Numerical control

Abstract

Numerical control, which controls a machine tool machining a workpiece with a tool, wherein the numerical control includes: a shape data acquisition unit that comprehensively captures tool information, including an image related to the shape of the tool; a machining condition capture unit that includes a machining condition in the machining of the workpiece; a machining result acquisition unit that acquires information relating to a machining result of the workpiece after the machining of the workpiece, wherein the information includes the presence of a cutting tool mark; and a machine learning device that performs machine learning on the tendency of the information associated with a machining result in relation to the tool information and the machining condition based on the tool information and machining condition used as input data and on the basis of information relating to a machining result used as teacher data in order to form a learning model, wherein The machine learning device is trained to determine whether a machining result is good or not by using the formed learning model based on the tool information acquired by the shape data acquisition unit and the machining condition acquired by the machining condition acquisition unit before the machine tool machines the workpiece.

Description

BACKGROUND OF THE INVENTION 1. Field of the invention

[0001] The present invention relates to a numerical control system and in particular to a numerical control system that learns a tool state that causes a machining error. 2. Description of the state of the art

[0002] With a standard tool used in machining, the cutting edge wears down, the cutting resistance increases over time, and the tool may eventually break. Furthermore, machining accuracy decreases with wear, making it impossible to achieve the required machining accuracy for a workpiece. To avoid such tool problems, it is primarily determined at the following intervals whether a tool has reached the end of its service life, and a tool change is performed when it is determined that the tool has reached this point. - At a point in time when the number of times or the time over which the tool is used exceeds a certain number (lifetime value). - At a point in time when it is determined by measuring the tool length that a wear quantity exceeds a certain number. - At a time when an operator determines, by visual inspection of the tool or a machined workpiece, that the tool has reached the end of its service life. - At a time when a break or chipping is detected by testing a spindle under load.

[0003] As a prior art method for determining whether a tool has reached the end of its service life, for example, Japanese patent application No. 2000-198047 discloses a method in which the amount of wear after machining is calculated on the basis of a tool shape, a cutting condition and an uncut amount of a workpiece in order to determine whether a tool has reached the end of its service life before machining.

[0004] The previously described method for determining tool life, however, has several problems. The method for determining tool life based on the number of times or the time the tool has been used generally sets a life value by more than a sufficient number of uses or usage time to avoid machining defects or breakage, so that the tool is often still usable and incurs costs. The method for determining tool life based on the tool length has a problem in that chip breakage cannot be detected when there are multiple cutting edges or when a cutting edge is located at the end.Furthermore, for a tool whose service life is determined based on wear in a section other than the end section, such as a tapping tool, the tool life cannot be determined by measuring the end section. The method for determining tool life based on the optical inspection of a machined workpiece by an operator requires an operator with technical knowledge and is therefore unsuitable for unattended operation. In the method for determining tool life by checking spindle load, slight chipping will hardly lead to changes in the spindle load, causing difficulties in accurate measurement.

[0005] Furthermore, in the previously described method for determining tool life, as described in Japanese patent application no. 2000-198047, the accuracy of determining tool life is improved by considering the tool shape, the cutting conditions, and the uncut portion of the workpiece. However, tool life cannot be determined solely based on the amount of wear on a tool. Tool life is also influenced by the tool material, the location of wear, the wear condition (e.g., chipping), and similar factors. Therefore, the method described in Japanese patent application no. 2000-198047 suffers from the problem that it is not designed for different situations. SUMMARY OF THE INVENTION

[0006] One object of the present invention is to provide a numerical control system that enables highly accurate determination of tool life.

[0007] In the present invention, the problems described above are solved by providing a numerical control with a function for learning the relationship between a machining condition and a machining result through supervised learning and improvement of the input data based on the determined learning result. The numerical control according to the present invention can continue machining while maintaining a certain level of quality by improving input data based on learning results.

[0008] A numerical control system according to the present invention controls a machine tool that machines a workpiece with a tool and comprises: a shape data acquisition unit that acquires tool information comprising a shape of the tool; a machining condition acquisition unit that comprises a machining condition in the machining of the workpiece; a machining result acquisition unit that acquires information relating to a machining result of the workpiece after the machining of the workpiece; a machine learning device that performs machine learning to determine the tendency of the information associated with a machining result in relation to the tool information and the machining condition based on the tool information and the machining condition used as input data and on the basis of information relating to a machining result used as teacher data in order to form a learning model.The machine learning device is trained to determine whether a machining result is good or not by using the formed learning model based on the tool information captured by the shape data acquisition unit and the machining condition captured by the machining condition acquisition unit, before the machine tool machines the workpiece.

[0009] The machine learning device can create a learning model through a multi-layered neural network.

[0010] Thus, the numerical control according to the present invention has the configuration described above. Accordingly, a reduction in tooling costs can be expected due to the improved determination of tool life, and unattended continuous operation can be supported. BRIEF DESCRIPTION OF THE DRAWINGS Fig. Figure 1 shows an overview of machine learning and the application of a learning result by a machine learning device integrated into a numerical control in the present invention. Fig. Figure 2 shows a configuration example (learning phase) of the machine learning device integrated into the numerical control according to the present invention. Fig. Figure 3 shows the numbering of images, sectional views and materials of tools, as well as the numbering of materials of a workpiece. Fig. Figure 4 shows a configuration example (operational phase) of the machine learning device integrated into the numerical control according to the present invention. Fig. Figure 5 shows a diagram (1) that is a schematic functional block diagram of the numerical control in the learning phase according to an embodiment of the present invention. Fig. Figure 6 shows a diagram (2) which is a schematic functional block diagram of the numerical control in the learning phase according to the embodiment of the present invention. Fig. Figure 7 shows a flowchart illustrating a company operating in Fig. 5 and Fig. 6 numerical control systems shown. Fig. Figure 8 shows a schematic functional block diagram of the numerical control in the operating phase according to the embodiment of the present invention. Fig. Figure 9 shows a flowchart illustrating a company operating in Fig. 8 numerical control shown. DETAILED DESCRIPTION OF THE PREFERRED EXECUTION FORM

[0011] Fig. Figure 1 shows an overview of machine learning and the application of a learning result by a machine learning device integrated into a numerical control system according to the present invention. The machine learning device integrated into the numerical control system according to the present invention operates in two phases, which are a learning phase and an operating phase.

[0012] During the learning phase, the machine learning device integrated into the numerical control according to the present invention uses as input data machining conditions such as tool information, a material of a workpiece, a cutting speed, a spindle speed, a quantity of the depth of cut and a cutting area, as well as a machining result such as a dimensional error of a workpiece machined on the basis of the machining conditions, in order to learn a tendency of the machining result with respect to the machining conditions.

[0013] During the operational phase, the numerical control according to the present invention provides machining conditions as input data for the machine learning device. The machine learning device outputs a predicted value of a machining result with respect to the input machining conditions, indicating the tendency of the machining result with respect to the machining conditions learned during the learning phase. Subsequently, the numerical control refers to the predicted value of a machining result output by the machine learning device to determine whether an error will occur in a subsequent machining operation.

[0014] In the numerical control system according to the present invention, the learning phase and the operating phase of the machine learning device can be executed at completely different times or in parallel. For example, after the machine learning device has been allowed to perform sufficient learning in the learning phase, operation can be carried out by a machine tool using only the learning results without performing new learning, i.e., by repeatedly using the same data, in the current machining operation. Alternatively, operation can be carried out in such a way that more accurate evaluation values ​​are sought (learned) during machining.

[0015] Furthermore, the machine learning device can be integrated into the numerical control of a machine tool or can be integrated into a PC connected to the numerical control.

[0016] Fig. 2 and Fig. Figure 3 shows a configuration example of the machine learning device integrated into the numerical control according to the present invention.

[0017] Fig. 2 and Fig. Figure 3 shows the configuration in which a multilayer neural network is used as the machine learning device. The multilayer neural network consists of an arithmetic device, a memory, etc., which execute a neural network trained as an imitation of a neuron model.

[0018] As in Fig. 2 shows that when the multilayer neural network is used as the machine learning device of the numerical control according to the present invention, the numerical control of the present invention uses machining conditions in the machine learning device as input data (X1 to X6 in the Fig. (as shown in the second example) before machining at each step (or at any time when a tool or cutting condition is changed, or at any time) during the learning phase. Examples of machining conditions include tool information (image, section view, material), a material of a numbered workpiece, a cutting speed (combined speed), a spindle speed, a depth of cut, and a cut quantity. Here, a variety of typical image and material data can be pre-numbered and stored for an image, a section view, or a material of a tool and a material of a workpiece, as shown in Fig. Figure 3 is shown, and the numbers can be used as input data for the machine learning device. This reduces the amount of input data and simplifies the structure of the multi-layered neural network.

[0019] Subsequently, according to the present invention, the numerical control system records output data W, which is output by the machine learning device when the machining conditions serving as input data are entered into the multilayer neural network, as a predicted machining result. Furthermore, after a workpiece has been machined based on the machining conditions entered into the multilayer neural network, the numerical control system records the dimensions of the machined workpiece, the number and size of the cutting tool marks, etc., as a machining result and calculates an error D between the recorded machining result and a design as teacher data. The numerical control system can automatically determine the dimensions of the machined workpiece, the number and size of the cutting tool marks, etc.For example, measuring by controlling a measuring instrument, or an operator can perform the measurement of the workpiece being machined and enter the measured result into the numerical control.

[0020] Subsequently, the numerical control according to the present invention enables the machine learning device to perform machine learning with respect to the machining results in relation to the machining conditions. This is based on the machining conditions serving as input data, the predicted machining result output by the multi-layer neural network, and the error D (teacher data) between the dimension of the machined workpiece, which serves as the recorded machining result, and the design for forming a learning model. Since the learning method (backpropagation or similar) is already well known as a prior art method with respect to a neural network, a detailed description of it is omitted in this patent specification. The learning process is carried out until a difference between the output data and a machining result is established within a predetermined value.The distance is calculated using a loss function. As an example of a loss function, a quadratic error (L² loss) as expressed in the following formula (1) can be used, for example, if the focus is on an error between a dimension of a machined workpiece and a design value. Furthermore, if the focus is on the presence of a cutting mark, a logical function can be used that simply expresses whether a cutting mark is present or not. Alternatively, any function can be used, provided that the function can adequately express a loss. L(y)(Xn;W),D)≡(y(Xn;W)−D)2

[0021] As in Fig. As shown in Figure 4, when the multilayer neural network is used as the machine learning device of the numerical control according to the present invention, the numerical control processing conditions in the machine learning device are used as input data (X1 to X6 in the Fig. (Example 2) before machining at each step (at a time when a tool or cutting condition is changed, or at any time) in the operation phase. The machining conditions input into the machine learning device in the operation phase are the same elements as those used in the learning phase.

[0022] The numerical control then captures the output data W, which is output by the machine learning device when the machining conditions serving as input data are entered into the multi-layer neural network, as a predicted machining result (an error of a predicted dimension of a machined workpiece in relation to a design, the presence of a cutting tool mark, etc.).

[0023] Subsequently, if it is determined that a machining result is poor when the predicted machining result, which serves as output data, is compared with a predetermined threshold (an error of a predicted dimension of a machined workpiece with respect to a design is equal to or greater than a predetermined specific value, the number of cutting marks is equal to or greater than a predetermined number, the size of a cutting mark is equal to or greater than a predetermined size, etc.), the numerical control according to the present invention determines that the machining result is a machining detection pattern and prompts an operator to perform a tool change.

[0024] Fig. 5 and Fig. Figure 6 shows functional block diagrams illustrating essential parts of a numerical control system according to an embodiment of the present invention during the learning phase. Figure 6 further shows Fig. 7 a flowchart to illustrate the process of an operation of a numerical control 1 in the learning phase.

[0025] The numerical control 1 according to the present embodiment comprises a machine learning device 10, a shape data acquisition unit 11, a machining condition acquisition unit 12, and a machining result acquisition unit 13. Furthermore, the machine learning device 10 comprises an input condition monitoring unit 110, an input condition storage unit 120, a unit for calculating a predicted machining result 130, a unit for storing a predicted machining result 140, a machining result monitoring unit 150, a machining result storage unit 160, and a comparison unit 170.

[0026] The numerical control 1 switches to the learning phase and begins learning when the time arrives to perform the learning of the machine learning device 10 (step SA01). Examples of the time to switch to the learning phase include a time when a tool is changed in use, a time when a machining condition (a machining content, a cutting speed, a rotational speed, etc.) is changed, and any time specified by a user.

[0027] The shape data acquisition unit 11 and the machining condition acquisition unit 12 each perform the acquisition of tool information (shape information (an image, a section view), a tool material, etc.) and the acquisition of machining conditions (a workpiece material, machining contents (flat surface / side surface, etc.), a spindle speed, a cutting speed, a cutting area) (steps SA02 and SA03). Tool information and machining conditions can be acquired based on setpoints or a machining program of the numerical control 1, or entered by an operator. Here, numerical values ​​are used for shape information, a tool material, and a workpiece material, as shown in Fig. 3. In particular, for the shape information of a tool, a tool attached to a machine tool can be subjected to imaging by an imaging device (not shown), and image processing can be performed to capture an image representing the shape of the tool to be used. The captured image can then be combined with typical tool image data as shown in Fig. 3 are shown and matched to use a number of image data to represent the next shape as the tool information.

[0028] The input condition monitoring unit 110 captures the tool information captured by the shape data acquisition unit 11 and the machining conditions captured by the machining condition acquisition unit 12 as input conditions for the machine learning device 10. The input condition monitoring unit 110 observes the acquisition states of the respective data captured by the shape data acquisition unit 11 and the machining condition acquisition unit 12. When the shape data acquisition unit 11 and / or the machining condition acquisition unit 12 acquire data, the input condition monitoring unit 110 captures the data as input data.

[0029] The input condition storage unit 120 stores the input conditions detected by the input condition monitoring unit 110 in a memory that is not shown. The input condition storage unit 120 can only overwrite and store input conditions already stored in the memory that are observed and detected by the input condition monitoring unit 110. Thus, only input conditions that change can be updated.

[0030] The unit for calculating a predicted processing result (130) calculates a predicted processing result based on the input conditions stored by the input condition storage unit (120) and a learning content (learning model) stored by the unit for storing a predicted processing result (140) (step SA04). The unit for storing a predicted processing result (140) consists, for example, of a multi-layered neural network as described previously.

[0031] After the numerical control 1 machines a workpiece while controlling a machine tool 2 based on the machining condition acquisition unit 12 and a machining program (not shown) (step SA05), the machining result acquisition unit 13 acquires a machining result from the machined workpiece (step SA06). A machining result to be acquired includes a deviation between a dimension of the machined workpiece and a design dimension, the presence of a cutting tool mark, etc.

[0032] The processing result monitoring unit 150 records the processing result, which is recorded by the processing result acquisition unit 13, in the machine learning device 10. The processing result monitoring unit 150 observes the acquisition state of all data acquired by the processing result acquisition unit 13. When a processing result is acquired, the processing result monitoring unit 150 records the data in the machine learning device 10.

[0033] The processing result storage unit 160 stores the processing result captured by the processing result monitoring unit 150 in a memory that is not shown.

[0034] The comparison unit 170 compares the predicted processing result calculated by the predicted processing result unit 130 with the actual processing result captured by the processing result observation unit 150 (and stored by the processing result storage unit 160) and calculates a difference between the predicted processing result and the actual processing result by using a loss function to compare the difference with a predefined threshold (step SA07). If the difference between the predicted processing result and the actual processing result is within the predefined threshold, no learning is performed.If the difference is equal to or greater than the threshold, learning is performed based on the predicted processing result and the actual processing result, and the determined learning result is stored in the unit for storing a predicted processing result 140 (steps SA08 and SA09).

[0035] Fig. Figure 8 shows a function block diagram to represent essential parts when the in Fig. 5 and Fig. The numerical control described in section 6 can operate during the operating phase. Furthermore, it shows Fig. Figure 9 shows a flow diagram illustrating the operation of the numerical control system 1 during the operating phase. During the operating phase, the numerical control system 1 further comprises a tool change instruction unit 14, and the machine learning device 10 further comprises a tool change determination unit 180.

[0036] During the operational phase, the numerical control 1, according to the present embodiment, performs a pre-machining check operation with the machine learning device 10 when the numerical control 1 is in a state for performing the pre-machining check (step SB01). Examples of the state for performing the pre-machining check include a time when a tool is changed into use, a time when a machining condition (a machining content, a cutting speed, a rotational speed, etc.) is changed, and an arbitrary time defined by a user.

[0037] The shape data acquisition unit 11 and the machining condition acquisition unit 12 each perform the acquisition of tool information (shape information (an image, a section view), a material of a tool, etc.) and the acquisition of machining conditions (a material of a workpiece, machining contents (flat surface / side surface, etc.), a rotational speed, a cutting speed, a cutting area) (steps SB02 and SB03).

[0038] The input condition observation unit 110 captures the tool information captured by the form data acquisition unit 11 and the machining conditions captured by the machining condition acquisition unit 12 as input conditions of the machine learning device 10.

[0039] The input condition storage unit 120 stores the input conditions detected by the input condition observation unit 110 in a memory that is not shown.

[0040] The unit for calculating a predicted processing result 130 calculates a predicted processing result based on the input conditions stored by the input condition storage unit 120 and learning content stored by the unit for storing a predicted processing result 140 (step SB04).

[0041] The previously described operational processes by the form data acquisition unit 11, the processing condition acquisition unit 12, the input condition observation unit 110, the input condition storage unit 120, the unit for calculating a predicted processing result 130 and the unit for storing a predicted processing result 140 are the same as those carried out by the respective functional means in the learning phase.

[0042] The tool change determination unit 180 determines the tool change based on the predicted machining result calculated by the predicted machining result calculation unit 130. The tool change determination unit 180 refers to the predicted machining result output by the predicted machining result calculation unit 130 (a deviation between a dimension of a machined workpiece and a design value, the presence of a cutting tool mark, etc.), and if the machining result is poor (the deviation is equal to or greater than a predetermined threshold, the number of cutting tool marks is equal to or greater than a predetermined number, the size of a cutting tool mark is equal to or greater than a predetermined size, etc.).), the tool change detection unit 180 determines that the tool needs to be changed (step SB05).

[0043] Based on the determination result of the tool change determination unit 180, if it is determined that the tool needs to be changed, the tool change instruction unit 14 indicates that the tool needs to be changed (step SB06). The tool change instruction unit 14 can issue a tool change instruction to the machine tool 2 for automatic tool changes or can display a tool change prompt on a display device of an operator console (not shown) to allow an operator to view the tool change indicator.

[0044] The embodiment of the present invention has been described so far. However, the present invention is not limited to the previously described example of the embodiment and can be carried out by making appropriate modifications.

[0045] For example, the example using the multilayer neural network as the machine learning device 10 is described in the preceding embodiment. However, various methods such as the least squares method, the step method, SVM, and decision tree learning are also well-known supervised learning algorithms, and any of these methods can be used accordingly as the machine learning device 10 of the present invention.

[0046] Furthermore, as with regard to Fig. 3. Images representing tool shapes are pre-numbered and stored in the numerical control 1 according to the present invention. This numbering can be done manually; however, automation is more easily performed, such that clustering is carried out by unsupervised learning with respect to a large number of tool images, representative images are determined based on the result of the clustering, and the numbering is carried out with respect to each of the representative images.

[0047] Furthermore, the operation of the numerical control 1 in the learning phase and the operation of the numerical control 1 in the operating phase are individually compared with regard to Fig. 5, Fig. 6, Fig. 7, Fig. 8 to Fig.9 described; however, the learning phase and the operational phase can also be carried out in parallel. In such a case, the pre-processing check takes place in the operational phase, and processing result information is captured at the same time as the completion of the processing by the machine learning device, whose learning has progressed to a certain level. In such a state, where a difference between a predicted processing result and an actual processing result is equal to or greater than a predefined threshold, learning can take place with respect to the unit for storing a predicted processing result 140.

Claims

[1] Numerical control that controls a machine tool machining a workpiece with a tool, wherein the numerical control comprises: a shape data acquisition unit that comprehensively captures tool information, including an image related to the shape of the tool; a machining condition capture unit that includes a machining condition in the machining of the workpiece; a machining result acquisition unit that acquires information relating to a machining result of the workpiece after the machining of the workpiece, wherein the information includes the presence of a cutting tool mark; and a machine learning device that performs machine learning on the tendency of the information associated with a machining result in relation to the tool information and the machining condition based on the tool information and machining condition used as input data and on the basis of information relating to a machining result used as teacher data in order to form a learning model, wherein The machine learning device is trained to determine whether a machining result is good or not by using the formed learning model based on the tool information acquired by the shape data acquisition unit and the machining condition acquired by the machining condition acquisition unit before the machine tool machines the workpiece. [2] Numerical control according to claim 1, wherein the machine learning device forms a learning model by means of a multilayer neural network.

Images