Tool Life Prediction System

Abstract

To provide a tool life prediction system capable of predicting a tool life with high accuracy by using a plurality of learned models without selecting one learned model from the plurality of learned models.SOLUTION: The tool life prediction system includes a remaining processing count prediction unit that outputs each of remaining processing count prediction values Na, Nb, and Nc by each of a plurality of life prediction models Ma1, Mb1, and Mc1 on the basis of a state data in an estimation phase, a scoring value output unit that outputs each of score values Va, Vb, and Vc by each of a plurality of scoring models Ma2, Mb2, and Mc2 on the basis of the state data in the estimation phase, a total remaining processing count prediction unit that calculates a total remaining processing count Nx on the basis of each of the plurality of predicted remaining processing count values Na, Nb, and Nc and each of the plurality of score values Va, Vb, and Vc.SELECTED DRAWING: Figure 3

Description

Technical Field

[0001] The present invention relates to a tool life prediction system.

Background Art

[0002] Predicting the life of tools used in machining is important from the perspective of tool cost. Conventionally, the tool life was determined in advance by specifying the number of machining operations that reach the life, taking into account the individual variations of the tools and the safety factor. However, with such a life determination method, even when the tool has not actually reached the life, it has been considered that the life has been reached and the tool has been replaced.

[0003] Therefore, in Patent Document 1, it has been proposed to predict the tool life using machining information when machining is performed by a tool, for example, the drive current of the motor of the spindle device. In this prediction method, the tool life is predicted using an arithmetic model determined based on the machining information. Also, the arithmetic model used for predicting the tool life is supposed to be selected from a plurality of arithmetic models according to, for example, the material of the workpiece and the machining conditions.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] Depending on factors such as the type of tool, the material of the workpiece, and the machining conditions, the state in which the tool reaches the life is different. Therefore, it is effective to set an arithmetic model used for predicting the tool life for each life case according to factors such as the type of tool, the material of the workpiece, and the machining conditions.

[0006] However, it is not easy to select one operation model from multiple operation models. For example, if there is an operation model in which all factors such as tools, workpieces, and processing conditions are the same, the operation model can be used. However, when some of the factors are different, for example, when the types of tools are different, it is not easy to determine which operation model should be selected. Furthermore, even if all factors such as tools, workpieces, and processing conditions are the same, the selected operation model may not be appropriate due to the influence of individual differences in tools, workpieces, and the processing machine body.

[0007] Also, consider the case of selecting one learned model from multiple learned models when applying machine learning. In this case, if all the explanatory variables of the multiple learned models are of the same type, it is also possible to select one learned model to be applied based on the explanatory variables in the estimation phase.

[0008] Generally, the selection of feature quantities as explanatory variables in machine learning has a great influence on the estimation accuracy of machine learning. Therefore, it may be effective to use different feature quantities for each learning model as explanatory variables. However, when the feature quantities as explanatory variables of multiple learned models are different for each model, it is not easy to select one learned model from the multiple learned models.

[0009] An object of the present invention is to provide a tool life prediction system that can accurately predict the tool life by using a plurality of learned models without selecting one learned model from the plurality of learned models.

Means for Solving the Problems

[0010] The tool life prediction system includes a machine tool main body that performs machining on a workpiece using a tool, a detector that detects observable state data in the machine tool main body during machining of the workpiece, a life prediction model generated by performing machine learning using the state data detected by the detector as explanatory variables and the remaining number of machining operations of the workpiece until the life of one tool is reached as the target variable, using a training data set including the explanatory variables and the target variable, a life prediction model storage unit that stores the life prediction model for each of a plurality of life cases, and a scoring model generated by performing machine learning using a training data set having each of the state data for each life case as explanatory variables, the scoring model outputting a score value of the state data. When the cases for each of the tools are regarded as separate life cases a scoring model storage unit that stores the scoring model for each of a plurality of life cases, a remaining number of machining operations prediction unit that outputs a predicted value of the remaining number of machining operations by each of the plurality of life prediction models based on the state data in the estimation phase, a score value output unit that outputs a score value by each of the plurality of scoring models based on the state data in the estimation phase, and a comprehensive remaining number of machining operations prediction unit that calculates a comprehensive remaining number of machining operations based on each of the plurality of predicted values of the remaining number of machining operations and each of the plurality of score values.

[0011] According to the above tool life prediction system, the life prediction model and the scoring model are stored in association with each other for each life case. The scoring model outputs a score value of the state data as an explanatory variable. The score value output by the scoring model corresponds to the score value of the corresponding life prediction model, that is, the degree of fitness indicating how well the life prediction model fits the state data. For example, when the score value is large, it is estimated that the fitness of the corresponding life prediction model is high, and when the score value is small, it is estimated that the fitness of the corresponding life prediction model is low.

[0012] Then, the total remaining machining times prediction unit calculates the total remaining machining times based on each of the plurality of remaining machining times prediction values and each of the plurality of score values. In other words, by considering the score values, the total remaining machining times prediction unit can use a plurality of remaining machining times prediction models. That is, the total remaining machining times prediction unit does not need to select only one remaining machining times prediction model from among the plurality of remaining machining times prediction models. Thus, since the total remaining machining times prediction unit predicts the total remaining machining times using a plurality of remaining machining times prediction models and a plurality of score values, the remaining machining times prediction unit can accurately predict the tool life.

Brief Description of the Drawings

[0013]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Embodiments for Carrying Out the Invention

[0014] (1. Types of Machine Tools and Tools to Which the Tool Life Prediction System is Applied) The tool life prediction system predicts the life of a tool in a processing machine that performs machining of a workpiece using the tool. Here, the tool life includes not only the state where the tool becomes completely unusable but also the state where the tool needs to be repaired. For example, tool repair includes regrinding the tool, truing and dressing of the grinding wheel, etc.

[0015] Examples of the processing machine include processing machines that perform cutting operations such as machining centers, lathes, milling machines, boring machines, and gear processing devices. In these cases, the tool becomes a cutting tool. For example, tools in a machining center include drills, milling tools, boring tools, gear processing tools, turning tools, etc. Tools in a lathe include at least turning tools, and tools in a combination lathe include, in addition to turning tools, tools similar to those in a machining center such as drills and milling tools. Examples of gear processing devices include gear skiving machines, hobbing machines, shaper machines, etc., and tools in the gear processing device include gear processing tools such as gear skiving tools, hob tools, and shaper tools.

[0016] In addition, the processing machine includes a grinding machine that performs grinding operations. Tools in a grinding machine include grinding wheels. Also, the processing machine includes forging machines, such as presses and rolling machines. Tools in a press or rolling machine include forging punches, forging dies, etc.

[0017] (2. Outline of the configuration of the tool life prediction system 1) The outline of the configuration of the tool life prediction system 1 will be described with reference to FIG. 1. The tool life prediction system 1 includes a processing machine 10 and arithmetic units (20, 30). The processing machine 10 may be one unit or, as shown in FIG. 1, may be a plurality of units. In this example, the case where the tool life prediction system 1 includes a plurality of processing machines 10 will be taken as an example.

[0018] The processing machine 10 includes at least a processing machine main body 11 that performs machining on a workpiece W using a tool T, and a detector 13 that detects observable state data in the processing machine main body 11 during machining of the workpiece W.

[0019] The arithmetic unit (20, 30) predicts the life of the tool T by applying machine learning using the state data detected by the detector 13. In FIG. 1, the arithmetic unit (20, 30) is composed of a learning processing unit 20 and a prediction arithmetic unit 30. The learning processing unit 20 and the prediction arithmetic unit 30 are shown as independent configurations, but they can also be integrated into one device. Further, part or all of the arithmetic unit (20, 30) can also be an embedded system in the processing machine 10.

[0020] In this example, the case where the learning processing unit 20 and the prediction arithmetic unit 30 have independent configurations will be taken as an example. The learning processing unit 20 executes the processing in the learning phase of machine learning, and the prediction arithmetic unit 30 executes the processing in the estimation phase of machine learning. Further, the learning processing unit 20 has a so-called server function and is communicably connected to a plurality of processing machines 10. On the other hand, the prediction arithmetic unit 30 is provided one-to-one for each processing machine 10 and is communicably connected to each processing machine 10. That is, a plurality of prediction arithmetic units 30 function as so-called edge computers, enabling high-speed arithmetic processing. That is, in this example, the tool life prediction system 1 includes one learning processing unit 20 having a server function and a plurality of prediction arithmetic units 30, 30 corresponding to each of the plurality of processing machines 10, 10.

[0021] (3. Details of the configuration of the tool life prediction system 1) The detailed configuration of the tool life prediction system 1 will be described in more detail with reference to FIG. 1. The tool life prediction system 1 includes a plurality of processing machines 10, 10, one learning processing unit 20 that functions as a part of the arithmetic unit, and a plurality of prediction arithmetic units 30, 30 that function as the other part of the arithmetic unit.

[0022] As described above, each processing machine 10 can apply various processing machines. The processing machine 10 includes a processing machine main body 11 that performs machining on a workpiece W using a tool T, a control device 12 that controls the processing machine main body 11, a detector 13, and an interface 14.

[0023] The processing machine main body 11 has a tool T, supports the workpiece W, and has a configuration for relatively moving the tool T and the workpiece W. That is, the processing machine main body 11 includes a structure and a driving device that drives the structure. The control device 12 includes a CNC device, a PLC device, etc., and is configured with a processor, a storage device, etc. The control device 12 controls the driving device and the like in the processing machine main body 11. The interface 14 is a device that enables communication between the processing machine main body 11, the control device 12, the detector 13, and the outside.

[0024] The detector 13 detects observable state data in the processing machine main body 11 during the machining of the workpiece W. The detector 13 detects, for example, time-series state data related to the machining load, the driving load of the driving device, etc. The detector 13 is, for example, a current sensor that detects the driving current data of a motor as a driving device, a vibration sensor that detects the vibration data of a component of the processing machine main body 11, a microphone that detects the sound data during machining, etc. That is, the state data is, for example, driving current data, vibration data, sound data, etc.

[0025] The learning processing device 20 is configured with a processor 21, a storage device 22, an interface 23, etc. The learning processing device 20 executes the processing in the learning phase of machine learning. Also, in this example, the learning processing device 20 has a server function and is communicably connected to a plurality of processing machines 10.

[0026] The learning processing device 20 generates a life prediction model for predicting the life of the tool T by performing machine learning based on the state data detected by the detector 13. In particular, the learning processing device 20 generates a life prediction model for each of a plurality of life cases. That is, the learning processing device 20 generates a plurality of life prediction models.

[0027] For example, in the learning processing device 20, a life prediction model is generated for each tool T. Whether for different types of tools T or the same type of tools T, assuming that the life cases for each tool T are separate life cases, a life prediction model is generated for each life case. That is, a life prediction model is generated for each tool T regardless of whether they are of the same type or different types. In this way, the learning processing device 20 generates a life prediction model for each of the multiple life cases with each tool T as each life case.

[0028] Furthermore, the learning processing device 20 generates a scoring model that outputs a score value for the state data by performing machine learning based on the state data for each life case detected by the detector 13. That is, the learning processing device 20 generates a scoring model for each of the multiple life cases. Therefore, the same number of scoring models as the life prediction models are generated, and each scoring model corresponds to each life prediction model.

[0029] Here, the score value is a numerical value representing the degree of normality of the state data input to the scoring model. The higher the degree of normality, the larger the score value, and the lower the degree of normality, the smaller the score value. Furthermore, the score value may also indicate a value representing an abnormality in addition to the value representing normality for the state data input to the scoring model. For example, the score value may use a positive value for the value representing normality and a negative value for the value representing an abnormality.

[0030] And, as described above, each scoring model corresponds to each life prediction model. Therefore, when a scoring model outputs a large score value, it means that the life prediction model corresponding to the scoring model has a high degree of fitness for the input state data. Conversely, when a scoring model outputs a small score value, it means that the life prediction model corresponding to the scoring model has a low degree of fitness for the input state data. That is, the scoring model outputs a score value as the degree (degree of fitness) to which the input state data fits the life prediction model corresponding to the scoring model. That is, the score value represents the degree of fitness between the life prediction model and the state data input to the life prediction model.

[0031] Each prediction arithmetic unit 30 includes a processor 31, a storage device 32, an interface 33, and the like. The prediction arithmetic unit 30 executes the processing in the estimation phase in machine learning. Further, the prediction arithmetic unit 30 is communicably connected to the learning processing device 20 as a server and the corresponding machine tool 10.

[0032] Each prediction arithmetic unit 30 is arranged at a position close to each machine tool 10 and functions as a so-called edge computer. The prediction arithmetic unit 30 predicts the life of the tool T based on the state data detected by the detector 13 during the machining of the workpiece W using a plurality of life prediction models and a plurality of scoring models generated by the learning processing device 20. In this example, the prediction arithmetic unit 30 predicts the remaining number of machining operations, that is, the remaining number of times the workpiece W can be machined.

[0033] The tool life prediction system 1 further includes a common display device 40 and a plurality of individual display devices 50. However, the tool life prediction system 1 may be configured not to include the common display device 40 or may be configured not to include the individual display devices 50. The common display device 40 is arranged corresponding to the learning processing device 20. Further, the individual display devices 50 are arranged corresponding to the respective machine tools 10.

[0034] (Example of the processing machine main body 11) As an example of the processing machine main body 11, a machining center will be described with reference to FIG. 2. Note that the processing machine main body 11 is an example of a processing machine that can perform various processes including gear processing as described above, and may be applied to other processing machines.

[0035] As shown in FIG. 2, the processing machine main body 11 as a machining center applies the configuration of a 5-axis machining center having three linear axes and two rotational axes as drive axes for changing the relative position and posture of the workpiece W and the tool T. In this example, the processing machine main body 11 has three orthogonal linear axes (X-axis, Y-axis, Z-axis) as linear axes, and a B-axis and a Cw-axis as rotational axes. The B-axis is a rotational axis around the Y-axis line, and the Cw-axis is a rotational axis around the central axis of the workpiece W.

[0036] The processing machine main body 11 includes a tool spindle 61 that supports the tool T (rotary tool) and is rotatable about the Ct-axis, and is movable in the Y-axis direction and the Z-axis direction, respectively. Further, the processing machine main body 11 includes a workpiece spindle 62 that supports the workpiece W and is rotatable about the Cw-axis, and is rotatable about the B-axis and is movable in the X-axis direction. The processing machine main body 11 has a motor as a drive device for driving each axis (X-axis, Y-axis, Z-axis, B-axis, Cw-axis, Ct-axis).

[0037] (5. Outline of tool life prediction processing) The outline of the tool life prediction processing of the tool T in the tool life prediction system 1 will be described with reference to FIGS. 3 and 4. Here, the following description is given by way of example in order to grasp the outline of the tool life prediction processing. Therefore, the tool life prediction processing is not limited to the following description. The tool life prediction processing applies machine learning and is a process that executes a learning phase by the learning processing device 20 and an estimation phase by the prediction calculation device 30.

[0038] As shown in FIG. 3, in the learning phase by the learning processing device 20, the state data Da during machining by the tool Ta, the state data Db during machining by the tool Tb, and the state data Dc during machining by the tool Tc are used. That is, each case by the tools Ta, Tb, and Tc is regarded as a separate life case.

[0039] First, the learning processing device 20 generates a life prediction model Ma1 by performing machine learning. The life prediction model Ma1 uses the state data Da during machining by the tool Ta as an explanatory variable, and the remaining number of machining operations of the workpiece W until the one tool Ta reaches the end of its life as the target variable, and is generated by performing machine learning using a training data set for life prediction including the explanatory variable and the target variable. That is, the life prediction model Ma1 is a learned model that outputs a predicted value Na of the remaining number of machining operations, which is the target variable, when the state data Da, which is the explanatory variable, is input.

[0040] The learning processing device 20 generates the life prediction model Ma1 by, for example, supervised learning. The life prediction model Ma1 can apply, for example, linear regression, ridge regression, Lasso, elastic net, random forest, support vector machine, etc. The life prediction model Ma1 may be any model that can define the relationship between the state data Da, which is the explanatory variable, and the remaining number of machining operations, which is the target variable.

[0041] Furthermore, the learning processing device 20 generates a scoring model Ma2 by performing machine learning using a training data set for scoring with the state data Da as the explanatory variable. The learning processing device 20 generates a scoring model Ma2 that outputs a score value Va for the state data Da by performing machine learning.

[0042] The learning processing device 20 generates a scoring model Ma2, for example, by unsupervised learning. The scoring model Ma2 can apply, for example, a one-class support vector machine (SVM). As shown in FIG. 4, the one-class SVM can perform machine learning using normal data, set the boundary between normal and abnormal, and output different score values for the score value representing normal and the score value representing abnormal. For example, the one-class SVM outputs a score value represented by a positive value in the case of normal and a score value represented by a negative value in the case of abnormal. Further, the one-class SVM outputs a score value representing the degree of normality in the case of normal. Each range of the score values representing the degree of normality is illustrated by the dashed line in FIG. 4.

[0043] Note that in addition to the one-class SVM, the scoring model Ma2 can apply models that can output the degree of normality, such as an autoencoder, Mahalanobis distance, likelihood, etc. The Mahalanobis distance is an index of distance obtained based on the dispersion of data. The likelihood is an index indicating "plausibility" used in Bayesian estimation, etc.

[0044] Here, it is preferable that the life prediction model Ma1 and the scoring model Ma2 corresponding to the life prediction model Ma1 use the same type of feature quantities as explanatory variables. By using the same type of feature quantities as explanatory variables, the correlation between the two models Ma1 and Ma2 can be increased. However, the feature quantities as explanatory variables of the two models Ma1 and Ma2 do not have to be all of the same type and may be partially different.

[0045] Various statistical quantities can be applied as the feature quantities. For example, the feature quantities can be statistical quantities such as mean, variance, standard deviation, skewness, kurtosis, maximum value, minimum value, median, first quartile, third quartile, difference between the maximum value and the minimum value, etc. Further, the feature quantities may include the above statistical quantities for the data obtained by differentiating the state data, the above statistical quantities for the data obtained by performing frequency analysis of the state data, etc.

[0046] Further, the learning processing device 20 similarly generates a remaining life prediction model Mb1 and a scoring model Mb2 for the state data Db regarding the tool Tb. Furthermore, the learning processing device 20 similarly generates a remaining life prediction model Mc1 and a scoring model Mc2 for the state data Dc regarding the tool Tc.

[0047] In this way, the learning processing device 20 generates remaining life prediction models Ma1, Mb1, Mc1 and scoring models Ma2, Mb2, Mc2 for each of the tools Ta, Tb, Tc, that is, for each remaining life case.

[0048] Here, each of the remaining life prediction models Ma1, Mb1, Mc1 can use different feature quantities for each of the state data Da, Db, Dc as explanatory variables. That is, at least a part of the feature quantity as the explanatory variable of the remaining life prediction model Ma1, the feature quantity as the explanatory variable of the remaining life prediction model Mb1, and the feature quantity as the explanatory variable of the remaining life prediction model Mc1 can be different types of feature quantities.

[0049] Of course, the explanatory variables of the remaining life prediction models Ma1, Mb1, Mc1 may all be the same type of feature quantity. However, by allowing different feature quantities to be explanatory variables, each of the remaining life prediction models Ma1, Mb1, Mc1 can be made into a good model, that is, a model capable of performing highly accurate prediction.

[0050] Next, in the estimation phase by the prediction calculation device 30, the state data Dx during processing by the new tool Tx is used. The prediction calculation device 30 uses the state data Dx regarding the tool Tx as input data and outputs remaining machining times prediction values Na, Nb, Nc as remaining life prediction results using a plurality of remaining life prediction models Ma1, Mb1, Mc1. For example, the remaining machining times prediction values Na, Nb, Nc are 500 times, 480 times, and 510 times, respectively.

[0051] Furthermore, the prediction calculation device 30 uses the state data Dx regarding the tool Tx as input data, and outputs score values Va, Vb, and Vc as scoring results by using a plurality of scoring models Ma2, Mb2, and Mc2. For example, the score values Va, Vb, and Vc are 5, 4, and 2 respectively.

[0052] Then, the prediction calculation device 30 calculates the total remaining processing times Nx by using each of the plurality of score values Va, Vb, and Vc as the weight of each of the plurality of predicted remaining processing times Na, Nb, and Nc. For example, the prediction calculation device 30 calculates the total remaining processing times Nx by weighted average using each of the plurality of score values Va, Vb, and Vc as weights. In this case, the total remaining processing times Nx is obtained by Equation 1. Nx = 500×5 / 11 + 480×4 / 11 + 510×2 / 11 = 495 (Equation 1)

[0053] (6. Functional Block Configuration of Tool Life Prediction System 1) Regarding the functional block configuration of the tool life prediction system 1 for realizing the above-described tool life prediction process, it will be described with reference to FIGS. 5-8. As shown in FIG. 5, the tool life prediction system 1 includes a detector 13, a counter 12a, a learning processing device 20, a prediction calculation device 30, and display devices 40 and 50.

[0054] As described above, the detector 13 detects observable state data in the processing machine main body 11 during the processing of the workpiece W. The state data includes, for example, drive load data in a motor that rotationally drives the rotary tool T. The state data further includes, for example, drive load data in a motor that rotationally drives the workpiece W. The drive load data corresponds to drive current data of the motor. Also, the state data may include vibration data, machining sound data, etc. The state data is time-series data from the start to the end of processing for one workpiece W.

[0055] The counter 12a is included in the control device 12 of the processing machine 10 and counts the number of times the workpiece W is processed for each tool T. That is, the counter 12a counts the number of times the workpiece W is processed for each tool T from the start of use of the tool T. Note that the counter 12a can be provided in the detector 13 itself in addition to the control device 12.

[0056] The learning processing device 20 generates a life prediction model for predicting the life of the tool T based on the state data detected by the detector 13 and the number of processing times by the counter 12a. Further, the learning processing device 20 generates a scoring model that outputs a score value of the input state data based on the state data detected by the detector 13.

[0057] As shown in FIG. 5, the learning processing device 20 includes a training data set acquisition unit 71, a life prediction training data set storage unit 72, a scoring training data set storage unit 73, a life prediction model generation unit 74, and a scoring model generation unit 75.

[0058] The training data set acquisition unit 71 acquires a training data set for performing machine learning for each of a plurality of tools T. The plurality of tools T includes the same type of tool T, different types of tool T, etc. Regardless of whether they are of the same type or different types, acquiring a training data set for each individual tool T means acquiring a training data set for each of a plurality of life cases.

[0059] The training data set acquisition unit 71 includes a state data acquisition unit 71a and a feature amount calculation unit 71b. The state data acquisition unit 71a acquires the state data detected by the detector 13 during the processing of the workpiece W. The state data acquisition unit 71a acquires state data of the detector for several minutes in one processing by one tool T (for example, the processing of one workpiece W).

[0060] Then, the state data acquisition unit 71a acquires the state data of each of a plurality of machining operations until one tool T reaches the end of its life. That is, the state data acquisition unit 71a acquires the state data of the plurality of detectors 13 for the number of machining operations until one tool T reaches the end of its life. And the state data acquisition unit 71a acquires the above state data for each of the plurality of tools T. That is, the state data acquisition unit 71a acquires the state data for each life case.

[0061] The feature amount calculation unit 71b calculates a plurality of feature amounts of the state data acquired by the state data acquisition unit 71a. The feature amount is, as described above, an average value or the like.

[0062] The training data set acquisition unit 71 further includes a machining number acquisition unit 71c, a life reach information acquisition unit 71d, and a remaining machining number calculation unit 71e. The machining number acquisition unit 71c acquires the number of machining operations from the start of machining for each individual tool T from the counter 12a. The life reach information acquisition unit 71d determines whether or not the operator determines that the target tool T has reached the end of its life, for example, input by the operator. The determination as to whether or not the tool T has reached the end of its life can be made, for example, by whether or not a scratch called a tool mark is formed on the surface of the workpiece W. Alternatively, the determination can also be made by whether or not the machining accuracy of the workpiece W has significantly decreased. In addition to being input by the operator, the life reach information acquisition unit 71d may be input by an inspection device when determined by the inspection device.

[0063] The remaining machining number calculation unit 71e determines the remaining machining number of the tool T based on the machining number acquired by the machining number acquisition unit 71c and the life reach information acquired by the life reach information acquisition unit 71d. The remaining machining number increases in ascending order while going back before the end of life with the end of life being zero. As shown in the left column of FIG. 6, when it is determined that the end of life is reached at the machining number N, as shown in the right column of FIG. 6, the remaining machining number becomes zero at the end of life and the first one is (N - 1).

[0064] As shown in the right column of FIG. 6, the training data set storage unit 72 for life prediction stores the feature amounts D(1) to D(N) of the state data and the remaining machining times (N-1) to (0) obtained by the training data set acquisition unit 71 as a training data set for life prediction for each of a plurality of life cases. Further, the training data set storage unit 72 for life prediction stores the feature amounts D(1) to D(N) of the state data and the remaining machining times (N-1) to (0) in an associated manner. The content shown in the right column of FIG. 6 is a training data set regarding one tool T. And in the training data set storage unit 72 for life prediction, a training data set is stored for each of a plurality of tools T, that is, for each of a plurality of life cases.

[0065] The training data set storage unit 73 for scoring stores the feature amounts D(1) to D(N) of the state data obtained by the training data set acquisition unit 71 as a training data set for scoring for each of a plurality of life cases. That is, in the training data set storage unit 73 for scoring, a training data set is stored for each of a plurality of tools T, that is, for each of a plurality of life cases.

[0066] Note that, although the training data set storage unit 72 for life prediction and the training data set storage unit 73 for scoring are separately described for convenience, they may be integrated. That is, an integrated training data set storage unit may store the feature amounts D(1) to D(N) of the state data and the remaining machining times (N-1) to (0) in an associated manner as a training data set for life prediction, and store the feature amounts D(1) to D(N) of the state data as a training data set for scoring. The feature amounts D(1) to D(N) of the state data are common data.

[0067] The life prediction model generation unit 74 performs machine learning using the life prediction training data set stored in the life prediction training data set storage unit 72. Specifically, for each life case, that is, for each individual tool T, the life prediction model generation unit 74 performs machine learning with the feature amount of the state data detected by the detector 13 as the explanatory variable and the remaining number of machining operations as the objective variable. Then, the life prediction model generation unit 74 generates a life prediction model for predicting the life of the tool T.

[0068] Here, the life prediction model is generated for each life case, that is, for each individual tool T. Therefore, the life prediction model generation unit 74 generates as many life prediction models as the number of a plurality of tools T. That is, the correspondence between the tool T and the life prediction model is as shown in FIG. 7. For example, the tool Ta corresponds to the life prediction model Ma1, and the following similar correspondence holds.

[0069] As the machine learning method, for example, regression may be used. For example, linear regression, ridge regression, Lasso, elastic net, random forest, support vector machine, etc. are useful. In particular, by using these methods, the influence degrees of a plurality of feature amounts can be grasped, and they can also be used for the selection of feature amounts as needed. Note that, as the machine learning method, other than regression, it is also applicable.

[0070] The scoring model generation unit 75 performs machine learning using the scoring training data set stored in the scoring training data set storage unit 73. Specifically, for each life case, that is, for each individual tool T, the scoring model generation unit 75 performs machine learning with the feature amount of the state data detected by the detector 13 as the explanatory variable. Then, the scoring model generation unit 75 generates a scoring model that outputs the score value of the state data.

[0071] Here, the scoring model is generated for each life case, that is, for each individual tool T. Therefore, the scoring model generation unit 75 generates the scoring model as many times as the number of a plurality of tools T. That is, the correspondence between the tool T and the scoring model is as shown in FIG. 7. For example, the tool Ta corresponds to the scoring model Ma2, and the following is the same correspondence. Further, as shown in FIG. 7, the scoring model also corresponds to the life prediction model. As the machine learning method, a model that can output the normality, for example, one-class SVM, autoencoder, Mahalanobis distance, likelihood, etc. may be used.

[0072] The prediction calculation device 30 predicts the remaining life of the tool T used for machining based on the state data during machining in the corresponding machine tool 10. The prediction calculation device 30 includes a life prediction model storage unit 81, a scoring model storage unit 82, a prediction data acquisition unit 83, a remaining machining times prediction unit 84, a score value output unit 85, a total remaining machining times prediction unit 86, and an output unit 87.

[0073] As shown in FIG. 8, the life prediction model storage unit 81 stores a plurality of life prediction models Ma1, Mb1, Mc1,... generated by the life prediction model generation unit 74. As described above, the plurality of life prediction models Ma1, Mb1, Mc1,... correspond to each life case, and in this example, they correspond to individual tools T.

[0074] As shown in FIG. 8, the scoring model storage unit 82 stores a plurality of scoring models Ma2, Mb2, Mc2,... generated by the scoring model generation unit 75. As described above, the plurality of scoring models Ma2, Mb2, Mc2,... correspond to each life case, and in this example, they correspond to individual tools T and also correspond to the respective life prediction models Ma1, Mb1, Mc1,...

[0075] The prediction data acquisition unit 83 acquires prediction data during machining by the tool Tx to be predicted. The prediction data acquisition unit 83 includes a state data acquisition unit 83a, a feature amount calculation unit 83b, and a machining times acquisition unit 83c.

[0076] The state data acquisition unit 83a acquires the state data detected by the detector 13 during the machining of the workpiece W by the tool Tx to be predicted. Here, the tool Tx to be predicted may be of the same type or a different type from the type of the tool T used when acquired by the training data set acquisition unit 71.

[0077] The feature amount calculation unit 83b calculates the feature amount of the state data acquired by the state data acquisition unit 83a. Here, various statistical amounts in the state data are used as the feature amount. And the feature amount is of the same type as the feature amount calculated by the feature amount calculation unit 71b of the training data set acquisition unit 71. The machining number acquisition unit 83c acquires the machining number from the start of machining in the tool Tx to be predicted from the counter 12a.

[0078] Here, the state data acquisition unit 83a and the feature amount calculation unit 83b perform the same processing as the state data acquisition unit 71a and the feature amount calculation unit 71b of the training data set acquisition unit 71. And in this example, the state data acquisition unit 83a and the feature amount calculation unit 83b are described as separate elements from the state data acquisition unit 71a and the feature amount calculation unit 71b of the training data set acquisition unit 71. However, it is also possible to use each element 71a, 71b of the training data set acquisition unit 71 as each element 83a, 83b of the prediction data acquisition unit 83. That is, the functions of the elements 71a, 71b in the learning processing device 20 are also used as part of the functions of the prediction calculation device 30.

[0079] The remaining machining times prediction unit 84 outputs predicted values Na, Nb, Nc, ··· of the remaining machining times of the tool Tx to be predicted by each of a plurality of life prediction models Ma1, Mb1, Mc1, ··· based on the state data of the target tool Tx acquired by the prediction data acquisition unit 83. Specifically, the remaining machining times prediction unit 84 uses, as input data, at least a part of the feature amounts calculated by the feature amount calculation unit 83b and the machining times from the start of machining acquired by the machining times acquisition unit 83c for each of the plurality of life prediction models Ma1, Mb1, Mc1, ···, and outputs a plurality of predicted values Na, Nb, Nc, ··· of the remaining machining times for the tool Tx to be predicted. In this way, the remaining machining times prediction unit 84 predicts the number of workpieces W that the tool Tx to be predicted can machine for a plurality of patterns.

[0080] The score value output unit 85 outputs score values Va, Vb, Vc, ··· by each of a plurality of scoring models Ma2, Mb2, Mc2, ··· based on the state data of the target tool Tx acquired by the prediction data acquisition unit 83. Specifically, the score value output unit 85 uses, as input data, at least a part of the feature amounts calculated by the feature amount calculation unit 83b for each of the plurality of scoring models Ma2, Mb2, Mc2, ···, and outputs score values Va, Vb, Vc, ··· for the state data of the target tool Tx. That is, the score value output unit 85 outputs score values Va, Vb, Vc, ··· for each of the plurality of predicted values Na, Nb, Nc, ··· of the remaining machining times.

[0081] The comprehensive remaining machining times prediction unit 86 calculates the comprehensive remaining machining times Nx based on each of the plurality of predicted values Na, Nb, Nc, ··· of the remaining machining times predicted by the remaining machining times prediction unit 84 and each of the plurality of score values Va, Vb, Vc, ··· calculated by the score value output unit 85. For example, the comprehensive remaining machining times Nx is calculated by weighted average using each of the plurality of score values Va, Vb, Vc, ··· as the weights of each of the plurality of predicted values Na, Nb, Nc, ··· of the remaining machining times. The comprehensive remaining machining times Nx is not limited to weighted average, and may be calculated using the plurality of score values Va, Vb, Vc, ··· as weights.

[0082] Further, when all of the plurality of score values represent normal positive values, as described above, the total remaining machining times prediction unit 86 can perform calculation by weighted average using all of the score values. However, when there are negative values representing abnormalities among a part of the plurality of score values, it is advisable to adopt a different processing method. For example, the predicted remaining machining times value that is a negative value representing an abnormality in the score value may be excluded, and the total remaining machining times Nx may be calculated by weighted average using the remaining predicted remaining machining times values and the score values corresponding to the predicted remaining machining times values. Further, when a predetermined number (a number of 1 or more) of score values are negative values representing abnormalities, it is also possible to predict that the target tool Tx is in an abnormal state rather than a normal wear state. The abnormal state of the target tool Tx is, for example, chipping, chipping, etc.

[0083] The output unit 87 outputs the total remaining machining times Nx of the target tool Tx predicted by the total remaining machining times prediction unit 86 to the display devices 40 and 50. In this case, the display devices 40 and 50 display the total remaining machining times Nx, and the operator can grasp the total remaining machining times Nx of the target tool Tx. Further, in addition to the total remaining machining times Nx, the output unit 87 can also output to the display devices 40 and 50 the current machining times, each score value of the life prediction model, and the like.

[0084] Further, the output unit 87 outputs the total remaining machining times Nx of the target tool Tx to the control device (control unit) 12, and the control device 12 may execute an operation of tool change or tool correction based on the total remaining machining times Nx. For example, when the total remaining machining times Nx reaches the life value (for example, zero), the control device 12 executes tool change or tool correction. Note that the tool change and the tool correction are set according to the type of the tool T, and in the case of a tool T for which tool correction is possible, when the number of tool corrections reaches a predetermined number, tool change is performed.

[0085] Here, the tool correction may be performed by removing the tool T from the machine tool main body 11, or may be performed by the machine tool main body 11. For example, when the tool T is a cutting tool, the tool correction is a regrinding process of the tool T, and is usually performed by removing the tool T from the machine tool main body 11. When the tool T is a grinding wheel, the tool correction is truing or dressing of the grinding wheel, and the process of the tool correction is performed on a grinding machine equipped with the grinding wheel.

[0086] Further, the output unit 87 may notify the timing of tool change or tool correction based on the total remaining machining times Nx. For example, the output unit 87 may notify not only at the timing when the total remaining machining times Nx reach the life value, but also at the timing when approaching the life value. Thereby, the operator can perform preparations for tool change or tool correction, preparations for the tool T to be replaced, and the like.

[0087] According to the tool life prediction system 1, the life prediction models Ma1, Mb1, Mc1 and the scoring models Ma2, Mb2, Mc2 are stored in association with each life case. The scoring models Ma2, Mb2, Mc2 output the score values Va, Vb, Vc of the state data as explanatory variables. The score values Va, Vb, Vc output by the scoring models Ma2, Mb2, Mc2 correspond to the score values of the corresponding life prediction models Ma1, Mb1, Mc1, that is, the degree of fitness indicating how well the life prediction models Ma1, Mb1, Mc1 fit the state data. For example, when the score values Va, Vb, Vc are large, it is estimated that the degree of fitness of the corresponding life prediction models Ma1, Mb1, Mc1 is high, and when the score values Va, Vb, Vc are small, it is estimated that the degree of fitness of the corresponding life prediction models Ma1, Mb1, Mc1 is low.

[0088] Then, the total remaining machining times prediction unit 86 calculates the total remaining machining times Nx based on each of the plurality of remaining machining times prediction values Na, Nb, Nc and each of the plurality of score values Va, Vb, Vc. In other words, the total remaining machining times prediction unit 86 can use the plurality of life prediction models Ma1, Mb1, Mc1 by considering the score values Va, Vb, Vc. That is, the total remaining machining times prediction unit 86 does not need to select only one life prediction model from the plurality of life prediction models Ma1, Mb1, Mc1. Thus, since the total remaining machining times prediction unit 86 predicts the total remaining machining times Nx using the plurality of life prediction models Ma1, Mb1, Mc1 and the plurality of score values Va, Vb, Vc, the life prediction of the tool T can be performed with high accuracy.

Explanation of Signs

[0089] 1: Tool life prediction system, 10: Machine tool, 11: Machine tool body, 12: Control device (control unit), 13: Detector, 20: Learning processing device, 30: Prediction calculation device, 40, 50: Display device, 71: Training data set acquisition unit, 72: Life prediction training data set storage unit, 73: Scoring training data set storage unit, 74: Life prediction model generation unit, 75: Scoring model generation unit, 81: Life prediction model storage unit, 82: Scoring model storage unit, 83: Prediction data acquisition unit, 84: Remaining machining times prediction unit, 85: Score value output unit, 86: Total remaining machining times prediction unit, 87: Output unit, Ma1, Mb1, Mc1: Life prediction models, Ma2, Mb2, Mc2: Scoring models, Na, Nb, Nc: Remaining machining times prediction values, Nx: Total remaining machining times, T: Tool, Va, Vb, Vc: Score values, W: Workpiece

Claims

1. A machining machine body for machining a workpiece using a tool, A detector for detecting observable state data in the machining machine body during machining of the workpiece, A life prediction model generated by performing machine learning using the state data detected by the detector as an explanatory variable and the remaining machining times of the workpiece until the life of one of the tools reaches its end as a target variable, and storing the life prediction model for each of a plurality of the life cases when each case by each of the tools is regarded as a separate life case in a life prediction model storage unit, A scoring model generated by performing machine learning using a training data set having each of the state data for each of the life cases as an explanatory variable and outputting a score value of the state data, and storing the scoring model for each of a plurality of the life cases in a scoring model storage unit, A remaining machining times prediction unit that outputs a predicted value of the remaining machining times by each of a plurality of the life prediction models based on the state data in the estimation phase, A score value output unit that outputs a score value by each of a plurality of the scoring models based on the state data in the estimation phase, A comprehensive remaining machining times prediction unit that calculates a comprehensive remaining machining times based on each of a plurality of the predicted values of the remaining machining times and each of a plurality of the score values, A tool life prediction system comprising the above components.

2. The comprehensive remaining machining times prediction unit calculates the comprehensive remaining machining times by weighted average, using each of a plurality of the score values as a weight for each of a plurality of the predicted values of the remaining machining times. The tool life prediction system according to Claim 1.

3. When a predetermined number of a plurality of the score values indicate values representing abnormalities, the comprehensive remaining machining times prediction unit predicts that the tool is in an abnormal state rather than a normal wear state. The tool life prediction system according to Claim 1 or 2.

4. At least a part of a plurality of the life prediction models uses different feature amounts of the state data as explanatory variables. The tool life prediction system according to any one of Claims 1 to 3.

5. The scoring model uses the same type of feature amounts as the life prediction model for the corresponding life case as explanatory variables. The tool life prediction system according to Claim 4.

6. The tool life prediction system according to any one of claims 1 to 5, wherein the scoring model is a one-class support vector machine and outputs the degree of normality as a score value. **Claim 7** The tool life prediction system according to claim 6, wherein the scoring model further outputs different values for the score value representing normality and the score value representing abnormality. **Claim 8** The tool life prediction system further comprises a display device for displaying the total remaining machining times, the tool life prediction system according to any one of claims 1 to 7. **Claim 9** The tool life prediction system further comprises a control unit for performing an operation of tool replacement or tool correction based on the total remaining machining times, the tool life prediction system according to any one of claims 1 to 8. **Claim 10** The tool life prediction system further comprises an output unit for notifying the timing of tool replacement or tool correction based on the total remaining machining times, the tool life prediction system according to any one of claims 1 to 9.

Images