor information during machining when machining is performed by the reference machining parameter in the second machining device as second reference information; A correction function learning unit learns the correction function used for the correction in the case where the sensor information is corrected in the second machining device and then input to the model so that the feature of the corrected second reference information is close to the feature of the first reference information; the correction function corrects the sensor information obtained from the second machining device during machining; and the model predicts the machining result obtained by the second machining device based on the sensor information corrected by the correction function.

14. A machining system characterized in that it has: A first machining device having an observation unit that obtains, as the sensor information, at least any one of a sensor signal representing an observation result of at least any one of the state of the workpiece and the state of the machining device and a feature amount of the sensor signal during machining while machining according to machining parameters; A second processing device, which has an observation unit identical to the observation unit of the first processing device, and a model that makes predictions based on the sensor information obtained by this observation unit; and A learning device that, when using the model learned by associating the sensor information obtained by the observation unit of the first processing device with a label corresponding to the sensor information in the second processing device, learns a correction function used for the correction in the case where the sensor information obtained by the observation unit of the second processing device is input to the model after being corrected, The learning device includes: A recording unit that records the sensor information when performing processing by a reference processing parameter, which is one of the processing parameters, in the first processing device as first reference information; A reference information acquisition unit that acquires the sensor information when performing processing by the reference processing parameter in the second processing device as second reference information; And A correction function learning unit that learns the correction function so that the characteristics of the corrected second reference information are close to the characteristics of the first reference information.

Images