Method and device for selecting machine learning models

Abstract

To enable quick and easy selection of a machine learning model that is appropriate for determining input machine information.SOLUTION: A machine learning model selection method stores condition range information that creates a relative frequency distribution for each condition item related to machining for a learning data group used in learning a machine learning model in advance, and includes the steps of: acquiring condition information that shows information about machining in current time (S1); calculating a priority which is a parameter used to select the machine learning model based on the acquired condition information and the stored condition range information (S2); selecting the machine learning model based on the calculated priority (S3); making determination using machine information acquired from a machine tool as input to the selected machine learning model (S4); and determining a next machine operation content of the machine tool based on the output determination result (S5).SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] This disclosure relates to a method and apparatus for selecting one machine learning model from a plurality of trained machine learning models for making decisions regarding the mechanical information of a machine tool. [Background technology]

[0002] A machine learning model is a mathematical model that extracts latent features from a vast amount of training data and optimizes its parameters to minimize the error in the output. Therefore, unlike physical models, machine learning models are known to be capable of making ambiguous judgments similar to human judgment, but their output cannot be controlled. To suppress the resulting instability in accuracy and improve judgment accuracy, machine learning models are trained using datasets that include multiple conditions, patterns, or noise. Regarding model training, for example, Patent Document 1 discloses an invention in which, in an identification system that stores a model for identifying objects represented by data, the model is retrained using labels and training data for data learned in another identification system, and the trained model is transmitted to the other identification system for updating. On the other hand, in order to suppress the instability of judgment accuracy, judgments may be made comprehensively from the output results of multiple machine learning models, or the optimal machine learning model may be used depending on the object of judgment. Regarding the selection of models, for example, Patent Document 2 describes an invention in which a data collection system that holds multiple classification models (learning models) classifies labeled input data transmitted from the user environment and presents an appropriate classification model to the user environment based on the classification result. [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] International Publication No. 2020 / 49636 [Patent Document 2] Japanese Patent Publication No. 2020-194355 [Overview of the project] [Problems that the invention aims to solve]

[0004] When using multiple machine learning models for decision-making, it is necessary to select which of the applied models to use for the decision. However, since machine learning models are often trained on data containing multiple conditions, the range of conditions in the training data given to each model may overlap. In this case, it becomes difficult to determine which machine learning model is most suitable for the event to be decided. In the field of machining, a wide range of conditions, including cutting speed, depth of cut, tool geometry, material properties, and machine rigidity, are intricately intertwined. As a result, the range of conditions learned by created machine learning models often overlaps. Consequently, selecting the right machine learning model is extremely difficult. In the invention described in Patent Document 2, although the classification model itself is provided by the collection system, the user has to go through the trouble of sending the processing data of the case they want to diagnose to the collection system themselves, and they also have to wait for the collection system to present a model, which may prevent quick model selection in the machining field.

[0005] Therefore, this disclosure has been made in consideration of these circumstances, and aims to provide a method and apparatus for selecting a machine learning model that enables the rapid and easy selection of a machine learning model suitable for determining newly input machine information from among a plurality of machine learning models that have been stored in advance. [Means for solving the problem]

[0006] To achieve the above objective, the first configuration of this disclosure is a method for selecting a machine learning model for determining newly input machine information from a plurality of trained machine learning models prepared for determining machine information of a machine tool, The proportion of each of the multiple predetermined condition items related to the judgment in the data set used to train the machine learning model. This was shown using a relative frequency distribution. A condition range information storage step in which condition range information is created and stored for each machine learning model, A condition information acquisition step involves acquiring condition information that indicates feature quantities related to the machine information newly input from the current operation or planned operation of the machine tool, Based on the aforementioned condition range information and the aforementioned condition information, the priority is a parameter that selects the machine learning model for the newly input machine information. By extracting the relative frequency of the value of the condition information from the relative frequency distribution for each of the condition items and multiplying them, The priority calculation step to be calculated, The method is characterized by performing a selection instruction step of comparing the priorities among the plurality of machine learning models and specifying one or more of the machine learning models. Another aspect of the first configuration is characterized in that, in the priority calculation step, when multiplying each of the relative frequencies for each of the condition items, the weight of each relative frequency is determined from the physical model related to machining acquired by the machine tool, and the priority is calculated. To achieve the above objective, the second configuration of this disclosure is a device for selecting a machine learning model for determining newly input machine information from a plurality of trained machine learning models prepared for determining machine information of a machine tool, The proportion of each of the multiple predetermined condition items related to the judgment in the data set used to train the machine learning model. This was shown using a relative frequency distribution. Condition range information storage means for creating and storing condition range information for each machine learning model, Condition information acquisition means for acquiring condition information that indicates feature quantities related to the machine information newly input from the current operation or planned operation of the machine tool, Based on the aforementioned condition range information and the aforementioned condition information, the priority is a parameter that selects the machine learning model for the newly input machine information. By extracting the relative frequency of the value of the condition information from the relative frequency distribution for each of the condition items and multiplying them, The priority calculation method used, Selection instruction means for comparing the priorities among the plurality of machine learning models and designating one or more of the machine learning models, characterized in that it comprises.

Advantages of the Invention

[0007] According to the present disclosure, it becomes possible to easily and quickly select and use the machine learning model that has been optimally learned for the determination of newly input machine information.

Brief Description of the Drawings

[0008] [Figure 1] It is a schematic diagram of a selection device for a machine learning model. [Figure 2] It is a flowchart of a method for selecting a machine learning model. [[ID=2E]] [Figure 3] It is an explanatory diagram showing an overview of condition range information of a machine learning model. [Figure 4] It is an explanatory diagram showing an overview of a priority calculation method for selecting a machine learning model.

Modes for Carrying Out the Invention

[0009] Hereinafter, embodiments of the present disclosure will be described based on the drawings. FIG. 1 is a schematic diagram showing an example of a selection device for a machine learning model. The selection device is provided in an NC device 101 that controls the operation of a machine tool main body 100. The NC device 101 is configured to include a CPU and a memory connected to the CPU, and the operation control of the machine tool main body 100 is realized by these. The NC device 101 includes a machine learning model control unit 110, a prediction execution unit 130, and an operation determination unit 140 as functional units for selecting a machine learning model. The machine learning model control unit 110 selects a machine learning model based on machine information, which is information related to the machining of the machine tool main body 100, and outputs it to the prediction execution unit 130. The prediction execution unit 130 uses the machine learning model output from the machine learning model control unit 110 to take machine information acquired by the machine tool body 100 as input and makes a determination (for example, normal / abnormal) about a predetermined phenomenon in the machine tool body 100. The operation determination unit 140 determines the operation content of the machine tool body 100 based on the determination of the prediction execution unit 130.

[0010] The machine learning model control unit 110 includes a condition information acquisition unit 111, a machine learning model holding unit 112, a priority calculation unit 113, and a selection instruction unit 114. The condition information acquisition unit 111 acquires condition information, which is information indicating characteristic quantities related to machining, from the machine tool body 100. The condition information acquisition unit 111 is an example of the condition information acquisition means of this disclosure. The machine learning model storage unit 112 stores multiple trained machine learning models in a file group 120 for making judgments on the machine information of the machine tool body. The file group 120 is provided with a condition range information storage unit 121 that stores condition range information for each machine learning model. The condition range information will be described later. The condition range information storage unit 121 is an example of the condition range information storage means of this disclosure. The priority calculation unit 113 calculates a priority, which is a parameter used to select a machine learning model, based on the condition information acquired by the condition information acquisition unit 111 and the condition range information of the machine learning model held in the condition range information holding unit 121. The priority calculation unit 113 is an example of the priority calculation means of this disclosure. The selection instruction unit 114 instructs the machine learning model holding unit 112 to select a machine learning model based on the priority calculated by the priority calculation unit 113 and send it to the prediction execution unit 130. The selection instruction unit 114 is an example of the selection instruction means of the present disclosure.

[0011] Next, the method by which the NC device 101 selects a machine learning model will be explained based on the flowchart in Figure 2. This selection method is executed by a program that is not temporarily stored in the memory of the NC device 101. The condition range information storage unit 121 is pre-stored with condition range information for the machine learning model (condition range information storage step). As shown in Figure 3, this condition range information is created by generating a relative frequency distribution for each machining condition item A, B, C, etc., for the training data set used when training the machine learning model. For example, for all information related to machining, such as the feed rate and rotational speed of each axis, and various information on tools and workpieces, the relative frequency distribution of the training data for the target machine learning model is stored as condition range information in the condition range information storage unit 121. This is done for all machine learning models fed into the NC device 101. However, the condition range information only needs to show the proportion of the data used for training, and can be expressed using probability distributions or probability densities. Also, although we expressed the condition range information for the training data in this example, it can also be done for the test data.

[0012] First, in step 1 (hereinafter referred to as "S"), the condition information acquisition unit 111 acquires condition information from the machine tool body 100 that indicates information regarding the machining process at the current time (condition information acquisition step). Conditional information includes the motor current value used for the axis movement of the machine tool body 100, log data acquired as an event when a predetermined machine operation is performed, and a coefficient for converting the motor current value into thrust. However, since various parameters exist in machining, it is also possible to acquire all information related to machining and machine operation as conditional information, such as temperature at various points on the machine tool body 100, vibration, tool specifications, and type of coolant. By utilizing this data, the condition information acquisition unit 111 can acquire characteristic quantities such as cutting speed, cutting cross-sectional area, and specific cutting resistance in machining. Other characteristic quantities that can be obtained include temperature, cutting ratio, tool tip angle, and friction force. Next, in S2, the priority calculation unit 113 calculates the priority, which is a parameter used to select a machine learning model, from the condition information acquired by the condition information acquisition unit 111 and the condition range information of the machine learning model held by the condition range information holding unit 121 (priority calculation step).

[0013] This priority will be explained with reference to Figure 4. For example, the resistance value F in turning is expressed by the following equation (1), based on the material coefficient Kc and the cutting cross-sectional area Ac. F = Kc·Ac ··(1) At this time, the priority calculation unit 113 determines the priority P from the relative frequency p obtained from the relative frequency distribution of the material coefficient Kc in the number of training data. Kc And the relative frequency p obtained from the relative frequency distribution of cutting cross-sectional area Ac in the number of training data. Ac Using these, the calculation is performed as shown in equation (2) below. P=p Kc ·p Ac (2) In other words, the relative frequencies of the material coefficient Kc and cutting cross-sectional area Ac, which are obtained as condition information, are determined from the relative frequency distributions created in advance using these as condition items, and the priority P is calculated by multiplying them as shown in equation (2). If, for any reason, data for a certain variable cannot be obtained, that term may be excluded from the calculation, or new variables such as parameters related to tool specifications or the presence or absence of cutting fluid may be added to the calculation. However, the calculations performed by the machine learning models being compared must be identical.

[0014] Priority P can also be calculated by weighting each relative frequency p according to the degree of the variable and then summing them up. For example, if a machine information f is a function X(A), Y(1 / B), Z(C) for each condition item A, B, C, then the priority P can be calculated as follows: n Let's consider the case where it is expressed by the following equation (3) using ). f = X(A)·Y(1 / B)·Z(C) n ) ··(3) Regarding variables A, B, and C, when the numerical values of the current operation are a, b, and c, the relative frequency distribution p A (A), p B (B), p C (C), the relative frequency p A (a), p B (b), p C (c) n is used to calculate the priority Pi of model i by the following formula (4). Pi = p A (a)·p B (b)·p C (c) n ··(4) In this way, the priority can be expressed by a physical model represented by a simple product.

[0015] Next, in S3, the selection instruction unit 114 selects a machine learning model based on the calculated priority (selection instruction step). In the selection instruction unit 114, for the machine learning models (model1, model2, model3, ···, modeli) held in the machine learning model holding unit 112, the priorities (P1, P2, P3, ···, P i ) are obtained. Among them, the machine learning model with the maximum priority is selected, and it is transmitted to the prediction execution unit 130 to perform prediction with that machine learning model. Note that due to the constraints of the control device, when it is determined that there is a margin, it is not always necessary to extract only the machine learning model with the maximum priority, and it may be transmitted to the prediction execution unit 130 to use a plurality of machine learning models. Next, in S4, the machine learning model holding unit 112 sends the selected machine learning model to the prediction execution unit 130, and the prediction execution unit 130 makes a determination using the selected machine learning model with the machine information acquired by the machine tool body 100 as input. Next, in S5, the operation determination unit 140 determines the machine operation content of the next machine tool body 100 based on the determination result output by the prediction execution unit 130.

[0016] In this manner, the machine learning model selection method and NC device 101 (an example of a selection device) create and maintain condition range information for each machine learning model, which is the proportion of multiple predetermined condition items related to judgment in the training data (an example of a data set) used to train the machine learning model. It acquires condition information that indicates the features of the machine information newly input from the current or planned operation of the machine tool. Based on the condition range information and condition information, it calculates a priority, which is a parameter for selecting a machine learning model for the newly input machine information, and compares the priorities among multiple machine learning models to specify one or more machine learning models. This configuration makes it possible to easily and quickly select and use an optimally trained machine learning model for determining newly input machine information.

[0017] In particular, since priority is calculated from the relative frequency distribution and condition information that are stored in advance as data proportions, it is possible to select a machine learning model that is likely to have been trained using data under the specified conditions, and thus use a machine learning model with a higher accuracy rate. Furthermore, since priority is calculated by multiplying relative frequencies, the model shows that all features have been trained evenly. Therefore, it outputs a higher priority than summation. For example, the probability density p of feature A. a , the probability density p of feature B b If the data is obtained as a percentage, then when the priority is calculated by adding it, p a =0.5, p b When = 0.5 and p a =0, p b In both cases, the priority is equal to 1. However, in the latter case, no learning has been performed on feature A, so false positives may occur in machine learning models selected based on sum. In contrast, machine learning models selected based on product will have different priority values, allowing for the selection of a machine learning model suitable for each feature.

[0018] Let's explain another example of how priority is calculated. For example, Taylor's life equation, which describes the life of a tool, is known to be expressed as shown in equation (5) below, using the cutting speed V, life time T, and multipliers n and C determined by the relationship between the material. VT n =C ··(5) From the above, the lifespan can be expressed as shown in equation (6) below. linT = lin(C / V) -n (6) In the above case, if one variable is a multiplier for another variable, the priority P is calculated by using the relative frequency as the multiplier, as shown in equation (7) below. P = (pc(C) × pv(V)) pn(n) (7)

[0019] Similarly, a priority ranking may be calculated and used for machine learning models that make judgments regarding theoretical surface roughness, wear amount, chip discharge amount, etc. Furthermore, when determining the priority of the conditions for which the relative frequency is multiplied, it is also possible to include tool specifications such as coating and L / D, lubrication conditions, friction coefficient, material geometry values ​​such as the second moment of area, machine structure, and cutting point temperature, which are not clearly expressed in the theoretical formula, and multiply them by the relative frequency.

[0020] In the above configuration, the machine learning model storage unit and the condition range information storage unit were described using common folder and file formats, but any format that can store machine learning models and condition range information may be used. Furthermore, instead of selecting a machine learning model based solely on the current operation of the machine tool, it may be beneficial to anticipate planned operations and select a machine learning model in advance to facilitate smooth machine control. In the above configuration, the procedure described is to select a machine learning model using the selection instruction unit, and then to perform judgment on machine information and action decisions based on those judgments in a continuous sequence. However, the procedure for selecting a machine learning model and the procedure for judgment and action decisions based on the selected machine learning model may be performed separately. In the above configuration, the machine learning model selection device is mounted on the NC device of the machine tool, but the selection device may be provided separately and independently from the NC device. [Explanation of symbols]

[0021] 100...Machine tool body, 101...NC device, 110...Machine learning model control unit, 111...Condition information acquisition unit, 112...Machine learning model storage unit, 113...Priority calculation unit, 114...Selection instruction unit, 120...File group, 121...Condition range information storage unit, 130...Prediction execution unit, 140...Operation determination unit.

Claims

1. A method for selecting a machine learning model from a plurality of trained machine learning models prepared for making judgments on machine tool information, for making judgments on newly input machine tool information, A condition range information storage step involves creating and storing condition range information for each machine learning model, which shows the proportion of each of several predetermined condition items related to the judgment in the data set used to train the machine learning model, as a relative frequency distribution. A condition information acquisition step involves acquiring condition information that indicates feature quantities related to the machine information newly input from the current operation or planned operation of the machine tool, A priority calculation step in which, based on the condition range information and the condition information, a priority, which is a parameter for selecting the machine learning model for the newly input machine information, is calculated by extracting the relative frequency of the value of the condition information from the relative frequency distribution for each of the condition items and multiplying them; A selection instruction step that compares the priorities among the plurality of machine learning models and specifies one or more of the machine learning models, A method for selecting a machine learning model, characterized by performing the following.

2. The method for selecting a machine learning model according to claim 1, characterized in that, in the priority calculation step, when multiplying the relative frequencies for each of the condition items, the weight of each relative frequency is determined from a physical model related to machining acquired by the machine tool, and the priority is calculated.

3. A device for selecting a machine learning model from a plurality of trained machine learning models prepared for making judgments on machine tool information, for making judgments on newly input machine tool information, Condition range information holding means for each machine learning model, which creates and holds condition range information for each machine learning model, which shows the proportion of each of a plurality of predetermined condition items related to the judgment in the data set used to train the machine learning model as a relative frequency distribution, Condition information acquisition means for acquiring condition information that indicates feature quantities related to the machine information newly input from the current operation or planned operation of the machine tool, Priority calculation means calculates a priority, which is a parameter for selecting the machine learning model for the newly input machine information, based on the condition range information and the condition information, by extracting the relative frequency of the value of the condition information from the relative frequency distribution for each condition item and multiplying it; A selection instruction means for specifying one or more machine learning models by comparing the priorities among the plurality of machine learning models, A machine learning model selection device characterized by comprising the following features.

Images