JP7646217B2 - Machine learning device, machine learning method, and mold maintenance timing prediction device - Google Patents

Description

The present invention relates to a machine learning device, a machine learning method, and a mold maintenance timing prediction device for use in predicting mold maintenance timing.

It has been proposed to set weights related to the degree of die wear depending on the material, plate thickness, and type of processing of the workpiece, and to calculate the number of times the die is used, which is proportional to the die wear, by taking into account the weights in the number of times the die is hit, and to know when the time to polish the die is reached when the number of times the die is used approaches a predetermined maximum number of times it can be used, which is the time when the die should be polished (for example, Patent Document 1).

Japanese Unexamined Patent Publication No. 61-74739

However, in Patent Document 1, the maximum usable number is uniquely determined, so depending on the setting, there is a risk that it may be determined that maintenance is due too soon, or that the appropriate maintenance time may be missed.

The present invention has been made in consideration of these circumstances, and provides a machine learning device, a machine learning method, and a mold maintenance timing prediction device for use in maintenance timing prediction that can predict the appropriate maintenance timing for a mold.

To solve this problem, the present invention provides a machine learning device that includes a data acquisition unit that acquires variation data of actual inspection values for each cumulative number of performance records of a mold, a maintenance record acquisition unit that acquires maintenance records of the mold corresponding to the variation data, and a learning unit that performs supervised learning using a set of the variation data acquired by the data acquisition unit and the maintenance records acquired by the maintenance record acquisition unit as training data, thereby constructing a learning model for predicting the timing of maintenance for the mold.

In the machine learning device, the variation data may indicate the degree of deviation of the trend of the actual inspection values for each cumulative number of results of the mold from the control reference value of the processed product toward either the upper or lower limit, and the maintenance record may indicate the timing of maintenance when maintenance was actually performed in relation to the degree of deviation of the trend.

In the machine learning device, the variation data may be calculated using actual measurements of the dimensions of the processed product input by the user and the number of times the mold has been used.

The present invention also provides a machine learning method including a data acquisition step of acquiring variation data of actual inspection values for each cumulative number of performance records of a mold, a maintenance record acquisition step of acquiring a maintenance record of the mold corresponding to the variation data, and a learning step of constructing a learning model for predicting the maintenance timing of the mold by performing supervised learning using a set of the variation data acquired in the data acquisition step and the maintenance record acquired in the maintenance record acquisition step as training data.

In the machine learning method, the variation data may indicate the degree of deviation of the trend of the actual inspection values for each cumulative number of results of the mold from the control reference value of the processed product toward either the upper or lower limit, and the maintenance record may indicate the timing of actual maintenance performed in relation to the deviation of the trend.

In the machine learning method, the variation data may be calculated using actual measurements of the dimensions of the processed product input by the user and the number of times the mold has been used.

The present invention also provides a mold maintenance timing prediction device that includes a mold identification information acquisition unit that acquires mold identification information of a mold, and an accumulated performance number acquisition unit that acquires an accumulated performance number of the mold, and applies a learning model constructed by the machine learning device to the accumulated performance number acquired by the accumulated performance number acquisition unit, and predicts the maintenance timing of the mold corresponding to the mold identification information acquired by the mold identification information acquisition unit.

The machine learning device, machine learning method, and mold maintenance timing prediction device of the present invention can detect signs of mold failure, prevent long-term production stoppages, and reduce mold repair costs.

In addition, by making maintenance schedules visible, the experience of experienced technicians can be passed on to less experienced technicians.

Furthermore, proper mold management through visualization like this will also contribute to achieving the SDGs ("8. Decent Work and Economic Growth", "9. Promote inclusive and sustainable industry, innovation and infrastructure", etc.).

1 is a block diagram showing a schematic configuration of a machine learning device and a mold maintenance timing prediction device according to a preferred embodiment of the present invention. FIG. 1 is a flowchart showing the flow of constructing a learning model according to a preferred embodiment of the present invention. FIG. 4 is an explanatory diagram showing variation data in a preferred embodiment of the present invention. FIG. 11 is an explanatory diagram showing an example of the degree of fluctuation in the tendency of variation data in a preferred embodiment of the present invention. FIG. 11 is an explanatory diagram showing another example of the degree of fluctuation in the tendency of variation data in the preferred embodiment of the present invention. 4 is a flowchart showing the operation of the mold maintenance timing prediction device according to the preferred embodiment of the present invention.

Below, preferred embodiments of the machine learning device, machine learning method, and mold maintenance timing prediction device of the present invention will be described with reference to the drawings. However, the present invention is not limited to the following description, and various modifications and changes are possible for those skilled in the art based on the gist of the invention described in the claims or disclosed in the description for implementing the invention. Such modifications and changes are also included in the scope of the present invention.

Figure 1 is a block diagram showing the schematic configuration of a machine learning device 1 and a mold maintenance timing prediction device 2 according to a preferred embodiment of the present invention. The machine learning device 1 includes a data acquisition unit 4 that acquires variation data 3 of actual inspection values for each cumulative number of performance records of a mold (not shown), a maintenance record acquisition unit 6 that acquires mold maintenance records 5 corresponding to the variation data 3, and a learning unit 7 that performs supervised learning using the set of the variation data 3 acquired by the data acquisition unit 4 and the maintenance records 5 acquired by the maintenance record acquisition unit 6 as teacher data to construct a learning model for predicting the maintenance timing of a mold.

The learning model constructed by the learning unit 7 is stored in the learning model storage unit 8. If new teacher data is acquired after the learning model is constructed, supervised learning is added to the learning model stored in the learning model storage unit 8, the previously constructed learning model is updated, and the new learning model is stored in the learning model storage unit 8.

During daily inspections, the user inputs and manages, for example, item code, item name, application area (inspection area), standard value of the dimension to be managed, lower limit value of the dimension to be managed, upper limit value of the dimension to be managed, actual measurement value of the processed product's dimension, average value of the actual measurement value, mold code, mold name, and the number of times the mold was used, into an input management system such as a master system connected for communication via a cloud server. Note that the same mold may be used even for different items.

The mold maintenance timing prediction device 2 includes a mold identification information acquisition unit 9 that acquires mold identification information of the mold, and an accumulated performance number acquisition unit 10 that acquires the accumulated performance number of the mold, and applies a learning model constructed by the machine learning device 1 to the accumulated performance number acquired by the accumulated performance number acquisition unit 10 to predict the maintenance timing of the mold corresponding to the mold identification information acquired by the mold identification information acquisition unit 9.

The mold maintenance timing prediction device 2 also includes a memory unit 11 that stores the prediction results, and a display unit 12 that displays the prediction results.

Figure 2 is a flowchart showing the flow of constructing a learning model according to a preferred embodiment of the present invention.

First, in step S11, the data acquisition unit 4 acquires the variation data 3. The data acquisition unit 4 outputs the variation data 3 to the learning unit 7. The variation data 3 indicates the degree to which the tendency of the inspection measurement value for each cumulative performance number deviates from the reference value of the dimension managed by the processed product toward either the upper or lower limit value.

Figure 3 shows the standard values and upper and lower limits of the dimensions to be managed within a histogram and probability distribution curve of the inspection measured values for each cumulative number of results. In addition to the peak of variation of the inspection measured values for each cumulative number of results approaching the standard value as shown in Figure 3, there are also cases where the peak of variation tends to fluctuate toward the lower limit as shown in Figure 4, and cases where the peak of variation tends to fluctuate toward the upper limit as shown in Figure 5.

The degree of deviation in the trend can be expressed numerically, for example, by taking the difference between the average value of the actual measurements or the value of the maximum point of the probability distribution curve and the reference value or upper and lower limits of the managed dimension. The sign of the difference makes it possible to determine whether the trend of the actual inspection values for each cumulative number of results in the variation data 3 deviates from the reference value of the managed dimension toward the upper or lower limit. The size of the difference makes it possible to determine to what extent the trend of the actual inspection values for each cumulative number of results in the variation data 3 deviates from the reference value of the managed dimension toward the upper or lower limit.

Each time a user inspects a processed product, the user inputs the actual measurements of the dimensions of the processed product and the number of times the mold has been used between the previous inspection and this inspection into the master system, which then calculates and records the difference and the cumulative number of inspections, and calculates variation data 3 as the degree of deviation in the inspection actual measurements for each cumulative number of inspections.

Next, in step S12, the maintenance record acquisition unit 6 acquires the maintenance record 5 of the mold corresponding to the variation data 3. The maintenance record acquisition unit 6 associates the maintenance record 5 with the variation data 3 and outputs it to the learning unit 7. Note that the two steps of step S11 and step S12 may be performed in parallel, with step S12 being performed first, or simultaneously.

The maintenance record 5 indicates, for example, the timing of when maintenance was actually performed in relation to the degree of deviation in the trend of the variation data 3. The maintenance timing is correlated with the degree of deviation in the trend of the variation data 3, and by using this correlation as training data to construct a learning model through supervised learning, it becomes possible to predict the appropriate timing of mold maintenance.

To explain the correlation in more detail, for example, when the peak of the variation shows a tendency to fluctuate toward the lower limit or toward the upper limit due to an increase in the cumulative number of results as described above, there is a correlation that maintenance will be performed more frequently. By collecting a large amount of data on the degree of deviation in the trend and the corresponding data on the timing of maintenance, it becomes possible to determine, for example, the degree of deviation that results in a large number of maintenances.

To apply such correlations to supervised machine learning, for example, as described above, the degree of deviation in the trend is expressed numerically as the difference between the average of the actual measurements or the maximum value of the probability distribution curve and the reference value or upper and lower limits of the dimension being managed, and the relationship between this numerical value and the maintenance timing is used as the training data.

The maintenance record 5 is entered into the master system by the user each time maintenance is performed, and is recorded in association with other data and variation data 3. It is preferable that the timing of maintenance used for the training data is determined by the user based on experience, etc.

Next, in step S13, the learning unit 7 generates training data by pairing each of the variation data 3 and the corresponding maintenance records 5.

Next, in step S14, the learning unit 7 performs machine learning based on the teacher data generated in step S13. This machine learning is supervised learning.

Next, in step S15, the learning unit 7 determines whether or not to end the machine learning. The condition for determining whether to end the machine learning can be, for example, that supervised learning has been performed a predetermined number of times.

If the conditions for terminating the machine learning are not met, the process returns to step S11 and the machine learning is repeated. If the conditions for terminating the machine learning are met, the process proceeds to step S16 and the constructed learning model is stored and saved in the learning model storage unit 8.

Figure 6 is a flowchart showing the operation of the mold maintenance timing prediction device 2 according to a preferred embodiment of the present invention.

First, in step S21, the mold maintenance timing prediction device 2 acquires mold identification information of the mold for which the maintenance timing is to be predicted, for example, from an input by a user, via the mold identification information acquisition unit 9. As the mold identification information, for example, a mold code, a mold name, etc. can be used.

Next, in step S22, the mold maintenance timing prediction device 2 acquires the accumulated number of records of the mold corresponding to the mold identification information acquired in step S21 by the accumulated number acquisition unit 10. The accumulated number of records of the mold can be acquired, for example, based on the number of records of the mold being used that was input to the master system described above. Alternatively, the accumulated number of records of the mold may be input by the user. The two steps of step S21 and step S22 may be performed in parallel, with step S22 being performed first, or simultaneously.

Next, in step S23, the mold maintenance timing prediction device 2 inputs the mold identification information and the cumulative number of mold performance records, and predicts the mold maintenance timing based on the learning model constructed by the above method.

Finally, in step S24, the mold maintenance timing prediction device 2 stores the prediction results from step S23 in the memory unit 11.

Supervised learning can use, for example, analytical methods such as regression analysis (the method of least squares) and decision trees, or it can be ensemble learning that combines a number of these analytical methods. In addition, the machine learning device 1 can detect features by itself using a neural network such as deep learning with an intermediate layer, allowing it to produce complex output that more closely matches the teacher data.

The machine learning device 1 can be realized, for example, by installing an application software program on a personal computer or a server device. That is, the machine learning device 1 includes at least an arithmetic processing device such as a CPU (Central Processing Unit) and a storage device that stores the application software program. Since the amount of calculations required for supervised learning is large, the machine learning device 1 may further include, for example, a GPU (Graphics Processing Unit).

The mold maintenance timing prediction device 2 can be realized by installing an application software program on, for example, a personal computer or a server device. That is, the mold maintenance timing prediction device 2 includes at least an arithmetic processing device such as a CPU (Central Processing Unit) and a storage device for storing the application software program. The mold maintenance timing prediction device 2 predicts the mold maintenance timing by calculation of an application software program that inputs mold identification information, the cumulative number of mold records, and the learning model constructed by the machine learning device 1, and outputs and displays the prediction result on a display unit 12 such as a display. The user can predict the timing of mold maintenance by inputting the actual measurement values of the dimensions of the processed product and the number of mold uses at each inspection, which can prevent long-term production stoppages due to sudden mold damage and reduce repair costs through regular maintenance. Preventing long-term production stoppages can reduce sales opportunity losses and the time required to re-plan production.

Reference Signs List 1 Machine learning device 2 Mold maintenance timing prediction device 3 Variation data 4 Data acquisition unit 5 Maintenance record 6 Maintenance record acquisition unit 7 Learning unit 8 Learning model storage unit 9 Mold identification information acquisition unit
10. Cumulative Achievement Number Acquisition Section
11 Memory section
12 Display unit

Claims (7)

a data acquisition unit that acquires variation data of actual inspection values for each cumulative number of molds;
a maintenance record acquisition unit that acquires a maintenance record of the mold corresponding to the variation data;
a learning unit that performs supervised learning using a set of the variation data acquired by the data acquisition unit and the maintenance record acquired by the maintenance record acquisition unit as teacher data to construct a learning model for predicting the timing of maintenance of the mold;
A machine learning device comprising: The machine learning device according to claim 1, wherein the variation data indicates the degree of deviation of the trend of the actual inspection values for each cumulative number of results of the mold from the control reference value of the processed product toward either the upper or lower limit, and the maintenance record indicates the timing of maintenance when maintenance was actually performed with respect to the degree of deviation of the trend. The machine learning device according to claim 1, wherein the variation data is calculated using actual measurements of dimensions of processed products input by a user and the number of times the mold has been used. A machine learning method using the machine learning device according to any one of claims 1 to 3,
a data acquisition step in which the data acquisition unit acquires variation data of the inspection measured values for each cumulative number of results of the mold;
a maintenance record acquisition step in which the maintenance record acquisition unit acquires a maintenance record of the mold corresponding to the variation data;
a learning step in which the learning unit performs supervised learning using a set of the variation data acquired in the data acquisition step and the maintenance record acquired in the maintenance record acquisition step as teacher data, thereby constructing a learning model for predicting the maintenance timing of the mold;
Machine learning methods, including The machine learning method according to claim 4, wherein the variation data indicates the degree of deviation of the trend of the actual inspection measurement value for each cumulative number of results of the mold from the control reference value of the processed product toward either the upper or lower limit, and the maintenance record indicates the timing of maintenance when maintenance was actually performed with respect to the degree of deviation of the trend. The machine learning method according to claim 4, wherein the variation data is calculated using actual measurements of dimensions of processed products input by a user and the number of times the mold has been used. a mold identification information acquisition unit for acquiring mold identification information of a mold;
A cumulative performance number acquisition unit for acquiring a cumulative performance number of the mold,
A mold maintenance timing prediction device that applies a learning model constructed by the machine learning device described in any one of claims 1 to 3 to the accumulated performance count acquired by the accumulated performance count acquisition unit, and predicts the maintenance timing of the mold corresponding to the mold identification information acquired by the mold identification information acquisition unit.

Images