JPH04135209A - Learning method for machining condition producing function of numerical controller - Google Patents

Abstract

PURPOSE:To learn a neural network without loading on the numerical control by using a numerical controller different from the one which performs the machining control based on the accumulated input data and teacher data in order to learn the coupling coefficient of the neural network with use of an off-line method. CONSTITUTION:A CPU 40 of a learning device inputs the learnt coupling coefficient data on the 1st and 2nd neural networks through a flexible disk FD 37 and stores the input coupling coefficient data in a coupling coefficient area 462 of a RAM 46. At the same time, the CPU 40 inputs and stores the teacher data in a teacher data area 463. Furthermore, the CPU 40 inputs the input data through the FD 37 to store it in an input data area 464 and also stores an arithmetic program in a neural network area 461. A numerical controller 30 learns a coupling coefficient by means of the high speed computing ability of an off-line learning device based on the accumulated teacher data in an actual application state.As a result, the precise learning of the neural networks can be attained without loading on the machining control of the controller 30.

Description

[Detailed description of the invention] [Industrial application field]

The present invention relates to a learning method for an automatic creation function of machining condition data of a numerical control device.

[Prior art]

Conventionally, for example, from fixed condition data such as the finished diameter of the workpiece, machining allowance, rigidity coefficient, and part type, the rotation speed of the workpiece, grinding start position, and grinding feed for each grinding mode such as rough grinding, fine grinding, and machine string has been calculated. Numerical control devices are known that automatically determine machining conditions such as speed and post-grinding feed stop time.

[Problem to be solved by the invention]

However, in reality, it is known from experience that the automatically determined machining condition data is not optimal for various fixed condition data. In such cases, when the operator looks at the automatically determined machining condition data and finds that it is not appropriate, the operator may correct the value based on the operator's experience or intuition, or perform the actual machining using the machining condition data. The operator judged the resulting machining results and corrected the machining condition data based on experience and intuition. How to correct which machining condition data from fixed condition data, or how to correct which machining condition data from the machining error obtained from the machining result when machining a workpiece with certain machining condition data. The correct correction depends on the experience and intuition of the operator, and the correct correction is not always performed. For this reason, it was necessary to machine the workpiece using the corrected machining condition data and to correct the machining condition data again according to the machining results. Furthermore, even if the operator corrects the machining condition data correctly, there is a problem in that the causal relationship between the machining condition data at that time and the fixed condition data or machining error data is not reflected in the subsequent correction of the machining condition data. Ta. For this reason, it is reported that it takes time to create appropriate processing condition data. The present invention has been made to solve the above-mentioned problems, and its purpose is to reduce the load on the numerical control device in order to automatically create subsequent machining condition data based on the state of modification of machining conditions by the operator. The goal is to do it efficiently without spending too much time.

[Means to solve the problem]

The structure of the invention for solving the above problem is that, in a learning method for an automatic creation function of machining condition data for controlling machining conditions in a numerical control device, machining factors that affect machining condition data during actual machining are used. Input data, calculate the correction amount of the processing condition data using a neural network, evaluate the correction amount that is the output of the neural network, and when the correction amount is corrected as a result of the evaluation, the correction value is The training data is stored in a storage device together with the input data of the neural network at that time, and the coupling coefficients of the neural network are calculated offline using a device separate from the numerical control device that performs processing control according to the stored input data and training data. It's about learning.

[Effect]

The amount of correction for machining condition data is input from data related to machining factors that affect machining condition data (for example, fixed condition data such as diameter and machining allowance for workpiece finishing, and machining error data obtained from machining results). Calculated by neural network. Since neural networks are trained to obtain appropriate outputs for various inputs, it is also possible to correct processing condition data caused by causal relationships that cannot be analyzed theoretically. On the other hand, since it is impossible to completely perform the initial learning of a neural network, when various actual processing is performed, the correction amount output by the neural network is not always appropriate. In such a case, the correction amount is corrected, and the corrected value is stored in the storage device as teacher data together with the input data of the neural network at that time. These data are processed in various ways and accumulated each time the amount of correction output by the neural network is not appropriate. Then, using this accumulated data, a neural network is trained off-line using a device different from the numerical control device. By doing this, learning of the neural network becomes possible without imposing a load on numerical control.

【Example】

The present invention will be described below based on specific examples. (1) Structure of Grinding Machine FIG. 1 is a block diagram showing the entire mechanical structure of a numerically controlled grinding machine having an automatic processing condition determination device according to the present invention. 50 is a grinding machine, and a table 52 that slides on the bed 51 of the grinding machine 50 is provided. The table 52 is moved in the left-right direction in the drawing by driving the table feed motor 53. Also, on the table 52 are a headstock 54 and a tailstock 56.
The headstock 54 has a main shaft 55 and the tailstock 56 has a psychological axis 57. The workpiece W is supported by a main shaft 55 and a psychological shaft 57,
It is rotated by the rotation of the main shaft 55. This rotation of the main shaft 55 is performed by a main shaft motor 59 disposed on the head stock 54. On the other hand, a grinding wheel 60 for grinding the workpiece W is pivotally supported by a drive shaft of a grinding wheel drive motor 62 provided on a grinding wheel head 61. Further, the grindstone head 61 is controlled to move in the vertical direction of the drawing by a grindstone head feed motor 63. Table feed motor 53, grindstone feed motor 63,
A numerical control device 30 is provided to drive and control the main shaft motor 59, grinding wheel drive motor 62, and the like. (2) Configuration of numerical control device The numerical control device 30 mainly consists of
It is composed of a CPU 31, a ROM 32, a RAM 33, and an IF (interface) 34. The RAM 33 has an NC data area 331 for storing NC programs, a fixed condition data area 332 for storing fixed condition data such as finishing diameter, machining allowance, rigidity coefficient, part type, rough finishing division, operator name, etc., and workpiece rotation speed, A machining condition data area 333 that stores machining condition data such as a grinding start position, a grinding feed rate, and a feed stop time after grinding, a neural network area 334 that stores a neural network calculation program, and a coupling coefficient that stores a neural network coupling coefficient. A region 335 is provided. In addition, the RAM 33 includes a required value area 336 that stores required values of machining results determined from required machining specifications, a reference value area 337 that stores standard values of machining condition data determined by self-study, and a neural network. A correction amount area 338 for storing a correction amount of the calculated machining condition data, a machining result data area 339 for storing machining result data, and a machining error data area 340 for storing machining error data are formed. Furthermore, when the correction amount of the output of the neural network is corrected, the RAM 33 stores the corrected value, that is, the teacher data area 341 that stores the teacher data, and the input data of the neural network at that time in correspondence with the teacher data. An input data area 342 for storage is formed. Note that the RAM 33 is backed up by a battery, and the learned coupling coefficients are stored therein. Further, the ROM 32 is formed with a control program area 321 storing a control program for operating the numerically controlled grinding machine according to NC data, and an automatic determination program area 322 storing a main automatic determination program for determining machining condition data. ing. In addition, the numerical control device 30 is connected to the operation panel 20 via the IF 54.
is installed. Operation panel 2 of the operation panel 20
1, there are provided a keyboard 22 for inputting data and a CRT display device 23 for displaying data. Further, an FD drive device 36 is connected to the numerical control device 30 via an IF 55, and an FD (flexible disk) 37, which is a storage medium, is read and written via the FD drive device 36. (3) Configuration of offline learning device The offline learning device is configured with a computer system as shown in FIG. An FD drive device 41 is connected to the CPU 40, and the FD
It is configured such that input/output to and from the 37 is possible. In addition, the CPU 40 is provided with C through an input/output interface 42.
An RT 43 and a keyboard 44 are connected. Further, a ROM 45 and a RAM 46 are connected to the CPU 40. A learning program area 451 is formed in the ROM 45 in which a learning program for learning the neural network is stored. Further, the RAM 46 includes a neural network area 461 for storing data regarding the structure of the neural network, a coupling coefficient area 462 for storing its coupling coefficient data, a teacher data area 463 for storing teacher data, and an input data area 464 for storing input data. is formed. (4) Operation Next, the processing procedure of the CPU 31 used in the device of this embodiment will be explained based on a flowchart. 1. General operation of the entire apparatus The general operation of the apparatus will be explained with reference to FIG. 13. In step S1, a first neural network calculation program used to actually obtain machining condition data is created from the FD37a that stores neural network calculation programs corresponding to many types of machining and their learned coupling coefficient data. and its coupling coefficient data are stored in RAM3.
3. Hereinafter, the calculation program of the first neural network will also be referred to as a condition matching engine. In step S2, fixed condition data is input to the RAM 33. In step S3, the reference value of the machining condition data is automatically determined from the fixed condition data. In step S4, the correction amount of the machining condition data is calculated by the calculation program (condition adaptation engine) of the first neural network that receives the fixed condition data as input. In step S5, the processing condition data corrected by the corrected quantum reference value is displayed on the CRT 23. In step S6, the machining condition data is judged based on the operator's experience and intuition to determine whether it is appropriate. If it is not appropriate, the processing condition data is corrected in step S7. In step S8, the correction value of the correction amount at this time is used as teacher data, and fixed condition data (input data) that is the input of the first neural network at that time.
saved with. If the machining condition data is determined to be appropriate in step S6, the workpiece is actually machined using the machining condition data in step S9. In step SIO, machining results such as finished dimensions of the workpiece are measured, and machining error data indicating an error from the required value is calculated. In step Sll, the machining error data is evaluated, and if appropriate, machining is completed in step S12. Furthermore, if the machining results are not appropriate, it is necessary to correct the machining condition data. In step S13, the above machining error data is stored in RAM3.
3 is input. In step S14, the second neural network calculation program corresponding to the processing type and its learned coupling coefficient are input to the RAM 33 from the FD 37b. The calculation program of the second neural network is also called a machining diagnosis engine. In step S15, the second o neural network which uses the machining error data as input data is activated, and a correction amount of the machining condition data is calculated. In step S1.6, new machining condition data is calculated using concave machining condition data + correction amount, and the influence of the new machining condition data on the machining result is predicted. Proceeding to step S6, the modified machining condition data is evaluated based on the predicted machining results. If the machining condition data is not appropriate, the machining condition data is corrected by an appropriate correction amount in step S7. Then, in step S8, the appropriate correction amount is stored as teacher data for the second neural network together with the machining error data (input data) at that time. As mentioned above, during actual processing, the first and second neural networks are activated, and each time a situation arises in which the output correction amount is inappropriate, the teacher data and input data at that time (these data are hereinafter referred to as " ``learning data'') is accumulated. This kind of learning data is used in step S1? In,
Output to offline learning device. In the offline learning device, during actual machining, learning is performed for the condition matching engine in step 318 and for the machining diagnosis engine in step S19, based on the respective accumulated learning data. Also, step S2
0. In S21, learning is also performed by specifying input data and teaching data through interaction with the user. The coupling coefficient data thus learned is updated and stored in the FD together with the neural network calculation program. In addition, in the above-mentioned learning of the first neural network, learning is performed in consideration of the amount of correction by the second neural network in the corrected machining condition data. That is, the coupling coefficient of the first neural network is also learned using the deviation of the machining condition data corrected by the second neural network from the reference value as teacher data. Therefore, since the amount of correction by the second neural network is also taken into account in the coupling coefficient of the first neural network, it is not possible to obtain optimal machining condition data using only the first neural network while learning is progressing. It becomes possible. That is, the machining error data gradually becomes smaller, and the second
The amount of correction by the neural network also gradually decreases. In addition, the coupling coefficient data of the first and second neural networks learned as described above during actual processing is F.
The data is output to D37, and the corresponding coupling coefficient data stored in FD37 is updated. As described above, in this embodiment, the machining condition data is corrected by the condition matching engine and the machining diagnosis engine. FIG. 4 shows a flowchart of a program that inputs a predetermined neural network calculation program and coupling coefficient data from the FD 37 of the storage medium and stores them in the RAM 33. The FD 37 stores a large number of first and second neural network calculation programs and learned coupling coefficients corresponding to the type of processing (for example, grinding of an engine camshaft, etc.). In step 700, a list of all neural network names is displayed on the CRT 23. In step 702, the operator looks at the list and selects the first and second neural networks (condition matching engine and processing diagnosis engine) corresponding to the type of processing to be performed. Next, in step 704, the calculation programs of the two selected neural networks and their coupling coefficient data are transferred from the FD 37 to the neural network area 334 and the coupling coefficient area 335 of the RAM 33, respectively.
is stored in a predetermined area. 3. Input of fixed condition data FIG. 5 is a flowchart of the main program that automatically generates machining condition data from fixed condition data. At step 100, fixed condition data input from the keyboard 22 is read and stored in the fixed condition data area 332 of the RAM 33. In this embodiment, the fixed condition data includes finishing diameter data row 5. Removal allowance data Da,
Rigidity coefficient data Os, parts type data 1. Rough finishing division data Ds, A worker name data D, B worker name data D, C worker name data D. 4. In step 102 of the calculation method for the reference value of machining condition data, the above fixed condition data (0 to D
, ) and other fixed condition data, the reference values v1 to V of the machining condition data are calculated. In this embodiment, the machining condition data includes coarse grinding rotation speed data of 1. Seiken rotation speed data 2. Equipment rotation speed data K
s, rough grinding start diameter data 1. Seiken starting diameter data 2. Equipment starting diameter data Ks, rough grinding feed speed data? , Seiken feed speed data 8. 9. Input the device feed speed data. The feed stop time data after coarse grinding Kl (1 + feed stop time data after fine grind K11) is calculated. This reference value is calculated as follows. Regarding the rotation speed data for each grinding mode, The circumferential speed of the workpiece is determined in advance as a function of the required surface roughness of the workpiece relative to the circumferential speed of the grinding wheel during a certain grinding process. The speed is calculated, and the rotation speed data of the workpiece is calculated based on the circumferential speed of the workpiece and the diameter of the workpiece.For the feed speed data of each grinding mode, the cut per rotation of the workpiece of the grinding wheel is calculated. The depth of cut is determined in advance as a function of the dimensional tolerance.The depth of cut is calculated from the dimensional tolerance commanded for each grinding mode, and the grinding feed rate data is calculated from the rotation speed data.Grinding for each grinding mode The starting diameter data has a standard feed amount set for each grinding mode, and is calculated based on the relationship between the commanded finishing diameter and this feed amount.The data on the feed stop time after grinding is determined by the standard feed amount for each grinding mode. The number of revolutions of the workpiece at which the feed is to be stopped is determined depending on whether or not the process is divided into parts for grinding.The number of revolutions at which the feed is stopped is determined from the fixed condition data, and the number of revolutions at which the feed is stopped is determined by using the number of revolutions data. Then, the post-grinding feed stop time is calculated.Next, in step 104 of FIG. 5, the first neural network is started and the fixed condition data D1 to D are calculated.
8 is input, and each correction amount 61 to δ of each processing condition data (K, . . . , . . . ) is calculated. The first neural network has the configuration shown in FIG. In this example, the neural network is
It has a three-layer structure: an input layer, an intermediate layer, and an output layer. As is well known in the neural network, each element of the second intermediate layer and the third output layer is defined as an element that performs the following calculation. The output of the third element in the first layer is 0. is calculated using the following formula. However, I≧2. 0, = f (1,) f (x) = 1/ (1+exp (-X) )
-(3) where V is the coupling coefficient between the j-th element of the first layer and the j-th element of the first layer, 0. represents the output value of the ``th'' element of the first layer. In other words, since it is the first layer, there is no calculation required.
Since the input is output as is, the j-th input layer (first layer)
It is also the input value of the th element. Therefore, 0,=D,
-(4) However, D is fixed condition data input to the third element of the input layer. The specific calculations of the neural network are executed according to the procedure shown in FIG. In step 200, each element of the intermediate layer (second layer) receives an output value from each element of the input layer (first layer), ~D,
This becomes the input value on which the product-sum operation of the following equation is performed. Next, in step 204, a product-sum operation of the input values of each element of the output layer (third layer) is performed. Next, in step 206, the output value of each element in the output layer is calculated using a sigmoid function, similar to equation (6). This output value becomes the correction amount δ of the machining condition data. That is, the correction amount δ is obtained by the following equation. The third element of the second layer is calculated using the following equation. In this embodiment, the bias is zero. Next, in step 202, the output of each element of the intermediate layer (second layer) is calculated using the sigmoid function of the product-sum function value of the input values of equation (5) using the following equation. jth layer of second layer
The output value of the th element is calculated using the following equation. This output value is 0. returns to step 106 in FIG. 5 for each element of the output layer (third layer), and calculates the reference values v1 to Vl+ of the processing condition data obtained in step 102 and the correction amounts δ, to δ11 obtained in step 104. By the sum of , - is calculated for the machining condition data. The machining condition data 1 to Kll are stored in the machining condition data area 333 of the RAM 33. Next, in step 108, ~Kl is added to the machining condition data.
1 is displayed on the CRT 23. In step 110, the operator looks at the displayed results and determines that the data needs to be corrected.In step 112, the operator inputs the correction value from the keyboard 22, and the corrected machining condition data is stored in the machining condition data area 333. be remembered. Next, in step 114, an appropriate correction amount is determined based on the correction value, and the correction amount is stored in the teacher data area 341 as teacher data. Next, in step 116, the fixed condition data that is the input of the first neural network corresponding to the teacher data at that time is stored in the input data area 342 in correspondence with the teacher data. Next, in step 118, NC data is calculated using 1 to K11 in the processing condition data, and the calculated NC data is stored in the NC data area 331 of the RAM 33. 7. Generation of machining error data Next, the workpiece is trial-machined according to this NC data. The machining results of the workpiece are measured and machining result data is obtained. Machining result data includes, for example, total grinding time, finished surface roughness, and finished dimensional accuracy. Finish roundness, degree of cracking, degree of chatter, etc. Next, in step 500 of FIG. 8, which describes a procedure for modifying machining condition data from these measured machining result data, the above-mentioned measured machining result data is inputted by the operator from the keyboard 22, The value is stored in the processing result data area 339 of the RAM 33. Next, in step 502, the required value of the machining result is input in the same way, and the value is stored in the required value area 336 of the RAM 33.
is memorized. Next, in step 504, machining error data, which is the deviation of the machining result data from the required value, is calculated, and these data are stored in the machining error data area 340 of the RAM 33. The machining error data includes a total grinding time error H1, a finishing roughness error H2, a finishing error H5, a finishing roundness error H4, a degree error Hs of scorching or cracking, a degree error H6 of chatter, and the like. 8. Second neural network (machining diagnosis engine) Next, in step 506, machining error data (II, ~ H
, ) to start the second neural network. Then, in step 508, the amount of correction (δ,
~δII) is stored in the correction amount area 338 of the RAM 33. This second neural network has the configuration shown in FIG. In this embodiment, the second neural network has a three-layer structure including an input layer, a middle layer, and an output layer. The arithmetic function of each element of this second neural network is completely the same as the arithmetic function of the first neural network described above. Next, in step 510, each concave machining condition data (K, ~kz
) by adding the correction amount (δ1-δ11) to obtain corrected new machining condition data (K, to 11). These data are stored by rewriting the contents of the machining condition data area 333 of the RAM 33. Further, the processing condition data is displayed on the CRT 23. In step 512, the operator looks at the displayed results and determines that the data needs to be corrected.In step 514, the corrected value is input from the keyboard 22, and the corrected machining condition data is stored in the machining condition data area 333. be remembered. Next, in step 516, an appropriate output (correction amount) of the second neural network is determined based on the correction value, and the output is stored in the teacher data area 34 as teacher data.
1 is stored. Next, in step 518, the machining error data that is the input of the second neural network corresponding to the teacher pig at that time is stored in the input data area 342 in correspondence with the teacher data. Next, in step 520, NC data is calculated using the corrected machining condition data, and the calculated NC data is stored in the NC data area 331 of the RAM 33.
is memorized. As described above, in this embodiment, the reference value of the machining condition data is corrected by the first neural network, and further, the second neural network is used to perform correction taking into account machining errors. Therefore, more accurate machining condition data is required. 99. Outputting learning data to the learning device As shown in step 800 in FIG. 10, the learning data stored in the RAM 33 is output to the FD 37. That is, the accumulated teacher data for the first and second neural networks stored in the teacher data area 341 is output to the FD 37. Further, the accumulated fixed condition data, which is the input data of the first neural network, stored in the input data area, and the accumulated processing error data, which is the input data of the second neural network, are output to the FD 37. Ru. Note that this FD 37 has a coupling coefficient area 33 of the RAM 33.
Coupling coefficient data of the first and second neural networks stored in 5 is stored. As shown in FIG. 11, the CPU 40 of the learning device inputs the learned coupling coefficient data of the first and second neural networks from the FD 37 in step 850, and stores the data in the RAM.
The data is stored in 46 coupling coefficient areas 462. Also, in step 852, the training data for the first and second neural networks is input from the FD 37, and R
The data is stored in the teacher data area 463 of AM46. Also, in step 854, the input data of the first and second neural networks are input from the FD 37, and
The data is stored in the input data area 464 of AM46. Further, in step 856, the calculation programs for the first and second neural networks are stored in the neural network area 461 from the FD 37. Next, a procedure for learning the coupling coefficients of the first neural network will be described. The coupling coefficients are determined using the well-known backpropagation method for the neural network shown in FIG. At step 600 in FIG. 12, a learning signal for each element in the output layer is calculated using the following equation. However, T is teacher data for the correction amount δ, which is the output, and is provided from the outside. f' (X) is the derivative of the sigmoid function. Next, in step 602, the learning signal of the intermediate layer is calculated using the following equation. Next, in step 604, the coupling coefficients of each element of the output layer are corrected. The amount of correction is determined by the following formula. The amount of change in the coupling coefficient between the child and the #1 element of the intermediate layer is the amount of change in the th calculation. Further, Δω+, i (t-1) is the previous correction amount of the coupling coefficient. P,0 is a proportionality constant. Therefore, the corrected coupling coefficient can be obtained as follows. Next, the process moves to step 606, and the coupling coefficients of the elements in the intermediate layer are corrected. The amount of correction of the coupling coefficient is determined by the following equation, as in the case of the output layer. Therefore, the corrected coupling coefficient can be obtained as follows. Next, in step 608, it is determined whether the correction amount of the coupling coefficient has become less than or equal to a predetermined value, and it is determined whether the coupling coefficient has converged. If the coupling coefficient has not converged, the process returns to step 600, similar calculations are repeated using the newly corrected coupling coefficient, and the coupling coefficient is again corrected. Furthermore, using this coupling coefficient, input data (fixed condition data) corresponding to the teacher data, start the first neural network, and calculate its output (correction amount).
seek. Using the output and the training data, the first
Learning is performed by executing the program shown in Figure 2. Then, the above learning is repeatedly performed for certain input data until the deviation between the output and the teacher data falls within an allowable range. Furthermore, the above learning is executed according to other teacher data and input data corresponding thereto. Then, the above learning is repeatedly performed until the outputs for all the accumulated input data converge to all the teacher data. Note that the training data of the first neural network is for processing condition data that has been corrected in consideration of the correction amount of the second neural network. Therefore, as the learning progresses, it is possible to obtain an appropriate correction amount for the machining condition data using only the first neural network. That is, the correction amount that is the output of the second neural network converges to zero as learning of the first neural network progresses. 11. Learning of the second neural network Learning of the second neural network is also performed in exactly the same way as the learning of the first neural network, using teacher data accumulated during actual use. In this way, when the numerical control device is actually used, the coupling coefficients can be learned using the accumulated teacher data and the high-speed calculation capability of the off-line learning device. As a result, sophisticated learning is possible without placing a burden on the processing control of the numerical control device. In the above embodiment, the neural network calculation program and its coupling coefficient data necessary for processing are stored on the FD37.
The learning data is inputted from the FD37 and outputted to the FD37. However, even if the calculation program of the two-dimensional network and its coupling coefficient data are input from the main computer via a communication line, or the learning data is output to the main computer, the coupling coefficients can be learned by the main computer. good.

【Effect of the invention】

In the present invention, in a numerical control device, when a correction occurs when the correction amount of machining condition data is actually obtained using a neural network, the appropriate correction amount at that time is accumulated as training data, The teacher data is output to an offline learning device, which performs sophisticated learning, and the learned neural network is re-inputted into the numerical control. As a result, sophisticated learning is possible without placing a load on the processing control of the numerical control device and interrupting the processing control.

[Brief explanation of drawings]

FIG. 1 is a configuration diagram showing the overall mechanical configuration of a numerically controlled grinding machine having an automatic processing condition determination device according to a specific embodiment of the present invention. FIG. 2 is a block diagram showing the electrical configuration of the numerical control device and operation panel according to the same embodiment. FIG. 3 is a block diagram showing the configuration of the offline learning device. FIG. 4 is a flowchart showing a procedure for inputting a neural network calculation program and its coupling coefficient data by the CPU used in the processing condition automatic determination device. FIG. 5 is a flowchart showing the calculation procedure of machining condition data by CPIJ used in the automatic machining condition determination device. FIG. 6 is a structural diagram showing the structure of the first neural network. FIG. 7 is a flowchart showing the calculation procedure by the first neural network. FIG. 8 is a flowchart showing a procedure for correcting machining condition data based on machining error data by the CPU used in the automatic machining condition determination device. FIG. 9 is a structural diagram showing the structure of the second neural network. FIG. 10 is a flowchart showing a procedure for saving coupling coefficient data learned by CPLJ used in the processing condition automatic determination device in a storage medium. FIG. 11 is a flowchart showing the procedure for inputting learning data by the CPU of the offline learning device. FIG. 12 is a flowchart showing the learning procedure of the first neural network by the off-line learning device. FIG. 13 is a flowchart showing the operational concept of the entire system. Operation panel 21 Operation panel keyboard 23 CRT display numerical control device 31 CPU ROM 33-RAM 50 ・Grinding machine bed
52 Table table feed motor 54 Headstock spindle 56 Tailstock 57 Tailstock spindle motor 60-grinding wheel 61 Grinding wheel head grinding wheel drive motor Grinding wheel head feeding motor Workpiece

Claims (1)

[Claims] In a learning method for an automatic creation function of machining condition data for controlling machining conditions in a numerical control device, data regarding machining factors that influence machining condition data are input during actual machining, and Calculate the correction amount of the processing condition data using a neural network, evaluate the correction amount that is the output of the neural network, and when the correction amount is corrected as a result of the evaluation, the correction value is used as training data. , and store it in a storage device along with the input data of the neural network at that time, and calculate the coupling coefficient of the neural network off-line in a device different from the numerical control device that performs processing control according to the stored input data and teacher data. How to learn the machining condition creation function of a numerical control device.