JP7225626B2 - Learning model generation device, estimation device and operation command data update device for grinding - Google Patents

Description

TECHNICAL FIELD The present invention relates to a learning model generating device, an estimating device, and an operation command data updating device related to grinding.

2. Description of the Related Art When grinding a workpiece with a grinding wheel on a grinder, it is required that the grinding quality of the workpiece satisfies a predetermined condition. For example, it is required to prevent the formation of a process-affected layer on the workpiece, to ensure that the surface properties (for example, surface roughness) of the workpiece are within a predetermined value, and to prevent the occurrence of chatter patterns on the workpiece. be done.

By inspecting the workpiece after grinding, an operator confirms whether or not the grinding quality satisfies a predetermined condition, and if the predetermined condition is satisfied, the workpiece is regarded as a non-defective product. Further, Patent Document 1 describes determining whether or not a work-affected layer is generated in a workpiece based on the grinding load measured during grinding.

Further, in grinding a workpiece with a grinding wheel in a grinder, truing and dressing of the surface of the grinding wheel are performed in order to maintain the sharpness of the grinding wheel. If the sharpness of the grinding wheel deteriorates, the quality of the workpiece may deteriorate. Therefore, truing and dressing are performed each time the number of ground workpieces reaches a predetermined number, and the predetermined number is determined so as not to deteriorate the quality of the workpiece. However, since it is determined by the operator, there is a risk that grinding will continue even though the sharpness has decreased, resulting in a decrease in the quality of the workpiece.

Therefore, in Patent Document 2, the vibration of the spindle head is detected by a vibration detector attached to the spindle head, and the amplitude of the spindle reaches a preset value according to the grinding accuracy of the grinding surface of the workpiece. Later, it is described to stop the grinding operation and dress the grinding wheel.

By the way, in recent years, along with the improvement of computer processing speed, artificial intelligence has developed rapidly. For example, Patent Document 3 describes generation of laser processing condition data by machine learning.

JP 2013-129028 A Japanese Patent Application Laid-Open No. 2002-307304 JP 2017-164801 A

However, as described in Patent Document 1, the presence or absence of a work-affected layer cannot be determined with high accuracy only by using the grinding load. This is because there are various factors that cause a work-affected layer. Among the various factors, there are factors that make the measurement itself difficult, unlike those that are easy to measure with sensors or acquire data from equipment. Therefore, it is required to obtain the grinding quality of the workpiece, such as the presence or absence of a work-affected layer, in consideration of various factors. In addition, it is required to obtain grinding conditions such that the grinding quality of the workpiece is good.

Further, as described in Patent Document 2, the sharpness of the grinding wheel cannot be sufficiently determined only by whether or not the vibration of the spindle head reaches the set value. As a result, the timing of grinding wheel modifications (truing and dressing) cannot be properly determined. Therefore, it is required to determine the surface condition of the grinding wheel using not only instantaneous vibration information but also more information.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a learning model generation device capable of acquiring the grinding quality of a workpiece, and an estimation device for estimating the grinding quality of the workpiece. Another object of the present invention is to provide an operation command data update device for updating operation command data so as to improve grinding quality.

Another object of the present invention is to provide a learning model generating device capable of acquiring the surface state of a grinding wheel and an estimating device for estimating the surface state of the grinding wheel. Another object of the present invention is to provide an operation command data update device for updating operation command data so as to improve the surface condition of a grinding wheel.

(1. First learning model generation device)
The first learning model generation device is sampling data when grinding a workpiece with a grinding wheel in a grinder, which is actual operation data related to the drive current or drive position of the drive device of the grinder, the grinder and the sampling data, which is at least one of the first measured data representing the vibration or deformation amount of the structural member and the second measured data related to the dimensions of the workpiece during grinding or the grinding point temperature , for each workpiece a sampling data acquisition unit that acquires a predetermined period of time, and based on the sampling data of the predetermined period of time , the sharpness of the grinding wheel, the dynamic pressure of the coolant supplied to the grinding point, or the static rigidity of the workpiece is measured for grinding. A grinding characteristic calculation unit that calculates a value representing a characteristic, and machine learning using the sampling data for the predetermined period and the value representing the grinding characteristic as first learning input data. a first learning model generation unit for generating a first learning model for estimating the surface texture of the workpiece or the chatter pattern state of the workpiece as the grinding quality of the workpiece.

(2. Second learning model generation device)
The second learning model generation device is sampling data when grinding a workpiece with a grinding wheel in a grinder, which is actual operation data related to the drive current or drive position of the drive device of the grinder, the grinder and the sampling data, which is at least one of the first measured data representing the vibration or deformation amount of the structural member and the second measured data related to the dimensions of the workpiece during grinding or the grinding point temperature , for each workpiece a sampling data acquisition unit that acquires a predetermined period of time, and based on the sampling data of the predetermined period of time , the sharpness of the grinding wheel, the dynamic pressure of the coolant supplied to the grinding point, or the static rigidity of the workpiece is measured for grinding. Grinding characteristic calculation unit that calculates values representing characteristics, and machine learning using the sampling data for the predetermined period and the values representing the grinding characteristics as first learning input data, causing the grinding wheel to become clogged and clogged. or a second learning model generation unit for generating a second learning model for estimating the surface state of the grinding wheel in which the grain is spilled .

(3. First estimation device)
A first estimating device includes a first learning model storage unit that stores the first learning model generated by the first learning model generating device, and grinds a new workpiece of the same type as the workpiece. a grinding quality estimating unit for estimating the grinding quality of the new workpiece using the input data for estimation, which is the sampling data for a predetermined period of time and the first learning model.

(4. Second estimation device)
A second estimating device includes a second learning model storage unit that stores the second learning model generated by the second learning model generating device, and grinds a new workpiece of the same type as the workpiece. Surface state for estimating the surface state of the grinding wheel when the new workpiece is ground using the input data for estimation, which is the sampling data for a predetermined period of time and the second learning model. and an estimating unit.

(5. Operation command data update device)
The motion command data update device includes a motion command data acquisition unit that acquires motion command data for a control device of the grinder for each workpiece when grinding a workpiece with a grinding wheel in the grinder; a remuneration determination unit that determines a remuneration for the operation command data according to the grinding quality of the work piece or the surface condition of the grinding wheel for each work piece; a third learning model storage unit for storing a third learning model for adjusting the motion command data to increase the reward by machine learning; Adjustment of the operation command data using the operation command data, the grinding quality or the surface state estimated by the first estimation device or the second estimation device, the reward, and the third learning model. and an operation command data adjuster for performing the operation.

(6. Effect)
The first and second learning models are generated by machine learning. The first learning input data in machine learning includes not only sampling data for a predetermined period, but also values representing grinding characteristics calculated from the sampling data for the predetermined period. Sampling data for a predetermined period is a data group (collection of many data) and may be affected in various ways. On the other hand, the values representing the grinding characteristics can be organized data based on the sampling data. In addition, it is difficult to directly measure values representing grinding characteristics.

That is, the first and second learning models are generated using the sampling data itself and the arranged values representing the grinding characteristics. In this way, by using the organized values representing the grinding characteristics, the first and second learning models become models that emphasize the relationship with the grinding characteristics. As a result, the estimated grinding quality or surface condition of the grinding wheel is the result of full consideration of the grinding characteristics, resulting in a more accurate result. Furthermore, the grinding characteristics, which are difficult to measure directly, are grasped by calculation from sampling data. By using grinding characteristics, which are difficult to obtain by measurement alone, as learning data, it is possible to obtain higher-precision grinding quality.

Further, the operation command data updating device is processed using the grinding quality or surface condition of the grinding wheel estimated as described above. In other words, the third learning model for adjusting the operation command data is generated using the grinding quality or the surface condition of the grinding wheel as a result of sufficiently considering the grinding characteristics, and the operation command data can be updated. can. Therefore, the operation command can be appropriately updated according to the grinding quality of the workpiece or the surface condition of the grinding wheel.

It is a top view which shows a grinder. 1 is a functional block diagram showing a schematic configuration of a machine learning device; FIG. 3 is a functional block diagram showing the detailed configuration of first and second learning phases of the machine learning device; FIG. 4 is a functional block diagram showing detailed configurations of an estimation phase and an update phase of the machine learning device; FIG.

(1. Configuration of Grinding Machine)
The configuration of the grinder 1 will be described with reference to FIG. The grinder 1 is a machine for grinding a workpiece W. As the grinder 1, grinders having various configurations such as a cylindrical grinder and a cam grinder can be applied. In the present embodiment, the grinding machine 1 is an example of a wheelhead traverse type cylindrical grinding machine. However, the grinding machine 1 can also apply a table traverse type.

The grinding machine 1 mainly includes a bed 11, a headstock 12, a tailstock 13, a traverse base 14, a grinding wheel head 15, a grinding wheel 16, a sizing device 17, a grinding wheel correction device 18, a coolant device 19, and a control device. 20.

The bed 11 is fixed on the installation surface. The headstock 12 is provided on the upper surface of the bed 11 on the front side in the X-axis direction (lower side in FIG. 1) and one end side in the Z-axis direction (left side in FIG. 1). The headstock 12 supports the workpiece W so as to be rotatable around the Z axis. The workpiece W is rotated by driving a motor 12 a provided on the headstock 12 . The tailstock 13 is located on the upper surface of the bed 11 at a position facing the headstock 12 in the Z-axis direction, i.e., on the front side in the X-axis direction (lower side in FIG. 1). That is, the headstock 12 and the tailstock 13 rotatably support the workpiece W at both ends.

The traverse base 14 is provided on the upper surface of the bed 11 so as to be movable in the Z-axis direction. The traverse base 14 is moved by driving a motor 14 a provided on the bed 11 . The wheelhead 15 is provided on the upper surface of the traverse base 14 so as to be movable in the X-axis direction. The wheel head 15 is moved by driving a motor 15 a provided on the traverse base 14 . The grinding wheel 16 is rotatably supported on the grinding wheel head 15 . The grinding wheel 16 is rotated by driving a motor 16 a provided on the grinding wheel head 15 . The grinding wheel 16 is configured by fixing a plurality of abrasive grains with a bond material.

The sizing device 17 measures the dimension (diameter) of the workpiece W. As shown in FIG. The grinding wheel modifying device 18 modifies the shape of the grinding wheel 16 . The grinding wheel correction device 18 is a device that performs either truing or dressing of the grinding wheel 16 . Further, the grinding wheel correction device 18 also has the function of measuring the dimension (diameter) of the grinding wheel 16 .

Here, the truing is a reshaping work, such as a work of shaping the grinding wheel 16 in accordance with the shape of the workpiece W when the grinding wheel 16 is worn by grinding, a work of removing runout of the grinding wheel 16 due to uneven wear, and the like. is. Dressing is a dressing (sharpening) work, and is a work of adjusting the protrusion amount of abrasive grains and creating a cutting edge of abrasive grains. Dressing is an operation for correcting blindness, clogging, spillage, etc., and is usually performed after truing. Also, truing and dressing may be performed without any particular distinction.

The coolant device 19 supplies coolant to the grinding point of the workpiece W by the grinding wheel 16 . The coolant device 19 cools the collected coolant to a predetermined temperature and supplies it to the grinding point again.

The control device 20 controls each driving device based on an NC program generated based on operation command data such as the shape of the workpiece W, machining conditions, the shape of the grinding wheel 16, coolant supply timing information, and the like. That is, the control device 20 receives operation command data, generates an NC program based on the operation command data, and controls the motors 12a, 14a, 15a, 16a, the coolant device 19, etc. based on the NC program. The workpiece W is ground. In particular, the control device 20 grinds the workpiece W based on the diameter of the workpiece W measured by the sizing device 17 until the workpiece W has a finished shape. Further, the control device 20 corrects (truing and dressing) the grinding wheel 16 by controlling the motors 14a, 15a, 16a, the grinding wheel correction device 18, etc. at the timing of correcting the grinding wheel 16. .

1, the grinder 1 includes various sensors 21, 22, 23 (shown in FIG. 3, etc.), which will be described later. For example, the grinder 1 includes a sensor for detecting actual operation data of each motor and the like, data indicating the state of structural members constituting the grinder 1, a sizing device 17, a grindstone diameter sensor, a temperature sensor, and the like. Details of various sensors and the like will be described later.

(2. Overview of machine learning device 100)
An overview of the machine learning device 100 will be described with reference to FIG. The machine learning device 100 performs (a) to (f) below.
(a) generating a first learning model for estimating the grinding quality of the workpiece W;
(b) estimating the grinding quality of the workpiece W using the first learned model;
(c) generating a second learning model for estimating the surface condition of the grinding wheel 16;
(d) estimating the surface condition of the grinding wheel 16 using the second learning model;
(e) generating a third learning model for adjusting the motion command data of the grinding machine 1 so as to improve the grinding quality and reduce the frequency of modification or replacement of the grinding wheel 16;
(f) using the third learning model to update the operation command data of the grinding machine 1 so as to improve grinding quality and reduce the frequency of repair or replacement of the grinding wheel 16;

The machine learning device 100 can be configured as a device separate from the grinding machine 1, or can be configured as a device incorporated in the control device 20 of the grinding machine 1 or the like. In this embodiment, the machine learning device 100 is connected to the grinder 1 via a network line, and transmits and receives various data.

(2-1. Overview of the first learning phase 101)
The first learning phase 101 corresponding to the above (a) and (c) will be described. The machine learning device 100 comprises elements 101a, 101b, 101c, 101d, 101e that function in a first learning phase 101 that generates a first learning model and a second learning model, as shown in FIG. The machine learning device 100 includes, as elements that function in the first learning phase 101, an element 101a that acquires first learning input data, an element 101b that acquires first teacher data, an element 101c that generates a first learning model, It has an element 101d that acquires the second training data and an element 101e that generates the second learning model.

The first learning input data acquired by the element 101a is input data used for machine learning, and includes operation command data for the control device 20 of the grinding machine 1, and a plurality of types of data for a predetermined period for each workpiece W. They are sampling data and values representing grinding characteristics calculated from a plurality of types of sampling data. The sampling data are, for example, actual operation data, first actual measurement data (data representing the state of the structural member), second actual measurement data (data concerning the ground portion), and the like.

The first teacher data acquired in element 101b is teacher data used for machine learning in supervised learning. The first teaching data is grinding quality data of the workpiece W, such as work-affected layer data of the workpiece W, surface texture data of the workpiece W, chatter pattern data of the workpiece W, and the like.

The first learning model generated in the element 101c is a model for estimating the grinding quality of the workpiece W by performing supervised learning among machine learning based on the first learning input data and the first teacher data. (function). However, the first learned model can also be generated by applying unsupervised learning for the purpose of classifying the grinding quality. However, when supervised learning is applied, grinding quality can be obtained with high accuracy.

The second teacher data acquired in element 101d is teacher data used for machine learning in supervised learning. The second teaching data is data representing the surface condition of the grinding wheel 16 (surface condition data of the grinding wheel 16). The surface condition data of the grinding wheel 16 includes, for example, data on the condition of crushing, clogging, spillage, etc., and data on the condition of over-grinding.

The surface of the grinding wheel 16 affects the grinding quality of the workpiece W. In other words, the surface condition of the grinding wheel 16 is the extent to which the grinding quality of the workpiece W is affected. The surface state of the grinding wheel 16 is, for example, a state where the grain is crushed, clogged, or loose, or excessively sharpened. If the surface condition of the grinding wheel 16 is not good, the grinding quality of the workpiece W may deteriorate. Therefore, it is required to grasp the surface condition of the grinding wheel 16 .

If the surface condition of the grinding wheel 16 is broken, clogged, loose, etc., it is necessary to perform dressing, or perform dressing after shaping by truing. Also, if the surface condition of the grinding wheel 16 is too sharp, it is necessary to perform truing. Dressing is usually performed after truing. When the number of times of truing reaches a predetermined number, or when a predetermined amount of truing is formed, the grinding wheel 16 needs to be replaced.

In order to extend the life of the grinding wheel 16, it is required to reduce the number of truing and dressing operations. In addition, the time required for truing, dressing and changing the grinding wheel 16 increases the grinding cycle time. Naturally, there is a desire to shorten the grinding cycle time. From this point of view as well, it is required to grasp the surface condition of the grinding wheel 16 . Therefore, the element 101d acquires the surface condition data of the grinding wheel 16 as the second teacher data. The surface condition data of the grinding wheel 16 is data indicating the degree of influence on the grinding quality of the workpiece.

The second learning model generated in the element 101e is a model for estimating the surface state of the grinding wheel 16 by performing supervised learning among machine learning based on the first learning input data and the second teacher data. (function). However, the second learned model can also be generated by applying unsupervised learning for the purpose of classifying the surface condition of the grinding wheel 16 . However, when supervised learning is applied, the surface state of the grinding wheel 16 can be obtained with high accuracy.

(2-2. Outline of second learning phase 102)
The second learning phase 102 corresponding to (e) above will be described. The machine learning device 100 comprises elements 102a, 102b, 102c, 102d functioning in a second learning phase 102 to generate a third learning model, as shown in FIG. As elements that function in the second learning phase 102, the machine learning device 100 includes an element 102a that acquires second learning input data, an element 102b that acquires first evaluation result data, and an element 102c that acquires second evaluation result data. , and an element 102d for generating a third learning model.

The second learning input data acquired by the element 102a is input data used for machine learning, and is, for example, operation command data. The first evaluation result data acquired in element 102b is evaluation result data for deriving a reward used for machine learning in reinforcement learning. The first evaluation result data is grinding quality data of the workpiece W, such as work-affected layer data of the workpiece W, surface texture data of the workpiece W, chatter pattern data of the workpiece W, and the like.

The second evaluation result data acquired in element 102c is evaluation result data for deriving a reward used for machine learning in reinforcement learning. The second evaluation result data is surface condition data of the grinding wheel 16 . The third learning model generated in the element 102d is based on the second learning input data, the first evaluation result data, and the second evaluation result data, and performs reinforcement learning of machine learning to generate an operation command for the grinding machine 1. It is a model (function) for adjusting the data.

(2-3. Outline of Estimation Phase 103)
The estimation phase 103 corresponding to the above (b) and (d) will be described. As shown in FIG. 2, the machine learning apparatus 100 includes, as elements that function in the estimation phase 103, an element 103a that acquires input data for estimation, and an element 103b that estimates the grinding quality and determines the quality of the workpiece W. . Furthermore, the machine learning device 100, as elements that function in the estimation phase 103, estimates the surface state of the grinding wheel 16, executes truing of the grinding wheel 16, executes dressing of the grinding wheel 16, and It has an element 103c that determines whether to perform an exchange.

The input data for estimation acquired by the element 103a is the same type of data as the first input data for learning, and is data acquired for a workpiece W different from the workpiece W used for learning (a new workpiece W). is. That is, the input data for estimation includes values representing multiple types of sampling data and grinding characteristics. The element 103b estimates the grinding quality using the input data for estimation and the first learning model, and determines the quality of the workpiece W based on the estimated grinding quality. The first learning model used in element 103 b is the first learning model generated by machine learning in the first learning phase 101 .

The element 103c estimates the surface state of the grinding wheel 16 using the estimation input data and the second learning model, and performs truing of the grinding wheel 16 based on the estimated surface state of the grinding wheel 16. Determines whether the dressing of the wheel 16 is to be performed and the replacement of the grinding wheel 16 is to be performed. The second learning model used in element 103 c is the second learning model generated by machine learning in the first learning phase 101 .

(2-4. Outline of update phase 104)
The update phase 104 corresponding to (f) above will be described. The machine learning device 100 includes an element 104a that acquires update input data and an element 104b that updates the action command data as elements that function in an update phase 104 that updates the action command data. The update input data acquired by the element 104a is the same type of data as the second learning input data, and is data acquired for a workpiece W different from the workpiece W used for learning (new workpiece W). is.

Element 104b updates the motion command data using the update input data, the third learning model, the estimated grinding quality, and the estimated grinding wheel 16 surface condition. The third learning model used in element 104b is the third learning model generated by machine learning in the second learning phase 102. Also, the estimated grinding quality is the grinding quality estimated in the estimation phase 103 . The estimated surface condition of the grinding wheel 16 is the surface condition of the grinding wheel 16 estimated in the estimation phase 103 .

(3. Detailed configuration of first learning phase 101)
A detailed configuration of the first learning phase 101 of the machine learning device 100 will be described with reference to FIG. The configuration of the first learning phase 101 constitutes a learning model generation device for grinding. The configuration of the first learning phase 101 includes a first input data acquisition unit 130, a grinding characteristic calculation unit 140, a teacher data acquisition unit 150, a first learning model generation unit 160, a first learning model storage unit 170, and a first learning model storage unit 170. It comprises a second learning model generation unit 180 and a second learning model storage unit 190 .

Table 1 shows the first learning input data, the first teacher data, and the second teacher data used in the first learning phase 101 .

The first input data acquisition section 130 includes an action command data acquisition section 110 and a sampling data acquisition section 120 . The action command data acquisition unit 110 acquires action command data for the control device 20 as first learning input data for machine learning. As shown in Table 1, the operation command data includes command cutting speed for each process, command positions of moving bodies 14 and 15 at process switching, command rotation speed of grinding wheel 16, command rotation speed of workpiece W, coolant supply information, etc. Here, the workpiece W is ground by a plurality of grinding processes such as coarse grinding, fine grinding, fine grinding, and spark out.

The sampling data acquisition unit 120 acquires a plurality of types of sampling data for a predetermined period regarding a plurality of workpieces W as first learning input data for machine learning. Sampling data constitutes a data group for each workpiece W at a predetermined sampling period. The sampling data acquisition unit 120 includes an actual operation data acquisition unit 121 that acquires actual operation data of the driving device 12a controlled by the control device 20 from the sensor 21, and a first measurement data acquisition unit that acquires first measurement data from the sensor 22. A section 122 and a second measured data acquisition section 123 that acquires second measured data from the sensor 23 .

The actual operation data, as shown in Table 1, includes the drive current of the motor 12a and the like, the actual position of the motor 12a and the like, and the like. The actual operation data acquisition unit 121 acquires actual operation data for each workpiece W for a predetermined period. The predetermined period is, for example, from the beginning of grinding to the end of grinding, from the beginning of rough grinding to the end of rough grinding, and the like. It should be noted that when grinding is in an unsteady state, it is unstable, so it is possible to obtain data only for the steady state.

The first actual measurement data is actual measurement data when the workpiece W is ground by the grinding wheel 16, and includes vibration of the structural member 15 and the like, deformation amount of the structural member 15 and the like. The second actual measurement data is actual measurement data when the workpiece W is ground by the grinding wheel 16, and includes the dimensions (diameter) of the workpiece W, the grinding point temperature, and the like.

The first measured data acquisition unit 122 acquires the first measured data for each workpiece W for a predetermined period. The second measured data acquisition unit 123 also acquires the second measured data for each workpiece W for a predetermined period. The first measured data and the second measured data are acquired for the same predetermined period as the actual operation data described above. As described above, the predetermined period is, for example, from the beginning of grinding to the end of grinding, or from the beginning of rough grinding to the end of rough grinding.

The grinding characteristic calculation section 140 calculates a value representing the grinding characteristic based on the operation command data and sampling data acquired by the first input data acquisition section 130 . In particular, the grinding characteristic calculator 140 calculates a value representing the grinding characteristic based on multiple types of sampling data. The values representing the grinding characteristics, as shown in Table 1, are the first learning input data, similar to the operation command data and the sampling data.

For example, the grinding characteristic calculation section 140 calculates a value representing the grinding characteristic by expressing the relationship between the plurality of types of sampling data for a predetermined period as an approximate relational expression. The approximate relational expression is represented by, for example, two types of parameters, and is a relatively low-order relational expression such as a linear expression, a secondary expression, or a cubic expression. Of course, the approximate relational expression may be represented by three or more types of parameters.

The value representing the grinding characteristic is a differential value, an extreme value, a value at which a predetermined axial component is zero, or the like in an approximate relational expression. For example, when the approximate relational expression is represented by a linear expression using two kinds of sampling data, the value representing the grinding characteristic is the slope (differential value) of the primary approximate relational expression. In other words, the sampling data is a group of data for a predetermined period of time, whereas the value representing the grinding characteristics is a single numerical value. Unlike sampling data, which is a group of data (aggregation of a large number of data), the values representing the grinding characteristics are organized data based on the sampling data.

Specific examples of the values representing the grinding characteristics are the sharpness of the grinding wheel 16, the dynamic pressure of the coolant supplied to the grinding point, the static rigidity of the workpiece W, and the like. Any one of the above three types may be applied as a value representing the grinding characteristics, or all three types may be applied. Furthermore, values other than the above three types can be included as values representing grinding characteristics.

The sharpness of the grinding wheel 16 and the dynamic pressure of the coolant are indicators of the state of the grinding wheel 16 . The sharpness of the grinding wheel 16 is determined from the relationship between the two types of sampling data (data group) in a two-dimensional coordinate system whose parameters are the grinding resistance and the depth of cut per unit time (per rotation of the workpiece W). is the value obtained. For the sharpness of the grinding wheel 16, the volume of the workpiece W removed per unit time (per rotation of the workpiece W) may be used as a parameter in place of the depth of cut.

A parameter similar to the sharpness of the grinding wheel 16 can also be used for the dynamic pressure of the coolant. That is, in a two-dimensional coordinate system whose parameters are the grinding resistance and the depth of cut per unit time (per rotation of the workpiece W), the action of the coolant is determined from the relationship between the two types of sampling data (data group). is the value obtained. For the dynamic pressure of the coolant, the removed volume of the workpiece W per unit time (per rotation of the workpiece W) may be used as a parameter instead of the cutting depth.

The static rigidity of the workpiece W is a value obtained from the relationship between the two types of sampling data (data group) in a two-dimensional coordinate system having parameters of the grinding resistance and the deflection amount of the workpiece W. The deflection amount of the workpiece W can also be obtained from the feeding position of the grinding wheel 16 and the diameter of the workpiece W. FIG. Here, when the workpiece W has a complicated shape such as a crankshaft, the static rigidity of the crankshaft can be obtained in grinding the crankpin.

The teacher data acquisition unit 150 includes a grinding quality data acquisition unit 151 that acquires grinding quality data, and a grinding wheel surface condition data acquisition unit 152 that acquires surface condition data of the grinding wheel 16 .

The grinding quality data acquisition unit 151 acquires the grinding quality data regarding the plurality of workpieces W acquired by the external device 2 as first teacher data for supervised learning. That is, the grinding quality data acquisition unit 151, for example, acquires process-affected layer data (data related to grinding burn, softened layer due to grinding, etc.), surface texture data (data such as surface roughness), chatter pattern data, etc., as the first teacher. Get it as data.

The grinding wheel surface condition data acquisition unit 152 acquires surface condition data of the grinding wheel 16 after grinding for each workpiece W as second teacher data for machine learning. The grinding wheel surface condition data acquisition unit 152 acquires surface condition data of the grinding wheel 16 corresponding to the grinding quality data of the workpiece W acquired by the external device 2 .

The surface condition data of the grinding wheel 16 includes first surface condition data corresponding to the condition of the work-affected layer of the workpiece W, second surface condition data corresponding to the surface condition of the workpiece W, and chatter patterns of the workpiece W. A third surface condition data corresponding to the condition is included. However, the first surface state data may be the work-affected layer data itself, or may be data calculated based on the work-affected layer data. The second surface condition data may be the surface condition data of the workpiece W itself, or may be data calculated based on the surface condition data. The third surface state data may be the chatter pattern data itself, or may be data calculated based on the chatter pattern data.

The first learning model generation unit 160 performs supervised learning to generate a first learning model. Specifically, the first learning model generation unit 160 converts the operation command data and the sampling data acquired by the first input data acquisition unit 130 and the values representing the grinding characteristics calculated by the grinding characteristics calculation unit 140 into the first Acquire as input data for learning. Also, the first learning model generation unit 160 acquires the grinding quality data regarding the plurality of workpieces W acquired by the grinding quality data acquisition unit 151 as first teacher data. Then, the first learning model generation unit 160 generates a first learning model for estimating the grinding quality of the workpiece W by machine learning using the first learning input data and the first teacher data.

That is, the first learning model generation unit 160 uses the operation command data, the actual operation data, the first measured data, the second measured data, and the value representing the grinding characteristic as the first learning input data, and the grinding quality data as the first learning input data. A first learning model is generated by machine learning using one teacher data. The first learning model is a model that indicates the relationship between the first learning input data and the first teacher data.

Here, the actual operation data, the first actual measurement data, and the second actual measurement data as the sampling data are data groups for each workpiece W for a predetermined period. Therefore, even sampling data about one workpiece W is a large amount of data. Furthermore, sampling data for a plurality of workpieces W would result in an extremely large amount of data. However, by using machine learning, the first learning model can be easily generated even with a large number of sampled data for a plurality of workpieces W. FIG. Therefore, as will be described later, the grinding quality of the workpiece W can be obtained by generating the first learning model in consideration of a large number of sampling data that affect the grinding quality of the workpiece W.

However, since the sampling data for the predetermined period is a data group (collection of many data), it may be affected in various ways. Therefore, the first learning input data includes not only sampling data for a predetermined period, but also values representing grinding characteristics calculated from sampling data for a predetermined period. Values representing grinding characteristics are data organized based on sampling data. In addition, it is difficult to directly measure values representing grinding characteristics.

That is, the first learning model is generated using the sampling data itself and the arranged values representing the grinding characteristics. In this way, by using the organized values representing the grinding characteristics, the first learning model becomes a model that emphasizes the relationship with the grinding characteristics. As a result, when estimating the grinding quality, the estimated grinding quality is the result of fully considering the grinding characteristics, resulting in a more accurate result. Furthermore, the grinding characteristics, which are difficult to measure directly, are grasped by calculation from sampling data. By using grinding characteristics, which are difficult to obtain by measurement alone, as learning data, it is possible to obtain higher-precision grinding quality.

Here, as the grinding quality of the workpiece W, the first learning model is used for estimating, for example, the state of the work-affected layer of the workpiece W, the surface texture of the workpiece W, and the chatter pattern state of the workpiece W. is a model. However, the first learning model is not limited to estimating the grinding qualities of all of them, and may estimate the grinding qualities of only some of them. The first learning model generated by first learning model generating section 160 is stored in first learning model storage section 170 .

Moreover, when the predetermined period to be acquired is from the initial stage of grinding to the final stage of grinding, the first learning model is a model that considers all grinding processes. On the other hand, if the predetermined period to be acquired is, for example, from the beginning of rough polishing to the end of rough polishing, the first learning model is a learning model that considers only the rough polishing process. If it is desired to specify the process that affects the grinding quality, the first learning model may be obtained for each process.

The second learning model generation unit 180 performs supervised learning to generate a second learning model. Specifically, the second learning model generation unit 180 converts the operation command data and the sampling data acquired by the first input data acquisition unit 130 and the values representing the grinding characteristics calculated by the grinding characteristics calculation unit 140 into the first Acquire as input data for learning. The second learning model generation unit 180 also acquires the surface condition data of the grinding wheel 16 for each workpiece W acquired by the grinding wheel surface condition data acquisition unit 152 as second teacher data. The second learning model generation unit 180 generates a second learning model for estimating the surface state of the grinding wheel 16 by machine learning using the first learning input data and the second teacher data.

That is, the second learning model generation unit 180 uses the operation command data, the actual operation data, the first actual measurement data, the second actual measurement data, and the values representing the grinding characteristics as the first learning input data, and the grinding wheel surface condition data. is used as second teacher data to generate a second learning model. The second learning model is a model that indicates the relationship between the first learning input data and the second teacher data. By applying machine learning, it is possible to generate a second learning model even if there is a large amount of sampling data. Further, by using the organized values representing the grinding characteristics, the second learning model becomes a model that emphasizes the relationship with the grinding characteristics. As a result, when estimating the surface state of the grinding wheel 16, the estimated surface state is the result of adequate consideration of the grinding characteristics, resulting in a more accurate result.

The second learning model is a model for estimating the extent to which the surface state of the grinding wheel 16 affects the grinding quality of the workpiece. This second learning model is a model for estimating, as the surface condition of the grinding wheel 16, the state in which the grinding wheel 16 is broken, clogged, or missing, and the state in which the grinding wheel 16 is over-dressed. be.

For example, the second learning model includes, as the surface conditions of the grinding wheel 16, a first surface condition corresponding to the condition of the work-affected layer of the workpiece W, a second surface condition corresponding to the surface properties of the workpiece W, and a machining This is a model for estimating the third surface state corresponding to the chatter pattern state of the object W. FIG. However, the second learning model is not limited to estimating all of these surface states, and may estimate only some of these surface states. The second learning model generated by the second learning model generation unit 180 is stored in the second learning model storage unit 190 .

(4. Detailed configuration of the second learning phase 102)
A detailed configuration of the second learning phase 102 of the machine learning device 100 will be described with reference to FIG. The configuration of the second learning phase 102 includes an operation command data acquisition unit 110, a grinding quality data acquisition unit 151, a grinding wheel surface condition data acquisition unit 152, a grinding cycle time calculation unit 210, and a grinding wheel shape information acquisition unit 220. , a reward determination unit 230 , a third learning model generation unit 240 , and a third learning model storage unit 250 .

Table 2 shows the second learning input data, the first evaluation result data, and the second evaluation result data used in the second learning phase 102 .

The motion command data acquisition unit 110 acquires motion command data regarding a plurality of workpieces W as second learning input data for machine learning. Further, the grinding quality data acquisition unit 151 acquires the grinding quality data regarding the plurality of workpieces W as the first evaluation result data of machine learning. The grinding wheel surface condition data acquisition unit 152 acquires surface condition data of the grinding wheel 16 after grinding each workpiece W as second evaluation result data of machine learning. Here, as shown in Table 2, the second learning input data is a large number of data, but it is not necessary to use all the data shown in Table 2, and only a part of the data may be used. .

The grinding cycle time calculator 210 calculates the grinding cycle time per workpiece W. FIG. The grinding cycle time includes the time required for grinding a plurality of workpieces W, the time required for replacing the grinding wheel 16 in the grinding, the time required for dressing the grinding wheel 16 in the grinding, and the time required for dressing the grinding wheel 16 in the grinding. It is a value obtained by dividing the total time required for truing by the number of workpieces W. That is, the less the grinding wheel 16 is replaced, the less the grinding wheel 16 is dressed, and the less the grinding wheel 16 is trued, the shorter the grinding cycle time.

The grinding wheel shape information acquisition unit 220 acquires shape information of the grinding wheel 16 . The grinding wheel shape information acquisition unit 220 acquires the dimension (diameter) of the grinding wheel 16 measured by the grinding wheel correction device 18 . That is, the grinding wheel shape information acquisition unit 220 is acquired when the grinding wheel correction device 18 performs truing or dressing of the grinding wheel 16 . Further, the grinding wheel shape information acquisition section 220 can acquire the dimensional change of the grinding wheel 16 and the deformation of the grinding wheel 16 as the shape information of the grinding wheel 16 .

The remuneration determination unit 230 acquires the operation command data as the second learning input data, the grinding quality data as the first evaluation result data, and the surface condition data of the grinding wheel 16 as the second evaluation result data, A reward for the motion command data is determined according to the grinding quality data and the surface condition data. Here, the reward is a reward for a combination of action command data in reinforcement learning.

In the remuneration determination unit 230, when the grinding quality data corresponding to the operation command data leads to a desired result, a large reward is given to the operation command data, and when an undesired result is derived, the operation Small rewards (including negative rewards) are given for command data.

For example, the remuneration determination unit 230 increases the remuneration when there is no work-affected layer in the work-affected layer data of the workpiece W, and decreases the remuneration when there is a work-affected layer. Further, the remuneration determination unit 230 increases the remuneration when the surface texture data of the workpiece W is equal to or less than the predetermined threshold, and decreases the remuneration when the surface texture data is greater than the predetermined threshold. Furthermore, in the chatter pattern data of the workpiece W, the remuneration determining unit 230 increases the remuneration when there is no chatter pattern, and decreases the remuneration when there is a chatter pattern. Note that the remuneration determination unit 230 may determine remuneration based on all of the work-affected layer data, surface texture data, and chatter pattern data, or may determine the remuneration based on only one of these. good.

Further, in the reward determination unit 230, when the surface state data corresponding to the action command data leads to a desirable result, a large reward is given to the action command data, and when the action command data leads to an undesirable result, the action A small reward is given for command data.

For example, the remuneration determination unit 230 increases the remuneration when there is no work-affected layer corresponding to the first surface state data, and decreases the remuneration when there is a work-affected layer. Further, the remuneration determination unit 230 increases the remuneration when the surface texture data of the workpiece W corresponding to the second surface condition data is equal to or less than the predetermined threshold, and decreases the remuneration when the surface texture is greater than the predetermined threshold. Further, the remuneration determination unit 230 increases the remuneration when there is no chatter pattern corresponding to the third surface state data, and decreases the remuneration when there is a chatter pattern.

Further, the remuneration determination unit 230 acquires the grinding cycle time calculated by the grinding cycle time calculation unit 210, and determines remuneration for the operation command data according to the grinding cycle time. Specifically, the remuneration determination unit 230 increases the remuneration as the grinding cycle time becomes shorter. That is, the remuneration determining unit 230 increases the remuneration as at least one of the time required for exchanging the grinding wheel 16, the time required for dressing the grinding wheel 16, and the time required for truing the grinding wheel 16 is shortened.

Further, the remuneration determination section 230 determines remuneration based on the shape information of the grinding wheel 16 acquired by the grinding wheel shape information acquisition section 220 . Specifically, the remuneration determination unit 230 increases the remuneration as the dimensional change of the grinding wheel 16 becomes smaller and the deformation of the grinding wheel 16 becomes smaller.

The third learning model generation unit 240 generates a third learning model for adjusting the action command data so as to increase the reward by machine learning. In the third learning model generation unit 240, for example, Q-learning, Sarsa, Monte Carlo method, etc. are applied as reinforcement learning.

Here, it is assumed that the operation command data before adjustment is data relating to the first workpiece W, and the operation command data after adjustment is data relating to the second workpiece W. A relationship between the operation command data relating to the first workpiece W and the grinding quality data of the first workpiece W is defined as a first data relationship. A relationship between the operation command data of the second workpiece W and the grinding quality data of the second workpiece W is defined as a second data relationship.

The third learning model is a model representing the mutual relationship between the first data relationship before adjustment and the second data relationship after adjustment. The third learning model generation unit 240 makes the grinding quality data of the second workpiece W after adjustment better than the grinding quality data of the first workpiece W before adjustment, that is, the reward increases. , the method of adjusting the motion command data of the first workpiece W before adjustment to the motion command data of the second workpiece W after adjustment is learned.

However, the adjusted operation command data can be adjusted with respect to the unadjusted operation command data within the limits set in advance. For example, for the command cutting speed, which is one of the adjustable parameters, the command cutting speed after adjustment is limited to within a predetermined ratio (eg ±3%) of the command cutting speed before adjustment. there is The predetermined ratio can be set arbitrarily. The same is true for other adjustable parameters. It is also possible to set adjustable parameters. The generated third learning model is stored in the third learning model storage unit 250 .

In addition, in the third learning model, the surface condition data of the grinding wheel 16 is also used in the same manner as the grinding quality data. That is, the third learning model generation unit 240 determines that the surface condition data of the grinding wheel 16 when grinding the second workpiece W after adjustment is the same as that for grinding the first workpiece W before adjustment. The second workpiece after adjustment from the operation command data of the first workpiece W before adjustment so that it is better than the surface condition data of the grinding wheel 16 when it is performed, that is, so that the reward increases Learn how to adjust W to the operation command data.

Note that the third learning model generation unit 240 can also learn the third learning model in the update phase 104, which will be described later. In this case, the grinding quality data obtained in the estimation phase 103 is used as the grinding quality data, which is the first evaluation result data. Further, the surface condition data of the grinding wheel 16, which is the second evaluation result data, uses the surface condition data obtained in the estimation phase 103. FIG.

(5. Detailed configuration of estimation phase 103)
A detailed configuration of the estimation phase 103 of the machine learning device 100 will be described with reference to FIG. The configuration of the estimation phase 103 corresponds to an estimation device for grinding. The configuration of the estimation phase 103 includes a first input data acquisition unit 130 , a first learning model storage unit 170 , a second learning model storage unit 190 , an estimation unit 310 and a discrimination unit 320 .

The estimating section 310 includes a grinding quality estimating section 311 and a grinding wheel surface state estimating section 312 . The grinding quality estimation unit 311 inputs first input data for a predetermined period during grinding of a new workpiece W and values representing the grinding characteristics calculated by the grinding characteristics calculation unit 140 for estimation. Get it as data. That is, the estimation input data includes operation command data, sampling data, and values representing grinding characteristics. Then, the grinding quality estimating section 311 inputs the estimation input data and estimates the grinding quality of the new workpiece W by using the first learning model stored in the first learning model storage section 170 .

Here, the first learning model is a model that indicates the relationship between the first learning input data and the first teacher data, as described above. The first learning model is a model relating to the condition of the work-affected layer of the workpiece W, the surface texture of the workpiece W, and the chatter pattern condition of the workpiece W as the grinding quality data, which is the first teaching data.

Therefore, the grinding quality estimator 311 estimates the state of the work-affected layer of the workpiece W, the surface texture of the workpiece W, and the chatter pattern state of the workpiece W as the grinding quality. However, the grinding quality estimator 311 may estimate only a part of the grinding quality instead of estimating the entire grinding quality. For example, the grinding quality estimator 311 may estimate only the state of the work-affected layer. In this case, the first learning model is generated as a model that estimates only the state of the work-affected layer.

The first learning model is generated using the sampling data itself and the arranged values representing the grinding characteristics. By using the organized values representing the grinding characteristics, the first learning model becomes a model that emphasizes the relationship with the grinding characteristics. As a result, the estimated grinding quality is a result that fully considers the grinding characteristics, resulting in a more accurate result.

The grinding wheel surface state estimating unit 312 estimates the first input data for a predetermined period during grinding of the new workpiece W and the value representing the grinding characteristic calculated by the grinding characteristic calculating unit 140. Get as input data for That is, the estimation input data includes operation command data, sampling data, and values representing grinding characteristics. Then, the grinding wheel surface state estimation unit 312 inputs the input data for estimation and uses the second learning model stored in the second learning model storage unit 190 to grind the new workpiece W. Estimate the surface condition of the grinding wheel 16 in the case. Here, the second learning model is a model indicating the relationship between the first learning input data and the second teacher data, as described above.

Therefore, the grinding wheel surface condition estimating unit 312 estimates the degree of influence on the grinding quality of the workpiece W as the surface condition of the grinding wheel 16 . For example, the grinding wheel surface state estimating unit 312 selects, as the surface state of the grinding wheel 16, a first surface state corresponding to the state of the work-affected layer of the workpiece W, a second surface state corresponding to the surface property of the workpiece W, Then, a third surface state corresponding to the chatter pattern state of the workpiece W is estimated. However, the grinding wheel surface condition estimator 312 may estimate only a part of the surface condition of the grinding wheel 16 instead of estimating the surface condition of all the grinding wheels 16 . For example, the grinding wheel surface state estimation unit 312 may estimate only the first surface state. In this case, the second learning model is generated as a model that estimates only the first surface state.

The second learning model is generated using the sampling data itself and the arranged values representing the grinding characteristics. In this way, by using the organized values representing the grinding characteristics, the second learning model becomes a model that emphasizes the relationship with the grinding characteristics. As a result, the estimated surface condition of the grinding wheel 16 is the result of fully considering the grinding characteristics, resulting in a more highly accurate result.

The determination unit 320 determines whether the workpiece W is good or bad based on the grinding quality of the workpiece W estimated by the grinding quality estimation unit 311 . For example, when it is determined that a work-affected layer exists in the workpiece W (does not satisfy a predetermined condition) based on the estimated state of the work-affected layer, the determination unit 320 W is determined to be defective. Further, when it is determined that the estimated surface texture does not satisfy the predetermined condition, the determination unit 320 determines that the workpiece W is defective. Further, when it is determined that a chatter pattern exists (does not satisfy a predetermined condition) based on the estimated chatter pattern state, the determining unit 320 determines that the workpiece W is defective. .

On the other hand, the discrimination unit 320 discriminates the workpiece W as a non-defective product when each condition regarding the state of the work-affected layer, the surface texture, and the chatter pattern state is satisfied. By using the first learning model generated by applying machine learning in this way, it is possible to easily determine a plurality of conditions.

Furthermore, based on the surface condition of the grinding wheel 16 estimated by the grinding wheel surface condition estimation unit 312, the determination unit 320 performs truing of the grinding wheel 16, dressing of the grinding wheel 16, and operation of the grinding wheel 16. Determine at least one of the exchange executions. For example, the determination unit 320 determines that a work-affected layer exists on the workpiece W (does not satisfy a predetermined condition) based on the first surface state corresponding to the estimated state of the work-affected layer. If so, it is determined that the grinding wheel 16 should be dressed. Further, when it is determined that the second surface state corresponding to the estimated surface texture does not satisfy the predetermined condition, the determining section 320 determines that the grinding wheel 16 should be trued. Further, if the determination unit 320 determines that the chatter pattern exists (does not satisfy the predetermined condition) based on the third surface state corresponding to the estimated chatter pattern state, the grinding wheel 16 dressing should be performed.

On the other hand, the discriminating section 320 discriminates that the grinding wheel 16 is in a good state when each condition regarding the first surface state, the second surface state and the third surface state is satisfied. In this case, it will be determined that the grinding wheel 16 does not require dressing and truing. By using the second learning model generated by applying machine learning in this way, it is possible to easily determine a plurality of conditions.

(6. Detailed configuration of update phase 104)
A detailed configuration of the update phase 104 of the machine learning device 100 will be described with reference to FIG. The configuration of the update phase 104 corresponds to an operation command data update device for grinding. The configuration of the update phase 104 includes an operation command data acquisition unit 110, a grinding quality estimation unit 311, a grinding wheel surface condition estimation unit 312, a grinding cycle time calculation unit 210, a grinding wheel shape information acquisition unit 220, and a remuneration determination unit. It comprises a unit 230 , a third learning model storage unit 250 and an action command data adjusting unit 330 .

The operation command data acquisition unit 110 acquires each data for grinding the new workpiece W, and the contents are substantially the same as those described in the second learning phase 102 . Also, the grinding cycle time calculation unit 210 and the grinding wheel shape information acquisition unit 220 are substantially the same as those described in the second learning phase 102 . The grinding quality estimator 311 and the grinding wheel surface state estimator 312 are the same as those described in the estimation phase 103 . That is, the grinding quality estimation unit 311 and the grinding wheel surface condition estimation unit 312 estimate the grinding quality of the workpiece W and the surface condition of the grinding wheel 16 in grinding the new workpiece W.

The remuneration determination unit 230 determines remuneration using the motion command data, grinding quality, and surface condition of the grinding wheel 16 acquired in grinding the new workpiece W. That is, the remuneration determination unit 230 determines remuneration for the operation command data according to the grinding quality and the surface condition of the grinding wheel 16 for grinding the new workpiece W. Further, the remuneration determination unit 230 determines a remuneration for operation command data according to the grinding cycle time and the shape information of the grinding wheel 16 . The third learning model storage section 250 stores the third learning model generated by the third learning model generation section 240 as described in the second learning phase 102 .

The motion command data adjustment unit 330 includes motion command data relating to grinding of the new workpiece W, estimated grinding quality of the new workpiece W, and estimated grinding wheel 16 when the new workpiece W is ground. A method for adjusting the action command data is determined using the surface state, reward, and the third learning model, and the action command data is adjusted based on the determined adjustment method. Here, the third learning model is a model generated by learning an adjustment method from the action command data before adjustment to the action command data after adjustment so as to increase the reward.

Specifically, the action command data adjuster 330 acquires the current action command data as action command data before adjustment, and acquires the reward at that time. In this case, the action command data adjuster 330 uses the current action command data, the reward for the current action command data, and the third learning model to determine the action command data to be adjusted. That is, the action command data to be adjusted is action command data that provides a larger reward than the current action command data.

In the processing of the action command data adjustment unit 330, there are cases where a plurality of candidates with equal rewards are output as action command data to be adjusted. In this case, for example, the action command data adjustment unit 330 can rank the plurality of candidates by setting the priority of parameters to be adjusted. For example, the order of priority of the parameters to be adjusted is such that the commanded cutting speed is first and the command rotation speed of the workpiece W is second.

Then, the action command data adjustment unit 330 determines the first candidate as the action command data to be adjusted, and updates the current action command data to the action command data to be adjusted. Then, the grinder 1 grinds the workpiece W based on the updated operation command data. Then, in the update phase 104 of the machine learning device 100, the operation command data for the next grinding is again adjusted based on each data in the grinding of the workpiece W. It is also possible to set the frequency of adjustment of the operation command data. For example, after grinding a set number of workpieces W, the operation command data may be adjusted.

That is, the action command data is updated using the third learning model generated by machine learning of the machine learning device 100 . Therefore, even if the grinding state changes, the operation command data is updated according to the current grinding state. By updating the operation command data in this manner, the grinding quality of the workpiece W can be improved.

Furthermore, by updating the operation command data, it is possible to perform grinding according to the surface condition of the grinding wheel 16 . That is, the surface condition of the grinding wheel 16 is improved by updating the operation command data. By improving the surface condition of the grinding wheel 16, the grinding quality of the workpiece W can be improved. Furthermore, by updating the operation command data, the time required for exchanging the grinding wheel 16, the time required for dressing the grinding wheel 16, and the time required for truing the grinding wheel 16 are reduced. As a result, the grinding cycle time is shortened. Furthermore, by updating the operation command data, the dimensional change of the grinding wheel 16 is reduced, and the deformation of the grinding wheel 16 is reduced.

In particular, in the update phase 104, the grinding quality of the workpiece W and the surface state of the grinding wheel 16 estimated in the estimation phase 103 are used. In other words, the third learning model for adjusting the operation command data is generated using the grinding quality or the surface condition of the grinding wheel as a result of sufficiently considering the grinding characteristics, and the operation command data can be updated. can. Therefore, the operation command can be appropriately updated according to the grinding quality of the workpiece W and the surface condition of the grinding wheel.

(7. Others)
The above-described machine learning device 100 generates a first learning model, generates a second learning model, estimates the grinding quality of the workpiece W, determines the quality of the workpiece W, estimates the surface condition of the grinding wheel 16, A case has been described in which determination of execution of truing No. 16, determination of execution of dressing of the grinding wheel 16, determination of execution of replacement of the grinding wheel 16, and updating of operation command data are all performed. In addition, the machine learning device 100 can also be applied as a device that performs only one element of the above processing. In that case, machine learning device 100 only has a configuration corresponding to the element.

1: grinding machine, 2: external device, 12: headstock (structural member), 12a, 14a, 15a, 16a: motor (driving device), 13: tailstock (structural member), 14: traverse base (structural member , moving body), 15: wheelhead (structural member, moving body), 16: grinding wheel, 17: sizing device (driving device), 18: grinding wheel correcting device (driving device), 19: coolant device, 20: Control device, 21, 22, 23: sensor, 100: machine learning device, 101: first learning phase (learning model generation device), 102: second learning phase (learning model generation device), 103: estimation phase (estimation device ), 104: Update phase (update device), 110: Operation command data acquisition unit, 120: Sampling data acquisition unit, 121: Actual operation data acquisition unit, 122: First measured data acquisition unit, 123: Second measured data acquisition Section 130: First input data acquisition section 140: Grinding characteristics calculation section 150: Teacher data acquisition section 151: Grinding quality data acquisition section 152: Grinding wheel surface condition data acquisition section 160: First learning model generation Section 170: First learning model storage section 180: Second learning model generation section 190: Second learning model storage section 210: Grinding cycle time calculation section 220: Grinding wheel shape information acquisition section 230: Reward determination Section 240: Third learning model generating section 250: Third learning model storage section 310: Estimating section 311: Grinding quality estimating section 312: Grinding wheel surface state estimating section 320: Discriminating section 330: Operation command Data adjustment unit, W: workpiece

Claims (8)

Sampling data when grinding a workpiece with a grinding wheel in a grinder, which includes actual operation data related to the drive current or drive position of the drive device of the grinder, and the amount of vibration or deformation of the structural members of the grinder. A sampling data acquisition unit that acquires the sampling data, which is at least one of the first measured data representing the size of the workpiece and the second measured data related to the dimensions of the workpiece or the grinding point temperature during grinding, for a predetermined period of time for each workpiece. and,
A grinding characteristic calculation unit that calculates the sharpness of the grinding wheel, the dynamic pressure of the coolant supplied to the grinding point, or the static rigidity of the workpiece as a value representing the grinding characteristic based on the sampling data for the predetermined period. and,
By machine learning using the sampling data for the predetermined period and the value representing the grinding characteristic as first learning input data, the work-affected layer of the workpiece, the surface properties of the workpiece, or the chatter pattern of the workpiece. a first learning model generation unit that generates a first learning model for estimating a state as the grinding quality of the workpiece;
A learning model generation device for grinding. Sampling data when grinding a workpiece with a grinding wheel in a grinder, which includes actual operation data related to the drive current or drive position of the drive device of the grinder, and the amount of vibration or deformation of the structural members of the grinder. A sampling data acquisition unit that acquires the sampling data, which is at least one of the first measured data representing the size of the workpiece and the second measured data related to the dimensions of the workpiece or the grinding point temperature during grinding, for a predetermined period of time for each workpiece. and,
A grinding characteristic calculation unit that calculates the sharpness of the grinding wheel, the dynamic pressure of the coolant supplied to the grinding point, or the static rigidity of the workpiece as a value representing the grinding characteristic based on the sampling data for the predetermined period. and,
The grinding wheel is in a state in which the grinding wheel is in a state where the grinding wheel is clogged, clogged, or missing due to machine learning using the sampling data for the predetermined period and the value representing the grinding characteristic as the first learning input data. a second learning model generation unit that generates a second learning model for estimating the surface state of the vehicle;
A learning model generation device for grinding. The sampling data acquisition unit acquires data on grinding resistance and data on depth of cut per unit time from the actual operation data, the first measured data and the second measured data,
Based on the data on the grinding resistance and the data on the depth of cut per unit time, the grinding characteristic calculation unit calculates the sharpness of the grinding wheel or the amount of coolant supplied to the grinding point as a value representing the grinding characteristic. 3. The learning model generation device for grinding according to claim 1, which calculates dynamic pressure. The sampling data acquisition unit acquires data on grinding resistance and data on the deflection amount of the workpiece from the actual operation data, the first measured data and the second measured data,
3. The grinding characteristic calculation unit according to claim 1, wherein the grinding characteristic calculation unit calculates the static rigidity of the workpiece as a value representing the grinding characteristic based on the data regarding the grinding resistance and the data regarding the deflection amount of the workpiece. A learning model generator for grinding. 5. The grinding characteristic calculator according to any one of claims 1 to 4, wherein the grinding characteristic calculation unit calculates the value representing the grinding characteristic by expressing the relationship of the plurality of types of sampling data for the predetermined period with an approximate relational expression. A learning model generation device for the described grinding process. a first learning model storage unit that stores the first learning model generated by the learning model generation device according to claim 1;
The grinding of the new workpiece using the input data for estimation which is the sampling data for a predetermined period and the first learning model when grinding the new workpiece of the same type as the workpiece. a grinding quality estimation unit for estimating quality;
An estimating device for grinding, comprising: a second learning model storage unit that stores the second learning model generated by the learning model generating device according to claim 2;
Grinding of the new workpiece is performed using the input data for estimation which is the sampling data for a predetermined period and the second learning model when grinding the new workpiece of the same type as the workpiece. a surface condition estimation unit for estimating the surface condition of the grinding wheel when the
An estimating device for grinding, comprising: an operation command data acquiring unit for acquiring operation command data for a control device of the grinder for each workpiece when grinding the workpiece with a grinding wheel in the grinder;
a remuneration determination unit that determines a remuneration for the operation command data according to the grinding quality of the workpiece or the surface condition of the grinding wheel for each workpiece;
a third learning model storage unit for storing a third learning model for adjusting the motion command data so as to increase the reward by machine learning using the motion command data and the reward for the plurality of workpieces; ,
Said motion command data relating to grinding of a new workpiece of the same type as said workpiece , said grinding quality or said surface condition estimated by said estimation device according to claim 6 or 7 , said reward, said third learning a motion command data adjustment unit that adjusts the motion command data using a model;
An operation command data update device for grinding.

Images