Learning device and cutting process evaluation system

Abstract

To evaluate accurately a cutting processing.SOLUTION: In a leaning device including an input processing part 2 and a leaning processing part 3, the input processing part 2 acquires a physical quantity 15 related to a cutting processing, and inputs a state variable 12 based on the physical quantity 15 into the leaning processing part 3, and the leaning processing part 3 updates based on a measured cutting result, an evaluation model 11 for outputting an evaluation result 13 of the cutting processing based on the state variable 12.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present disclosure relates to a learning device used in an evaluation system for a workpiece manufactured by cutting, and a cutting evaluation system using the same.

Background Art

[0002] In cutting, generally, while holding a workpiece placed on a member called a die with a member called a stripper, a tool called a punch presses the workpiece into the die to punch it out to obtain a predetermined shape. Cutting is generally widespread in various manufacturing fields such as the manufacture of household appliances, the manufacture of precision instruments, or the manufacture of automotive parts.

[0003] In cutting using such a die, it is common to adjust the die position or die shape by trial and error according to each individual die. However, there are cases where adjustment by such trial and error cannot be handled, and in such cases, a processed product of a predetermined quality cannot be obtained. Therefore, an evaluation method has been proposed, such as the evaluation method disclosed in Patent Document 1, which measures a physical quantity generated by a cutting process and performs an abnormality diagnosis by comparing the measured value of this physical quantity with a reference value.

[0004] Also, as a method for determining the quality of a processed product in a general processing apparatus, an evaluation method has been proposed, such as the evaluation method disclosed in Patent Document 2, which performs a quality determination by comparing a measured value of internal information of the processing apparatus with a threshold value set in a provisional determination unit, and feeds back the quality of an actual processed product to update the threshold value of the provisional determination unit.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Patent Document 2

[0006] However, in the evaluation method disclosed in Patent Document 1, the criterion for determining whether something is normal or not is whether the acquired value (measured value) falls within an acceptable range (tolerance range) relative to the value obtained from normal processing (reference value). For this reason, the setting of the tolerance range must be considered each time depending on the specifications of the workpiece and processed product. Moreover, even if an abnormality is detected in the evaluation method disclosed in Patent Document 1, the cause cannot be identified, making it time-consuming to deal with the abnormality.

[0007] Furthermore, the evaluation method described in Patent Document 2 has the following problems. In other words, cutting processes generally involve a fracture mechanics aspect in which crack growth occurs after compressive deformation and plastic deformation, followed by fracture separation, and this fracture process takes place in an extremely short time of less than 0.1 seconds. Therefore, internal information such as the motor's current value and rotational speed cannot capture the state of the fracture process, such as the load profile at the time of fracture separation, and it is difficult to obtain the necessary number of data samples in a short time. For this reason, it is difficult to accurately determine the quality of the processing process (and consequently the quality of the processed product) based solely on the internal information of the processing equipment.

[0008] This disclosure aims to enable accurate evaluation of cutting processes in light of the above-mentioned conventional problems. [Means for solving the problem]

[0009] The learning device of this disclosure comprises an input processing unit and a learning processing unit, wherein the input processing unit acquires physical quantities related to cutting, inputs state variables based on the physical quantities to the learning processing unit, and updates an evaluation model that derives an evaluation of the cutting process based on the state variables, based on the measured cutting results.

[0010] The cutting process evaluation system of this disclosure comprises a sensor for measuring the physical quantity, a learning device, and an output processing unit for deriving the evaluation using the evaluation model updated by the learning device. [Effects of the Invention]

[0011] According to the learning device and cutting process evaluation system of this disclosure, the processing quality can be evaluated by an evaluation model using state variables based on physical quantities related to the cutting process, and the evaluation model can be updated. Therefore, the processing quality can be evaluated with high accuracy. [Brief explanation of the drawing]

[0012] [Figure 1] A block diagram illustrating the outline of the cutting process evaluation system used in the embodiments of this disclosure. [Figure 2] Block diagram showing an overview of the learning process in the evaluation system used in the embodiments of this disclosure. [Figure 3] A general cutting load-punch stroke diagram illustrating the correlation between load and processing quality used in the embodiments of this disclosure. [Figure 4] A general sound-time diagram illustrating the correlation between sound and processing quality used in the embodiments of this disclosure. [Figure 5] A general temperature-time diagram illustrating the correlation between temperature and processing quality used in the embodiments of this disclosure. [Figure 6] Diagram of the convolutional neural network used in the embodiments of this disclosure [Figure 7] An overall diagram showing the layout of a cutting apparatus to which the cutting process evaluation system used in the embodiments of this disclosure is applied. [Figure 8] Functional block diagram relating to the control unit of the embodiment of the present disclosure [Figure 9] A flowchart illustrating the learning steps performed by the learning device according to the embodiment of this disclosure. [Figure 10] A flowchart illustrating the evaluation steps performed by the cutting process evaluation system according to the embodiment of this disclosure. [Figure 11]Figure showing an example of cutting result

Mode for Carrying Out the Invention

[0013] Hereinafter, embodiments of the present disclosure will be described with reference to FIGS. 1 to 10.

[0014] FIG. 1 shows a block principle diagram showing an overview of a cutting process evaluation system 1 according to an embodiment of the present disclosure. In FIG. 1, a cutting process evaluation system 1 according to the present disclosure includes, as functions, a learning device 21 and an output processing unit 4 that executes an output step. The learning device 21 includes, as functions, an input processing unit 2 that executes an input step and a learning processing unit 3 that executes a learning step. In the input step, acquisition of a physical quantity 15 measured by a sensor 150 and generation of a state variable 12 based on the physical quantity 15 are performed. In the output step, an output of a cutting evaluation result 13 obtained by inputting the state variable 12 to an evaluation model 11 is performed. Further, the cutting process evaluation system 1 may further include a sensor 150 that measures a physical quantity 15 related to the cutting process. The sensor 150 includes at least one or more of a load sensor 151, a sound sensor 152, a position sensor 153, and a temperature sensor 154 described later.

[0015] The input processing unit 2 acquires the physical quantity 15 measured during the cutting process as the state variable 12 and inputs it to an evaluation model 11 (described later) of the learning processing unit 3.

[0016] The learning processing unit 3 is included in the learning device 21 and includes a learned evaluation model 11 and a dataset group 14.

[0017] The output processing unit 4 outputs the cutting evaluation result 13.

[0018] With this configuration, the cutting process evaluation system 1 is configured to take state variables 12 as input to a pre-trained evaluation model 11 and output a cutting evaluation result 13. The evaluation model 11 is optimized through a learning step by the learning processing unit 3 using a dataset 14. The cutting evaluation result 13 is a prediction of whether or not there was an abnormality in the cutting process when the physical quantity 15 was measured, and the cause of the abnormality if it occurred.

[0019] The state variables 12 input to the evaluation model 11 include at least one of the following: cutting load, noise generated during cutting, vibration generated during cutting, shear rate, clearance between die and punch, and workpiece temperature generated during cutting (hereinafter referred to as "processing temperature"). Physical quantities 15 measured in real time from the start to the end of a single cutting operation are converted as needed and input to the evaluation model 11 as state variables 12. Since a sufficient number of samples are needed to capture the characteristics of the process, such as the curvature of the curve of the load profile, and the trends of local values ​​of physical quantities 15, the sampling period (measurement period) of physical quantities 15 should preferably be 1 / 100th or less of the time required for cutting.

[0020] The evaluation model 11 receives state variables 12 based on physical quantities 15 as input. The evaluation model 11 is a model equipped with a function that processes the input state variables 12 and converts them into an output (i.e., a function that calculates and outputs a segmented evaluation result 13 based on the state variables 12). By optimizing the function in the learning processing unit 3, which will be described later in Figure 2, the evaluation model 11 can obtain a highly accurate segmented evaluation result 13.

[0021] The cutting evaluation result 13 evaluates abnormalities in the cutting process in stages and is classified into n+1 patterns, which is the sum of one pattern for cases where there are no abnormalities in the process and n patterns for the number of abnormality causes when there are abnormalities in the process. Examples of abnormality causes include known defects such as excessive clearance, insufficient clearance, tool wear, and incorrect mold installation.

[0022] The cutting evaluation result 13 is specifically a one-dimensional vector that holds the probability of the above pattern for each of the n+1 elements, and the output processing unit 4 outputs the element with the largest value among the elements, i.e., the pattern with the highest probability of occurring (cause of defect).

[0023] The dataset group 14 used to train the evaluation model 11 is a collection of two sets of data: input data and output data. Specifically, the dataset group 14 is a collection of state variables 12 based on physical quantities 15 measured during a single cutting process, as input data, and the cutting results 16 (see Figure 2) when those state variables 12 were measured, as output data, with each set of data collected for each cutting process.

[0024] State variable 12 is a variable based on physical quantity 15, which is determined by actual measurement for each processing step. State variable 12 may be obtained as the physical quantity 15 as is, but it is preferable to obtain a transformed version of the physical quantity 15. For example, as will be described later with reference to Figure 7, when the load is measured by the load sensor 151 shown in Figure 7 to measure the cutting load, which is an example of physical quantity 15, the detected load includes not only the cutting load but also the load from the stripper 107. Therefore, the load from the stripper 107 is calculated from the spring constant of the spring included in the stripper 107 and the linear load approximation line. Then, the cutting load is calculated by removing the load from the stripper 107 from the load measured by the load sensor 151. In this way, only the calculated load, i.e., the cutting load, is obtained as state variable 12.

[0025] In other words, the information obtained as physical quantity 15 may contain information other than that necessary for evaluating the cutting process. Therefore, it is preferable to obtain the state variable 12 by appropriately converting the physical quantity 15.

[0026] The cutting result 16 is determined by actual measurement for each processing step and is associated with the physical quantity 15 during cutting. That is, when a certain processing step is performed, the detection of the physical quantity 15 and the determination of the cutting result 16 are performed together and input as a set into the dataset group 14. The cutting result 16 is judged, for example, as the quality of the processed product, by a person or by mechanical means using conventional technology, and input into the dataset group 14. The cutting result 16 is measured, for example, by a device that measures cutting results and input as a signal. Alternatively, the cutting result 16 may be evaluated by a person in an inspection process, for example, and input via an input device such as a keyboard. The quality of the processing process is most easily judged based on the quality of the processed product, but the quality of the processing process may also be judged based on other criteria.

[0027] The cutting results 16 evaluate abnormalities in the cutting process in stages and are classified into n+1 patterns, which is the sum of one pattern for cases where there are no abnormalities in the processing and n patterns for the number of abnormality causes when there are abnormalities in the processing. Examples of abnormality causes include clearance deviation, tool wear, chipping, burr height, debris clogging, and incorrect die installation. Chipping refers to the phenomenon in machining and press working where the cutting edge of a cutting tool or cutting blade breaks into small pieces. Burr height is the length of the burr generated on the processed workpiece along the thickness direction of the workpiece. Debris clogging is the accumulation of chips in the clearance between the die and the punch. An example of the cutting results 16, which classifies abnormalities in the cutting process in stages, is explained based on Figure 11.

[0028] Figure 11 shows an example of cutting results 16. In the example shown in Figure 11, the causes of abnormalities are evaluated as follows: the magnitude of tool wear, the maximum width of chipping, the burr height, debris clogging, and the magnitude of clearance deviation. As shown in Figure 11, the degree of each cause of abnormality is classified into three stages, labeled 1 to 3. That is, label 1 indicates no abnormality, label 2 indicates a yellow light, i.e., caution is required, and label 3 indicates an abnormality.

[0029] Specifically, if the tool wear is less than R20μm, it is evaluated as normal and labeled with label 1. If the tool wear is between R20μm and R25μm, it is evaluated as a warning sign and labeled with label 2. If the tool wear is greater than R25μm, it is evaluated as abnormal and labeled with label 3.

[0030] Furthermore, if the maximum chipping width is less than 5 μm, it is evaluated as normal and labeled as label 1. If the maximum chipping width is 5 μm or more but less than 10 μm, it is evaluated as a yellow light and labeled as label 2. If the maximum chipping width is 10 μm or more, it is evaluated as abnormal and labeled as label 3.

[0031] Furthermore, if the burr height is less than 15 μm, it is evaluated as normal and labeled with label 1. If the burr height is 15 μm or more but less than 20 μm, it is evaluated as a yellow light and labeled with label 2. If the burr height is 20 μm or more, it is evaluated as abnormal and labeled with label 3.

[0032] Furthermore, if there is no or virtually no clogging, it is evaluated as normal and labeled as label 1. If there is virtually clogging, it is evaluated as abnormal and labeled as label 3. Here, "virtually no clogging" or "clogged" means that there is no or present clogging that would cause an abnormality.

[0033] Furthermore, if the clearance deviation is less than 1 μm, it is evaluated as normal and labeled with label 1. If the clearance deviation is 1 μm or more but less than 2 μm, it is evaluated as a yellow light and labeled with label 2. If the clearance deviation is 2 μm or more, it is evaluated as abnormal and labeled with label 3.

[0034] In the example shown in Figure 11, the magnitude of tool wear, the maximum chipping width, the burr height, and the clearance deviation are evaluated numerically (scalar), and classified into multiple stages based on these values, but this is not the only method. For example, tool wear, chipping, burrs, and clearance deviation may be evaluated in two stages (presence or absence of abnormality due to each cause of abnormality). Alternatively, the numerical values ​​of the magnitude of tool wear, the maximum chipping width, the burr height, and the clearance deviation may be used as the cutting result 16. In other words, the cutting result 16 represents the degree of abnormality in the cutting process in stages or numerically for each type of abnormal cause. The cutting result 16 only needs to include one or more types of abnormal causes.

[0035] Figure 2 shows a block diagram illustrating the overview of the learning steps performed by the learning processing unit 3 of the evaluation system 1 used in the embodiment of this disclosure. From the dataset (state variables 12 and cut results 16) extracted from the dataset group 14, the state variables 12 based on the input data, i.e., the physical quantities 15 during the cut, are input to the evaluation model 11. The error 17 between the cut evaluation result 13 output from the evaluation model 11 and the cut results 16 extracted from the dataset group 14 is calculated using the loss function 18. Then, the weight coefficients of the evaluation model 11 are updated using the optimization algorithm 19 based on this error 17. This series of operations to update the weight coefficients is performed using all the datasets accumulated in the dataset group 14. Specifically, learning is performed by repeating the update of the weight coefficients of the evaluation model 11 using the entire dataset group 14 until the error 17, which is the sum of the differences between all cut results 16 in the entire dataset group 14 and the cut evaluation result 13 estimated by the evaluation model 11, is minimized and convergence occurs. As a loss function 18 suitable for deriving the error 17, it is desirable to use the cross-entropy error, which is a loss function suitable for the group classification algorithm, since the evaluation model 11 outputs the truncated evaluation result 13 by classifying according to the aforementioned anomaly cause patterns. Furthermore, it is desirable to use the steepest descent method or RMSprop as the method used in the optimization algorithm 19.

[0036] Here, we will explain the specific method for calculating the error using the example of the segmentation result 16 in Figure 11. As shown in Figure 11, if the segmentation result 16 contains multiple causes of anomalies, the learning processing unit 3 uses multiple evaluation models 11, each associated with one of the multiple causes of anomalies. In other words, the learning processing unit 3 uses the same number of evaluation models 11 as there are types of causes of anomalies.

[0037] As shown in Figure 11, when evaluating the segmentation results 16 in stages, the segmentation results 16 and the segmentation evaluation results 13 are expressed as the probability of each label's event occurring. For example, if the segmentation result 16 (measured value) for a certain abnormal cause is label 1, the segmentation result 16 is represented as

[0100] . Also, if the segmentation evaluation result 13 (predicted value) for a certain abnormal cause has a probability of 0.2 for label 1, a probability of 0.7 for label 2, and a probability of 0.1 for label 3, the segmentation evaluation result 13 is represented as [0.2 0.7 0.1].

[0038] When evaluating the truncation result 16 step by step, as described above, the cross-entropy error E of equation (1) below is used as the loss function 18. When the above measured value

[0100] and predicted value [0.2 0.7 0.1] are obtained, the error 17 is calculated using equation (1) below, as shown in equation (2) below.

[0039]

number

[0040]

number

[0041] When numerically evaluating the cutting result 16, the mean squared error E shown in equation (3) below may be used as the loss function 18. For example, if the measured value of the burr height is 13 μm and the predicted value is 18 μm, the error 17 may be calculated using equation (3) below, as shown in equation (4) below. Alternatively, another function capable of deriving errors of two scalar values ​​may be used instead of the mean squared error.

[0042]

number

[0043]

number

[0044] In this way, the learning processing unit 3 calculates an error 17 for each type of anomaly cause, and based on the calculated error 17, updates the weight coefficients of the evaluation model 11 associated with each anomaly cause using the optimization algorithm 19.

[0045] The learning step of the learning processing unit 3 can be performed in parallel with the machining process. That is, it is possible to optimize the evaluation model 11 through learning in real time during machining. However, the dataset 14 requires the measured cutting results 16. Therefore, it is preferable to perform the learning step after the completion of the cutting process, which is the timing when the input state variables 12 and the measured cutting results 16 are obtained.

[0046] For the evaluation model 11 to be effectively optimized through the learning steps, a strong correlation must exist between the input data and output data of the dataset 14. Optimizing the evaluation system 1 requires the appropriate selection of state variables 12 based on physical quantities 15 that have a strong correlation with machining anomalies. Examples of state variables 12 with a strong correlation to machining anomalies include the correlation between cutting load and machining quality, the correlation between sound generated during cutting and machining quality, and the correlation between machining temperature and machining quality, which are described below.

[0047] Figure 3 shows a general cutting load-punch stroke diagram illustrating the correlation between cutting load (shear load) and processing quality used in the embodiments of this disclosure. The cutting load-punch stroke diagram is a diagram in which the vertical axis represents the shear load generated during workpiece cutting and the horizontal axis represents the punch stroke during workpiece cutting. This cutting load-punch stroke diagram shows that there are five processes in the progression of the cutting load.

[0048] The section from point 31 to point 32 in Figure 3 is generally referred to as the compression deformation process, where the workpiece is compressed by the punch and die, causing sagging and the tool to bite into both the die and the punch. For example, if the tool is worn, it is expected that there will be less biting into the workpiece, and a large punch stroke (hereinafter also simply referred to as "stroke") will be required before moving on to the next shear deformation process.

[0049] The section from point 32 to point 33 in Figure 3 is generally referred to as the shear deformation process, where sliding deformation occurs in the workpiece, causing it to curve and generating a bending moment and tensile force. Generally, the larger the clearance between the die and the punch (hereinafter, when simply referred to as "clearance," it refers to the clearance between the die and the punch), the larger the bending moment generated in the workpiece. Therefore, if the clearance is excessive, it is expected that the strong bending moment generated in the workpiece in this section will act on the inner wall of the die (die hole), increasing the breaking load. Conversely, if the clearance is insufficient, it is expected that the bending moment acting on the inner wall of the die will decrease, reducing the breaking load.

[0050] The section from point 33 to point 34 in Figure 3 is generally referred to as the crack growth process, where cracks develop in the workpiece and the shear load begins to decrease. Generally, the larger the clearance, the earlier the crack develops. Therefore, if the clearance is excessive, it is expected that the punch stroke required to reach this section will be shorter than usual. In other words, if the clearance is excessive, the crack growth process will be reached with a shorter stroke.

[0051] The section from point 34 to point 35 in Figure 3 is generally referred to as the fracture separation process, where cracks propagating from both tool (punch and die) sides of the workpiece converge, causing the workpiece to separate (fracture). Generally, if the clearance is small, the crack may become stationary and secondary shear may occur; therefore, if the clearance is too small, secondary shear will occur. For this reason, the diagram is expected to have multiple slopes in this section.

[0052] The section from point 35 onwards in Figure 3 represents the process of the punch passing through the die after the cutting process is complete. If cutting debris from the workpiece adheres to the punch or die, or if the punch and die are misaligned and the punch comes into contact with the die, it is expected that a cutting load will still remain even after the workpiece has been punched out.

[0053] Furthermore, the punch stroke length in the section from point 31 to point 35 in Figure 3 is expected to correlate with the workpiece thickness. If the workpiece thickness is thicker than expected, the stroke length from point 31 to point 35 will increase, and if the workpiece thickness is thinner than expected, the stroke length from point 31 to point 35 will decrease. Therefore, it is expected that anomalies, such as differences in workpiece thickness, can be detected from the punch stroke length.

[0054] Figure 4 shows a typical sound-time diagram illustrating the correlation between sound (processing sound) generated during cutting and processing quality, as used in the embodiments of this disclosure. The vertical axis represents the gain of the processing sound generated when cutting the workpiece, and the horizontal axis represents the time from the moment the punch member starts moving. The presence of two peaks is shown during the cutting process. The peak at point 36 is due to the sound generated at the moment the stripper presses down on the workpiece, and the peak at point 37 is due to the sound generated when the workpiece is pressed and cut by the punch, that is, from the moment the punch contacts the workpiece until the cutting of the workpiece is completed.

[0055] If a peak exists between points 36 and 37, it is expected that the punch is in sliding contact with a wall surface that defines a guide for the punch provided in the stripper, and that punch wear will progress prematurely.

[0056] If the peak at point 37 is high, there is a high possibility that a crack has occurred in the workpiece, and the workpiece is expected to be a defective product.

[0057] If a peak exists after point 37, it indicates that the punch is rubbing against the inner wall (die hole) of the die, and that wear on the punch and die is likely to progress prematurely.

[0058] Figure 5 shows a typical processing temperature-time diagram illustrating the correlation between temperature and processing quality used in the embodiments of this disclosure. The vertical axis represents the processing temperature generated during workpiece cutting, and the horizontal axis represents the time from the start of processing. A peak point 38 in the processing temperature occurs only during cutting (i.e., from the moment the punch contacts the workpiece until the cutting of the workpiece is completed).

[0059] If the peak at point 38 in the processing temperature is high, it indicates that the energy required for processing is high, placing an excessive load on the punch and die, and it is expected that the lifespan of the punch and die will be reduced prematurely.

[0060] As described above, by capturing the characteristics of the cutting load-stroke diagram, processing sound-time diagram, and processing temperature-time diagram, it is expected that the presence and cause of abnormalities in the cutting process can be identified. Therefore, the evaluation model 11 in the cutting process evaluation system 1 is preferably a model that can capture the characteristics of the state variables 12 (cutting load, processing sound, processing temperature, etc.), and preferably a convolutional neural network model applied to image recognition algorithms is used.

[0061] Figure 6 shows a diagram of the configuration of the convolutional neural network 51 applied to the evaluation model 11 in the embodiment of this disclosure. First, the state variable 12 is input to the input layer 52, and the convolutional layer 53 repeatedly captures local data features of the state variable 12 for various target data and data locations, such as the curvature of the curve in the compression deformation process section shown from point 31 to point 32 in the cutting load-punch stroke diagram of Figure 3, and the slope of the fracture separation process shown from point 34 to point 35. In this way, the convolutional layer 53 extracts the features of the entire state variable 12. Next, the pooling layer 54 processes the features extracted by the convolutional layer 53 to make them more prominent. Finally, the fully connected layer 55 classifies using the features summarized by the pooling layer 54 and outputs the results to the output layer 56.

[0062] Thus, the convolutional neural network 51 is a model that excels at grasping and classifying the features of input data. Therefore, by applying the convolutional neural network 51 to the evaluation model 11 explained in Figure 3, the ability to grasp the features of the state variable 12 is improved, and the accuracy of the output truncated evaluation result 13 is greatly improved.

[0063] Figure 7 shows an overall view of a cutting apparatus 101 to which the cutting process evaluation system 1 used in the embodiment of this disclosure is applied. In the cutting apparatus 101, a workpiece (not shown) is placed on a die 102 and, while being held down by a stripper 107, is punched into the inner diameter (die hole 102a) of the die 102 by a descending punch 103. The cutting apparatus 101 is also equipped with a load sensor 151, a sound sensor 152, a position sensor 153, and a temperature sensor 154 as sensors to measure the physical quantity 15 generated when cutting the workpiece. The cutting process evaluation system 1 is configured with these load sensor 151, sound sensor 152, position sensor 153, and temperature sensor 154, along with the control unit 112 described later.

[0064] The load sensor 151 is preferably capable of measuring with high sensitivity the load (cutting load) that the punch 103 applies to the workpiece placed on the die 102. For this reason, it is desirable to install the load sensor 151 directly below the base 108 on which the die 102 is mounted. The specific number of load sensors 151 is preferably 2 to 4, with 3 being optimal to ensure that the cutting load is reliably distributed to all load sensors 151. The position of the load sensors 151 should be such that they are arranged at equal intervals and no part of the load sensor 151 protrudes from the bottom surface of the base 108. As for the load sensor 151, a quartz piezoelectric sensor is preferable because high-speed measurement (fast response measurement) is desirable, and a three-component load sensor that can measure cutting load in both the vertical and horizontal directions is even more preferable.

[0065] Since it is desirable that the sound sensor 152 does not detect any sound other than that generated during cutting, it is desirable to install it directly above the stripper 107. Specifically, because the punch 103 is nearby, it is desirable that the sound sensor 152 does not protrude from the top surface of the stripper 107. As for the sound sensor 152, due to the limited space above the stripper 107, a microphone or AE (acoustic emission) sensor with a diameter of 6 mm or less is preferable.

[0066] The position sensor 153 measures the amount of descent of the upper base 109 and, consequently, the punch stroke. It is desirable to install the position sensor 153 in a location that is less susceptible to vibrations generated during the cutting process. For this reason, it is desirable to install it inside the device cover 110. Specifically, it is desirable to install the position sensor 153 at a position 0.5 mm inward from the upper base 109 when the punch 103 is at its bottom dead center. As for the position sensor 153, considering that it measures the position of the descending member (upper base 109) and that the upper base 109 may be made of a non-metallic material, a non-contact capacitive sensor is preferable.

[0067] The temperature sensor 154 is preferably used to measure the processing temperature near the punch 103. For this reason, it is desirable that the temperature sensor 154 be embedded in the punch plate 111 so that its tip (detection end) faces the processing point (in other words, the die hole 102a). Specifically, in order to prevent contact with the stripper 107 during cutting, it is preferable to position the temperature sensor 154 such that the length of the tip of the temperature sensor 154 protruding from the lower surface of the punch plate 111 is 5 mm or less. Also, if the temperature sensor 154 is a radiation-type thermometer, it is desirable that the angle of the temperature sensor 154 with respect to the vertical direction be 10 degrees or less in order to accurately measure the processing temperature. Since the temperature sensor 154 is installed near the punch 103 in an extremely narrow space, it is necessary to prevent contact between the temperature sensor 154 and the punch 103 during cutting. For this reason, a radiation-type temperature sensor is preferable for the temperature sensor 154.

[0068] The cutting device 101 further includes a control unit 112. This control unit 112 functions as a learning device for the cutting evaluation system 1.

[0069] The control unit 112 will be described with reference to Figures 1 and 8. Figure 8 is a functional block diagram relating to the control unit 112 in the embodiment of this disclosure.

[0070] As shown in Figure 8, the control unit 112 includes a learning unit 104, a storage unit 105, and a calculation unit 106.

[0071] During cutting, the calculation unit 106 acquires state variables 12 based on information (physical quantities 15) obtained from the load sensor 151, sound sensor 152, position sensor 153, and temperature sensor 154, respectively. In other words, the calculation unit 106 functions as an input processing unit 2 as shown in Figure 1. The calculation unit 106 outputs a cutting evaluation result 13 using the evaluation model 11. In other words, the calculation unit 106 functions as an output processing unit 4. After cutting the workpiece, the state variables 12 and cutting result 16 input to the calculation unit 106 during cutting are sequentially stored in the storage unit 105 as a dataset. In other words, the storage unit 105 functions as a dataset group 14. The learning unit 104 performs learning of the evaluation model 11 based on all the datasets (dataset group 14) stored in the storage unit 105, and feeds back the learned evaluation model 11 to the calculation unit 106. In other words, the learning unit 104, storage unit 105, and calculation unit 106 cooperate to function as a learning processing unit 3.

[0072] Figure 9 is a flowchart showing the learning steps performed by the learning device 21 according to this embodiment. As shown in Figure 9, the learning processing unit 3 extracts a dataset from the dataset group 14 to which state variables 12 and truncation results 16 are associated (S1). The learning processing unit 3 uses the evaluation model 11 to identify the truncation evaluation results 13 for the state variables 12 included in the extracted dataset (S2). The learning processing unit 3 calculates an error from the truncation results 16 included in the extracted dataset and the identified truncation evaluation results 13 (S3). The learning processing unit 3 updates the evaluation model 11 based on the calculated error (S4). The learning processing unit 3 determines whether the calculated error has converged to the minimum (S5). If the error has not converged to the minimum (S5; No), the learning processing unit 3 executes the processes S1 to S4 again. On the other hand, if the error has converged to the minimum (S5; Yes), the learning processing unit 3 terminates the learning step.

[0073] Figure 10 is a flowchart showing the evaluation steps performed by the cutting process evaluation system 1 according to this embodiment. As shown in Figure 10, the input processing unit 2 acquires physical quantities 15 measured by each of the sensors 151 to 154 (S11). The input processing unit 2 generates state variables 12 based on the acquired physical quantities 15 (S12). The output processing unit 4 outputs the cutting evaluation result 13 obtained by inputting the state variables 12 into the evaluation model 11 (S13).

[0074] The cutting process evaluation system 1 of this disclosure comprises a sensor 150 for measuring physical quantities related to the cutting process, a learning device 21, and an output processing unit 4 that derives an evaluation using an evaluation model 11 updated by the learning device 21.

[0075] According to the learning device and cutting process evaluation system of this disclosure, the processing quality can be evaluated by an evaluation model using state variables based on physical quantities related to the cutting process, and the evaluation model can be updated. Therefore, the processing quality can be evaluated with high accuracy. [Industrial applicability]

[0076] According to this disclosure, by using an evaluation model to accurately evaluate machining quality and updating the evaluation model, it becomes possible to predict machining anomalies that are independent of skill level. Therefore, a reduction in the number of defects due to early response to anomalies and an improvement in productivity due to reduced equipment downtime are expected. [Explanation of symbols]

[0077] 1. Cutting Process Evaluation System 2 Input Processing Unit 3. Learning Processing Unit 4. Output Processing Unit 11 Evaluation Models 12 State Variables 13 Amputation Evaluation Results 14 datasets 15 Physical quantities 16 Cutting results 17 error 18. Loss Function 19 Optimization Algorithms 21 Learning device 51 Convolutional Neural Networks 52 Input Layers 53 Convolutional Layer 54 Pooling Layer 55 Fully connected layer 56 Output Layer 101 Cutting equipment 102 Die 103 Punch 104 Learning Department 105 Storage section 106 Calculation section 107 Strippers 108 Bass 109 Upper Base 110 Device cover 111 Punch Plate 112 Control Unit 150 sensors 151 Load Sensor 152 Sound Sensor 153 Position Sensor 154 Temperature Sensor

Claims

1. Input processing unit, It comprises a learning processing unit, The input processing unit is, Physical quantities related to the cutting process are acquired with a sampling period of 1 / 100th or less of the time required for the cutting process. The state variables based on the aforementioned physical quantities are input to the evaluation model of the learning processing unit. From the evaluation model, the cutting evaluation results are output, which are the results of evaluating whether or not there is an abnormality in the cutting process when the physical quantity is measured, and if there is an abnormality, the type of cause and the stepwise degree of the abnormality. The aforementioned learning processing unit, A dataset is accumulated for each processing step, in which the state variables based on the physical quantities measured during a single processing step are used as input data, and the actual processing results when those state variables were measured are used as output data. The error between the truncation evaluation result output from the evaluation model and the truncation result included in the dataset is calculated using a loss function. Based on the calculated error, the weight coefficients of the evaluation model are updated using an optimization algorithm. Learning device.

2. The aforementioned cutting process is a punching process in which the workpiece is punched out. The input processing unit uses the following as the physical quantity: The learning device according to claim 1, which acquires at least one of the load acting on the workpiece during punching, the shear rate during punching, the clearance between the punch and the die, and the temperature of the workpiece during punching.

3. A sensor that measures the aforementioned physical quantity, A learning device according to claim 1 or 2, A cutting process evaluation system comprising: an output processing unit that derives the cutting evaluation result using the evaluation model updated by the learning device; and

4. The aforementioned cutting process is a punching process in which the workpiece is punched out. As the aforementioned sensor, The workpiece is equipped with at least one load sensor for measuring the load applied during the punching process, a position sensor for measuring the position of the punch, and a sound sensor for measuring the sound generated by the punching process. The cutting process evaluation system according to claim 3.

Images