Inference device, machine learning device, and processing system

Abstract

An inference device (3) comprises: a data acquisition unit (35) that acquires data for inference including processing condition time-series data including time-series data indicating transitions of a processing condition applied to a processing machine up to a specific time point and processing conditions to be used in the next processing, and specific point-of-time degradation degree data (D14) indicating transitions of a degree of degradation of a consumable item used in a processing device up to a specific time point; and an inference unit (36) that uses a learning model for inferring next degradation degree data indicating the degree of degradation of the consumable item at the time of completion of the next processing from the processing condition time-series data and the specific point-of-time degradation degree data (D14), to infer the next degradation degree data from the data for inference having been input from the data acquisition unit (35), and outputs the next degradation degree data.

Description

Inference device, machine learning device, and processing system

[0001] The present disclosure relates to an inference device, a machine learning device, and a processing system that infer the degree of deterioration of consumables used by a processing machine.

[0002] In machining machines such as wire electric discharge machines, workpieces are machined using consumables such as machining fluid permeable filters, and it is desirable to predict the degree of deterioration of these consumables and replace them at an appropriate time based on the predicted degree of deterioration.

[0003] The machine learning device described in Patent Document 1 learns the deterioration level of consumables used in a machining machine through machine learning. This machine learning device generates a learning model through supervised learning using the machining conditions of electric discharge machining and the deterioration level of the consumables after electric discharge machining under these machining conditions. A prediction device then inputs the electric discharge machining conditions into the generated learning model and predicts the deterioration level of the consumables after electric discharge machining under these machining conditions.

[0004] International Publication No. 2021 / 177237

[0005] However, the technology of Patent Document 1 does not take into account the transition of processing conditions that affect the deterioration level of the consumables, and only learns the deterioration level of the consumables corresponding to the processing conditions used in one processing. Therefore, the technology of Patent Document 1 has a problem in that it cannot accurately infer the deterioration level of the consumables.

[0006] The present disclosure has been made in view of the above, and aims to provide an inference device that can accurately infer the degree of deterioration of a consumable item.

[0007] In order to solve the above-mentioned problems and achieve the object, the inference device of the present disclosure includes a data acquisition unit that acquires inference data including processing condition time-series data including time-series data indicating the transition of processing conditions applied to a processing machine up to a specific point in time and processing conditions to be used in the next processing, and specific time degradation level data indicating the transition of the deterioration level of a consumable used by the processing machine up to a specific point in time. The inference device of the present disclosure also includes an inference unit that infers and outputs next degradation level data from the inference data input from the data acquisition unit using a learning model for inferring next degradation level data indicating the deterioration level of the consumable at the time the next processing is completed from the processing condition time-series data and the specific time degradation level data.

[0008] The inference device according to the present disclosure has the effect of being able to accurately infer the degree of deterioration of a consumable item.

[0009] FIG. 1 is a diagram showing the configuration of a learning inference system according to an embodiment; FIG. 2 is a diagram for explaining a neural network used by a machine learning device according to an embodiment; FIG. 3 is a flowchart showing the processing procedure of a learning process executed by a machine learning device according to an embodiment; FIG. 4 is a flowchart showing the processing procedure of an inference process executed by an inference device according to an embodiment; FIG. 5 is a diagram for explaining input data and output data used for learning by a machine learning device according to an embodiment;

[0010] An inference device, a machine learning device, and a processing system according to embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.

[0011] 1 is a diagram showing the configuration of a learning inference system according to an embodiment. The learning inference system 1 includes a machine learning device 2, an inference device 3, and a learning model storage unit 5. The inference device 3 is a lifespan prediction device that predicts the lifespan of a consumable by inferring the deterioration level (deterioration degree) of the consumable. The lifespan of a consumable is the time when replacement of the consumable is recommended, and corresponds to the deterioration level of the consumable.

[0012] The inference device 3 predicts the lifespan of consumables used in machining machines such as wire electric discharge machines, die-sinking electric discharge machines, small-hole electric discharge machines, and laser machining machines. Consumables in wire electric discharge machines include wire electrode wires, filtration filters, ion exchange resins, power supply dies, and electrode wire guide rollers. Consumables in die-sinking electric discharge machines include tool electrodes and filtration filters. Consumables in small-hole electric discharge machines include pipe electrodes and filtration filters. Consumables in laser machining machines include dust collection filters, machining lenses, and bend mirrors.

[0013] In the following, we will explain the case where the machining machine for which the inference device 3 predicts the deterioration level of consumables is a wire electric discharge machine, but the machining machine for which the inference device 3 predicts the deterioration level of consumables may be any type of machining machine.

[0014] In the following, an example of the deterioration level of a consumable item will be described, in which the pressure loss of a machining fluid filter in a wire electric discharge machine is taken into consideration. The filter is a machining fluid permeable filter that filters the machining fluid by blocking machining debris that has become mixed in the machining fluid and allowing the machining fluid to pass through.

[0015] In wire electrical discharge machining (EDM) using a wire-cut electrical discharge machine, machining fluid is used to remove machining debris generated during machining. Any machining debris mixed into the machining fluid is removed by a filter, allowing the machining fluid to be reused. When machining fluid passes through a filter, a differential pressure occurs between the primary side (upstream side) and secondary side (downstream side) of the filter. The value of this differential pressure is the pressure loss of the filter. When machining debris accumulates on the filter, the pressure loss increases, and if the pressure loss exceeds a certain value, the filter will burst. Therefore, the filter must be replaced before it bursts. The pressure loss of the filter corresponds to the degree of deterioration of the filter, which is a consumable item.

[0016] In the embodiment, learning and inference system 1 learns the deterioration level of the filtration filter based on the pressure loss of the filtration filter, and infers the deterioration level of the filtration filter when the next processing is completed.

[0017] In the learning and inference system 1, a machine learning device 2 generates a learning model by learning the deterioration degree of a filtration filter, and an inference device 3 infers the deterioration degree of the filtration filter using the learning model.

[0018] The learning model generated by the machine learning device 2 is a learning model for inferring the degree of deterioration of the filtration filter when the next machining is completed, from time series data showing the changes in machining conditions, which are the machining conditions applied to the wire electric discharge machine, and time series data showing the changes in the accumulated amount of machining waste, which is an example of a physical quantity of a consumable that has a significant impact on the degree of deterioration of the consumable.

[0019] The machine learning device 2 is disposed in a machining machine such as a wire electric discharge machine. Note that the machine learning device 2 may also be disposed outside the wire electric discharge machine. The machine learning device 2 includes a schedule creation unit 21, a calculation unit 22, a measurement unit 23, a deterioration level acquisition unit 24, a data acquisition unit 20, and a model generation unit 27. The data acquisition unit 20 includes an input data acquisition unit 25 and a label acquisition unit 26.

[0020] The schedule creation unit 21 acquires the planned machining condition data D1 by creating data of planned machining conditions including a planned machining speed and a planned machining time (hereinafter referred to as planned machining condition data D1). The planned machining speed is the planned machining speed in wire electric discharge machining. The planned machining time is the planned machining time in wire electric discharge machining.

[0021] The schedule creation unit 21 may create the planned machining condition data D1, for example, based on a machining program used to control the wire electric discharge machine, or may create the planned machining condition data D1 based on machining conditions set by the user in the wire electric discharge machine.

[0022] The schedule creation unit 21 may create planned machining condition data D1 including a machining path length instead of the planned machining condition data D1 including a planned machining speed and a planned machining time. The schedule creation unit 21 may also create time-series data of planned machining conditions (planned machining conditions) instead of the planned machining condition data D1. In this case, the planned machining condition data D1 includes data on machining conditions from the past up to a specific point in time and data on machining conditions scheduled for the next execution.

[0023] The schedule creation unit 21 sends the planned machining condition data D1 to the calculation unit 22 and the input data acquisition unit 25. The planned machining condition data D1 sent by the schedule creation unit 21 to the input data acquisition unit 25 may include the surface finish accuracy of the workpiece. The surface finish accuracy is the finishing accuracy of the surface of the workpiece set for wire electric discharge machining.

[0024] The measurement unit 23 acquires time-series data (hereinafter referred to as measurement result time-series data D3) of measurement data (measurement results) including the machining speed and machining time actually measured during wire electric discharge machining. The measurement result time-series data D3 is data of actual measurements (measurement values) acquired during each machining operation from the past up to a specific point in time. The specific point in time is, for example, the present time. Note that the specific point in time is not limited to the present time, but may be a specific point in the recent past or a point in the past slightly earlier than the present time. The measurement unit 23 acquires the measurement result time-series data D3 from the wire electric discharge machine.

[0025] The measurement data included in the measurement result time-series data D3 is time-series data that indicates the transition of machining conditions such as machining speed and machining time. The machining conditions (machining speed and machining time) included in the measurement result time-series data D3 are actual machining conditions, and the machining conditions (machining speed and machining time) included in the planned machining condition data D1 are planned machining conditions.

[0026] Note that when the schedule creation unit 21 creates time-series data of planned machining conditions, the machine learning device 2 does not need to include the measurement unit 23. The machine learning device 2 may use the planned machining speed and planned machining time included in the time-series data of planned machining conditions created by the schedule creation unit 21, or may use the machining speed and machining time measured by the measurement unit 23. That is, the machine learning device 2 may use the planned machining conditions as the machining conditions from the past to a specific point in time, or may use the actual machining conditions measured by the measurement unit 23. The machining condition time-series data is data including time-series data indicating the transition of the machining conditions from the past to a specific point in time and the machining conditions to be used in the next machining.

[0027] The measurement data may include data on the temperature during wire electric discharge machining, data on the surface finish accuracy, etc. The measurement unit 23 sends the measurement result time-series data D3 to the calculation unit 22 and the input data acquisition unit 25.

[0028] The calculation unit 22 calculates time-series data of calculation data (calculation results) (hereinafter referred to as calculation result time-series data D2) based on the processing conditions included in the planned processing condition data D1 acquired by the schedule creation unit 21 and the processing conditions (measurement data) included in the measurement result time-series data D3 acquired by the measurement unit 23. The calculation data acquired by the calculation unit 22 through calculation is, for example, the accumulated processing waste amount (accumulated amount of processing waste). The calculation result time-series data D2 is an example of physical quantity time-series data that indicates the transition of the physical quantity of a consumable up to a specific point in time.

[0029] The cumulative amount of machining debris is the cumulative amount of machining debris that has been mixed into the machining fluid since the filter was replaced and the new filter was started to be used. The cumulative amount of machining debris that has been mixed into the machining fluid corresponds to the cumulative amount of machining debris that has been filtered by the filter.

[0030] The calculation result time series data D2 includes calculation data of the accumulated machining waste amount from the past to a specific time point and calculation data of the accumulated machining waste amount after the next machining scheduled to be performed is completed. The accumulated machining waste amount from the past to a specific time point is not just the accumulated machining waste amount at the specific time point, but also the transition of the accumulated machining waste amount from the past to the specific time point.

[0031] The calculation unit 22 calculates the cumulative amount of machining waste, for example, from the product of the machining depth calculated from the surface finishing accuracy, the machining speed, and the machining time. When the calculation unit 22 uses the surface finishing accuracy, data on the surface finishing accuracy is included in the planned machining condition data D1 or the measurement result time-series data D3. Note that the calculation unit 22 may also calculate the cumulative amount of machining waste from the product of the machining speed and the machining time without using the surface finishing accuracy.

[0032] When the surface finish accuracy is included in the planned machining condition data D1, the surface finish accuracy is the planned surface finish accuracy specified in the machining program, etc. When the surface finish accuracy is included in the measurement result time-series data D3, the surface finish accuracy is the surface finish accuracy actually measured after machining.

[0033] The calculation unit 22 uses the machining speed and machining time acquired from the measurement unit 23 as the machining speed and machining time from the past up to a specific point in time. Furthermore, the calculation unit 22 uses the planned machining speed and planned machining time acquired from the schedule creation unit 21 as the machining speed and machining time for the next machining.

[0034] The accumulated amount of processing waste calculated by the calculation unit 22 increases as the number of processing days increases. The calculation unit 22 calculates the time-series data of the accumulated amount of processing waste as calculation result time-series data D2. The calculation unit 22 sends the calculation result time-series data D2 to the input data acquisition unit 25.

[0035] The machining speed and machining time from the past to a specific point in time are, for example, the machining speed and machining time applied between the (T-N)th day and the Tth day (specific point in time), where T and N (T>N) are natural numbers. The next machining speed and machining time are, for example, the planned machining speed and planned machining time applied between the Tth day and the (T+1)th day. The cumulative machining waste amount from the past to a specific point in time is the cumulative machining waste amount applied between the (T-N)th day and the Tth day. The next cumulative machining waste amount is, for example, the cumulative machining waste amount at the time when machining applied between the Tth day and the (T+1)th day is completed. The specific point in time, the Tth day, is the time when machining on the Tth day is completed, and the (T+1)th day is the time when machining on the (T+1)th day is completed. Therefore, the specific point in time (present) in this embodiment is the time when the Tth day ends.

[0036] In this embodiment, the case where the deterioration level of the filtration filter is learned and inferred on a daily basis will be described, but the deterioration level of the filtration filter may also be learned and inferred on an hourly or minutely basis.

[0037] It is assumed that the processing from the (T-N)th day to the (T+1)th day has been performed in advance before the machine learning device 2 generates the learning model. In other words, the machine learning device 2 generates the learning model after the processing from the (T-N)th day to the (T+1)th day has been performed.

[0038] The machine learning device 2 acquires in advance the processing speed and processing time from the (T-N)th day to the Tth day (specific time point) and the planned processing speed and planned processing time for the (T+1)th day. The reason why the processing speed and processing time for the (T+1)th day are the planned processing speed and planned processing time is that when the deterioration level of the filtration filter is inferred, the actual processing speed and processing time have not been measured at the time of inference. In other words, since the deterioration level of the filtration filter is inferred using the planned processing speed and planned processing time during inference, the machine learning device 2 also learns the deterioration level of the filtration filter using the planned processing speed and planned processing time during learning.

[0039] The deterioration level acquiring unit 24 acquires time-series data (hereinafter referred to as "specific time-point deterioration level data D4") of the deterioration level of the filtration filter after each processing from the past to a specific time point. The deterioration level from the past to a specific time point is the deterioration level from the (T-N)th day to the Tth day. When the deterioration level of the consumable item is the deterioration level of the filtration filter, this deterioration level is the pressure loss of the filtration filter.

[0040] The deterioration level acquiring unit 24 also acquires time-series data (hereinafter referred to as next deterioration level data D5) of the deterioration level of the filtration filter at the time when the next processing is completed. The next deterioration level is, for example, the deterioration level at the time when processing on the (T+1)th day is completed.

[0041] The degradation level acquiring unit 24 sends the specific time degradation level data D4 and the next degradation level data D5 to the data acquiring unit 20. That is, the degradation level acquiring unit 24 sends the specific time degradation level data D4 to the input data acquiring unit 25 and sends the next degradation level data D5 to the label acquiring unit 26.

[0042] The input data acquisition unit 25 acquires planned processing condition data D1 from the schedule creation unit 21, acquires measurement result time series data D3 from the measurement unit 23, acquires calculation result time series data D2 from the calculation unit 22, and acquires specific point-in-time deterioration degree data D4 from the deterioration degree acquisition unit 24.

[0043] In this manner, the input data acquiring unit 25 acquires the machining conditions, measurement data, cumulative machining waste amount, and pressure loss of the filtration filter. The input data acquiring unit 25 acquires data from the (T-N)th day to the (T+1)th day from the schedule creating unit 21, the calculation unit 22, the measurement unit 23, and the deterioration level acquiring unit 24. Specifically, the input data acquiring unit 25 acquires the machining conditions for the (T+1)th day from the schedule creating unit 21, and acquires the machining conditions from the (T-N)th day to the Tth day from the measurement unit 23. The input data acquiring unit 25 also acquires the cumulative machining waste amount from the (T-N)th day to the (T+1)th day from the calculation unit 22, and acquires the deterioration level (pressure loss of the filtration filter) from the (T-N)th day to the Tth day from the deterioration level acquiring unit 24.

[0044] The input data acquiring unit 25 sends the planned processing condition data D1, the measurement result time series data D3, the calculation result time series data D2, and the point-in-time degradation degree data D4 to the model generating unit 27 as input data.

[0045] The label acquisition unit 26 acquires the next degradation level data D5, which is the pressure loss of the filtration filter on the (T+1)th day, from the degradation level acquisition unit 24, and sends the next degradation level data D5 to the model generation unit 27 as label data (result data).

[0046] In this way, the data acquisition unit 20 acquires learning data including planned processing condition data D1, measurement result time series data D3, calculation result time series data D2, specific point-in-time deterioration degree data D4, and next deterioration degree data D5.

[0047] Model generation unit 27 learns the pressure loss of the filtration filter based on learning data created based on a combination of input data sent from input data acquisition unit 25 and label data sent from label acquisition unit 26. That is, model generation unit 27 learns the pressure loss of the filtration filter based on learning data created based on the processing conditions, measurement data, and calculation data (accumulated processing waste amount) output from input data acquisition unit 25 and the deterioration level (pressure loss of the filtration filter) acquired by deterioration level acquisition unit 24. Model generation unit 27 learns the pressure loss of the filtration filter based on learning data created based on a combination of processing data from the past up to a specific point in time and data for the next processing. Here, the learning data is data in which processing data from the past up to a specific point in time and data for the next processing are associated with each other.

[0048] The model generation unit 27 generates a learning model based on the processing conditions from the past to a specific time point and the next time, the accumulated amount of processing waste from the past to a specific time point and at the time of completion of the next processing, and the pressure loss of the filtration filter from the past to a specific time point and at the time of completion of the next processing, which are received from the input data acquisition unit 25 and the label acquisition unit 26.

[0049] In this way, model generation unit 27 generates a learning model that predicts the pressure loss of the filtration filter with high accuracy from the machining conditions of the wire electric discharge machine, the accumulated amount of machining waste, and the pressure loss of the filtration filter.

[0050] The learning model generated by the model generation unit 27 is a learning model for inferring the next deterioration level at the time of completion of the next processing from the processing conditions from the past to the specific point in time, the calculation data from the past to the specific point in time, the measurement data from the past to the specific point in time, the specific point in time deterioration level from the past to the specific point in time, the next processing conditions, and the next calculation data.In this way, the learning model generated by the model generation unit 27 is a learning model for predicting the deterioration level of the consumable after processing under the processing conditions scheduled to be performed next time.

[0051] For example, if T is 100 and N is 30, the learning model learns the correspondence between the data from the 70th to 100th days and the data from the 101st day that corresponds to the data from the 70th to 100th days.

[0052] The model generation unit 27 stores the generated learning model in the learning model storage unit 5. The learning model storage unit 5 may be located inside the machine learning device 2 or outside the machine learning device 2. The learning model storage unit 5 may also be located inside the inference device 3. The learning model storage unit 5 may also be located both inside the machine learning device 2 and inside the inference device 3.

[0053] The learning data used by the model generation unit 27 does not necessarily have to include the calculation data (accumulated amount of processing waste). In this case, the input data acquisition unit 25 does not have to acquire the accumulated amount of processing waste. Furthermore, the machine learning device 2 does not have to include the calculation unit 22.

[0054] The inference device 3 is disposed in a machining device such as a wire electric discharge machine. The inference device 3 may also be disposed outside the wire electric discharge machine. The inference device 3 includes a schedule creation unit 31, a calculation unit 32, a measurement unit 33, a deterioration level acquisition unit 34, a data acquisition unit 35, and an inference unit 36.

[0055] The schedule creation unit 31 has the same function as the schedule creation unit 21, and the calculation unit 32 has the same function as the calculation unit 22. Furthermore, the measurement unit 33 has the same function as the measurement unit 23, and the deterioration level acquisition unit 34 has the same function as the deterioration level acquisition unit 24.

[0056] The schedule creation unit 31 acquires the planned machining condition data D11 by creating data of planned machining conditions including a planned machining speed and a planned machining time (hereinafter referred to as planned machining condition data D11). The planned machining condition data D11 is the same data as the planned machining condition data D1.

[0057] The schedule creation unit 31 may create the planned machining condition data D11, for example, based on a machining program used to control the wire electric discharge machine, or may create the planned machining condition data D11 based on machining conditions set by the user in the wire electric discharge machine.

[0058] The schedule creating unit 31 may create scheduled processing condition data D11 including a processing path length instead of the scheduled processing condition data D11 including the scheduled processing speed and the scheduled processing time. The schedule creating unit 31 may also create time-series data of the scheduled processing condition data D11.

[0059] When the schedule creation unit 31 creates the machining condition time-series data, the inference device 3 does not need to include the measurement unit 33. The inference device 3 may use the planned machining speed and planned machining time included in the machining condition time-series data created by the schedule creation unit 31, or may use the machining speed and machining time measured by the measurement unit 33.

[0060] The schedule creation unit 31 sends the planned machining condition data D11 to the calculation unit 32 and the data acquisition unit 35. The planned machining condition data D11 sent by the schedule creation unit 31 to the data acquisition unit 35 may include the surface finish accuracy of the workpiece.

[0061] The measurement unit 33 acquires time-series data (hereinafter referred to as measurement result time-series data D13) of measurement data (measurement results) including machining speeds and machining times actually measured during wire electric discharge machining from the past to a specific point in time. The measurement result time-series data D13 is the same data as the measurement result time-series data D3.

[0062] The measurement result time-series data D13 is measurement data during inference, and the measurement result time-series data D3 is measurement data during learning. The measurement unit 33 acquires the measurement result time-series data D13 from the wire electric discharge machine. Note that the measurement data may include data on the temperature during wire electric discharge machining, data on surface finish accuracy, etc. The measurement unit 33 sends the measurement result time-series data D13 to the calculation unit 32 and the data acquisition unit 35.

[0063] The calculation unit 32 calculates time-series data of calculation data (calculation results) (hereinafter referred to as calculation result time-series data D12) based on the processing conditions included in the planned processing condition data D11 acquired by the schedule creation unit 31 and the processing conditions (measurement data) included in the measurement result time-series data D13 acquired by the measurement unit 33. The calculation result time-series data D12 is the same data as the calculation result time-series data D2.

[0064] Similar to the calculation unit 22, the calculation unit 32 calculates, for example, the amount of machining waste from the product of the machining depth calculated from the surface finish accuracy, the machining speed, and the machining time. When the calculation unit 32 uses the surface finish accuracy, time-series data of the surface finish accuracy is included in the planned machining condition data D11 or the measurement result time-series data D13. Note that the calculation unit 32 may also calculate the cumulative amount of machining waste from the product of the machining speed and the machining time without using the surface finish accuracy.

[0065] When the surface finish accuracy is included in the planned machining condition data D11, the surface finish accuracy is the planned surface finish accuracy specified in the machining program, etc. When the surface finish accuracy is included in the measurement result time-series data D13, the surface finish accuracy is the surface finish accuracy actually measured after machining.

[0066] The calculation unit 32 uses the machining speed and machining time acquired from the measurement unit 33 as the machining speed and machining time from the past up to a specific point in time. Furthermore, the calculation unit 32 uses the planned machining speed and planned machining time acquired from the schedule creation unit 31 as the machining speed and machining time for the next machining.

[0067] The accumulated amount of processing waste calculated by the calculation unit 32 increases as the number of processing days increases. The calculation unit 32 calculates time-series data of the accumulated amount of processing waste as calculation result time-series data D12. The calculation unit 32 sends the calculation result time-series data D12 to the data acquisition unit 35.

[0068] The machining speed and machining time from the past to a specific time point are, for example, the machining speed and machining time from the (t-n)th day to the tth day (specific time point), where t and n (t>n) are natural numbers. The next machining speed and machining time are, for example, the planned machining speed and planned machining time for the (t+1)th day. The cumulative machining waste amount from the past to a specific time point is the cumulative machining waste amount from the (t-n)th day to the tth day. The next cumulative machining waste amount is, for example, the cumulative machining waste amount at the time when machining on the (t+1)th day is completed.

[0069] It is assumed that the processing from the (t-n)th day to the tth day has been executed in advance by the time the inference device 3 infers the deterioration level of the filtration filter after the next processing. That is, the inference device 3 infers the deterioration level of the filtration filter on the (t+1)th day after the processing from the (t-n)th day to the tth day has been executed. The tth day is, for example, the present, and the (t+1)th day is tomorrow.

[0070] The deterioration level acquisition unit 34 acquires time-series data (hereinafter referred to as "specific time-point deterioration level data D14") of the deterioration level of the consumable (filtration filter) after each processing from the past to a specific point in time. The deterioration level from the past to a specific point in time is the deterioration level from the (t-n)th day to the tth day. The deterioration level acquisition unit 34 sends the specific time-point deterioration level data D14 to the data acquisition unit 35.

[0071] The data acquiring unit 35 acquires the planned machining condition data D11 from the schedule creating unit 31, acquires the measurement result time-series data D13 from the measuring unit 33, acquires the calculation result time-series data D12 from the calculation unit 32, and acquires the specific time point deterioration degree data D14 from the deterioration degree acquiring unit 34. In this way, the data acquiring unit 35 acquires the machining conditions, the measurement data, the accumulated amount of machining waste, and the pressure loss of the filtration filter from the past to the specific time point.

[0072] The data acquiring unit 35 acquires data from the (t-n)th day to the (t+1)th day from the schedule creating unit 31, the calculation unit 32, the measurement unit 33, and the deterioration degree acquiring unit 34. Specifically, the data acquiring unit 35 acquires the processing conditions for the (t+1)th day from the schedule creating unit 31, and acquires the processing conditions from the (t-n)th day to the tth day from the measurement unit 33. The data acquiring unit 35 also acquires the accumulated amount of processing waste from the (t-n)th day to the (t+1)th day from the calculation unit 32, and acquires the deterioration degree (pressure loss of the filtration filter) from the (t-n)th day to the tth day from the deterioration degree acquiring unit 34.

[0073] The data acquisition unit 35 included in the inference device 3 is a first data acquisition unit, and the data acquisition unit 20 included in the machine learning device 2 is a second data acquisition unit. The data acquisition unit 35 sends inference data including the planned processing condition data D11, the measurement result time-series data D13, the calculation result time-series data D12, and the point-in-time degradation degree data D14 to the inference unit 36 ​​as input data.

[0074] The inference unit 36 ​​reads out the learning model from the learning model storage unit 5. The inference unit 36 ​​applies the inference data, which is input data, to the learning model to calculate an inference result corresponding to the input data. That is, the inference unit 36 ​​infers the deterioration level of the filtration filter on the (t+1) day corresponding to the processing conditions and accumulated amount of processing waste from the (t-n)th day to the (t+1)th day, and the deterioration level of the filtration filter from the (t-n)th day to the t day.

[0075] t is a value greater than T. For example, when t is 101 and n is 30, the inference unit 36 ​​uses the learning model to infer the data for the 102nd day, which corresponds to the data from the 70th day to the 101st day. When t is 200 and n is 50, the inference unit 36 ​​uses the learning model to infer the data for the 201st day, which corresponds to the data from the 150th day to the 200th day.

[0076] In this way, the inference unit 36 ​​predicts the pressure loss of the filtration filter obtained using the learning model and outputs inference data that is the prediction result. That is, by inputting the processing conditions, accumulated processing waste amount, and deterioration degree of the filtration filter acquired from the data acquisition unit 35 into the learning model, the inference unit 36 ​​can infer next deterioration degree data D15 that is the deterioration degree (pressure loss) of the filtration filter after the next processing is completed, and output the next deterioration degree data D15 that is the inference data.

[0077] The inference data used by the inference unit 36 ​​does not necessarily have to include the calculation data (accumulated amount of processing waste). In this case, the data acquisition unit 35 does not have to acquire the accumulated amount of processing waste. Furthermore, the inference device 3 does not have to include the calculation unit 32.

[0078] The machine learning device 2 and the inference device 3 are used to learn the pressure loss of the filtration filter of the processing machine, but may be devices connected to the processing machine via a network and separate from the processing machine. The machine learning device 2 and the inference device 3 may also be built into the processing machine. Furthermore, the machine learning device 2 and the inference device 3 may reside on a cloud server.

[0079] The learning algorithm used by the model generation unit 27 may be a known algorithm such as supervised learning, unsupervised learning, reinforcement learning, etc. As an example, a case where a neural network is applied to the model generation unit 27 will be described.

[0080] The model generation unit 27 learns the pressure loss of the filtration filter by, for example, supervised learning in accordance with a neural network model. Here, supervised learning refers to a method of providing a learning device with sets of data, each set consisting of an input and a result (label), learning the features of the learning data, and predicting a result from the input.

[0081] A neural network consists of an input layer consisting of multiple neurons, an intermediate layer (hidden layer) consisting of multiple neurons, and an output layer consisting of multiple neurons. The intermediate layer may be one layer, or two or more layers.

[0082] 2 is a diagram illustrating a neural network used by a machine learning device according to an embodiment. For example, in a three-layer neural network as shown in FIG. 2, when multiple inputs are input to the input layer (X1 to X3), the input values ​​are multiplied by a weight W1 (shown as w11 to w16 in FIG. 2) and input to the intermediate layer (Y1 to Y2). The result is then further multiplied by a weight W2 (shown as w21 to w26 in FIG. 2) and output from the output layer (Z1 to Z3). This output result varies depending on the values ​​of the weights W1 and W2.

[0083] The neural network used by the machine learning device 2 learns the degradation level of the filtration filter corresponding to the input data by so-called supervised learning in accordance with learning data generated based on a combination of the input data and label data acquired by the input data acquiring unit 25 and the label acquiring unit 26. In other words, the neural network used by the machine learning device 2 learns the output data (degradation level of the filtration filter) corresponding to the input data by so-called supervised learning in accordance with the input data and label data generated based on a combination of the first input and the second input (correct answer) acquired by the input data acquiring unit 25 and the label acquiring unit 26.

[0084] In other words, the neural network learns by adjusting the weights W1 and W2 so that the result output from the output layer after receiving the input data, which is the first input, approaches the label data, which is the second input (correct answer).

[0085] In this way, the neural network learns by inputting input data to the input layer and adjusting the weights W1 and W2 so that the result output from the output layer approaches the label data. By learning the correspondence between the input data and the label data, the neural network generates a learning model that can output an appropriate deterioration level of the filtration filter that corresponds to the label data when input data is input. In this way, the machine learning device 2 learns a learning model that can output the correct deterioration level of the filtration filter when input data such as the accumulated amount of processing waste is input.

[0086] The model generation unit 27 generates a learning model by executing the above-described learning, and outputs the learning model to the learning model storage unit 5. The learning model storage unit 5 stores the learning model output from the model generation unit 27.

[0087] The machine learning device 2 may divide the data from the (t-N)th day to the (t+1)th day into multiple parts and perform multiple rounds of machine learning. That is, if the machine learning device 2 has acquired L days' worth of data (L is a natural number), it may generate M (M is a natural number) rounds of learning data from this L days' worth of data and perform M rounds of machine learning.

[0088] For example, when the machine learning device 2 acquires 100 days' worth of data, the machine learning device 2 may divide the 100 days' worth of data into data from the 1st to 50th days and data from the 50th to 100th days. In this case, the machine learning device 2 performs machine learning using the data from the 1st to 50th days as first-round learning data, and performs machine learning using the data from the 51st to 100th days as second-round learning data.

[0089] If the data from the 1st to 50th days is the learning data, the degradation level from the 1st to 49th days is the specific time degradation level, and the data on the 50th day is the next degradation level. Similarly, if the data from the 51st to 100th days is the learning data, the degradation level from the 51st to 99th days is the specific time degradation level, and the data on the 100th day is the next degradation level.

[0090] Furthermore, the machine learning device 2 may perform machine learning using data from the 1st to 50th days as first-round learning data, data from the 25th to 75th days as second-round learning data, and data from the 50th to 100th days as third-round learning data. In other words, the machine learning device 2 may use the acquired data for multiple machine learning processes.

[0091] Next, a description will be given of the processing procedure of the process of learning a learning model performed by the machine learning device 2. Fig. 3 is a flowchart showing the processing procedure of the learning process executed by the machine learning device according to the embodiment.

[0092] The input data acquisition unit 25 and label acquisition unit 26, which are the data acquisition unit 20, acquire learning data used for learning (step S10). Specifically, the input data acquisition unit 25 acquires the processing conditions (processing speed and processing time) from day (T-1) to day T as measurement result time-series data D3. The input data acquisition unit 25 also acquires the planned processing conditions (planned processing speed and planned processing time) for day (T+1) as planned processing condition data D1. The input data acquisition unit 25 also acquires the cumulative processing waste volume from day (T-1) to day (T+1) as calculation result time-series data D2. The input data acquisition unit 25 also acquires the deterioration level of the consumables from day (T-1) to day T as specific time-point deterioration level data D4. The label acquisition unit 26 also acquires the deterioration level of the consumables on day (T+1) as next-time deterioration level data D5.

[0093] The input data acquiring unit 25 sends the planned processing condition data D1, the measurement result time-series data D3, the calculation result time-series data D2, and the specific point-in-time degradation degree data D4 as input data to the model generating unit 27. In addition, the label acquiring unit 26 sends the next degradation degree data D5 as label data to the model generating unit 27. A combination of the input data acquired by the input data acquiring unit 25 and sent to the model generating unit 27 and the label data acquired by the label acquiring unit 26 and sent to the model generating unit 27 is learning data.

[0094] Although the input data acquisition unit 25 and the label acquisition unit 26 are assumed to acquire data simultaneously here, it is sufficient for the input data acquisition unit 25 and the label acquisition unit 26 to acquire data in an associated manner. Therefore, the input data acquisition unit 25 and the label acquisition unit 26 may acquire data at different times. That is, the data acquisition unit 20 (the input data acquisition unit 25 and the label acquisition unit 26) may acquire the planned processing condition data D1, the measurement result time-series data D3, the calculation result time-series data D2, the specific time degradation degree data D4, and the next degradation degree data D5 at different times. The input data acquisition unit 25 and the label acquisition unit 26 transmit the acquired data to the model generation unit 27.

[0095] The model generation unit 27 executes a learning process using the input data sent from the input data acquisition unit 25 and the label data sent from the label acquisition unit 26 (step S20). Specifically, the model generation unit 27 learns label data corresponding to the input data by so-called supervised learning in accordance with learning data generated based on a combination of the input data and the label data acquired by the input data acquisition unit 25 and the label acquisition unit 26, and generates a learning model.

[0096] The input data here includes the following information: Processing speed and processing time from the past up to a specific point in time; Accumulated processing waste volume from the past up to a specific point in time (calculated value); Pressure loss of the filtration filter from the past up to a specific point in time; Planned processing speed and planned processing time for the next processing; Accumulated processing waste volume at the time the next processing is completed (calculated value).

[0097] The label data here also includes the following information: Pressure loss of the filtration filter at the time the next processing is completed

[0098] The model generation unit 27 sends the generated learning model to the learning model storage unit 5. The learning model storage unit 5 stores the learning model generated by the model generation unit 27 (step S30).

[0099] Next, a process procedure for the inference device 3 to infer the pressure loss of the filtration filter at the time when the next processing is completed using the learning model will be described. Fig. 4 is a flowchart showing the process procedure for the inference process executed by the inference device according to the embodiment.

[0100] The data acquisition unit 35 acquires inference data to be used to infer the pressure loss of the filtration filter at the time the next processing is completed (step S110). Specifically, the data acquisition unit 35 acquires the processing conditions from the (t-1)th day to the tth day as measurement result time-series data D13. The data acquisition unit 35 also acquires the planned processing conditions for the (t+1)th day as planned processing condition data D11. The data acquisition unit 35 also acquires the accumulated processing waste volume from the (t-1)th day to the (t+1)th day as calculation result time-series data D12. The data acquisition unit 35 also acquires the deterioration level of the consumables (such as the pressure loss of the filtration filter) from the (t-1)th day to the tth day as specific time-point deterioration level data D14.

[0101] Although the data acquisition unit 35 acquires the data simultaneously in this example, it is sufficient for the data acquisition unit 35 to acquire the data in an associated manner. Therefore, the data acquisition unit 35 may acquire each piece of data at a different timing. That is, the data acquisition unit 35 may acquire the planned processing condition data D11, the measurement result time-series data D13, the calculation result time-series data D12, and the point-in-time degradation degree data D14 at different timings. The data acquisition unit 35 sends the acquired data to the inference unit 36 ​​as inference data (input data).

[0102] The inference unit 36 ​​acquires inference data from the data acquisition unit 35. The inference unit 36 ​​also acquires a learning model from the learning model storage unit 5. The inference unit 36 ​​inputs the inference data into the learning model (step S120), and acquires an inference result corresponding to the inference data (pressure loss of the filtration filter at the time when the next processing is completed).

[0103] The inference unit 36 ​​outputs the inference result, which is the pressure loss of the filtration filter at the time when the next processing is completed, to a display device (not shown) or the like (step S130). As a result, the display device displays the pressure loss of the filtration filter at the next processing as the life of the consumable item (step S140).

[0104] In this embodiment, the inference device 3 infers the pressure loss of the filtration filter using a learning model learned by the machine learning device 2 of the processing machine. However, the inference device 3 may also acquire a learning model from an external source, such as another processing machine, and output the pressure loss of the filtration filter based on this learning model.

[0105] Here, the input data and output data used for learning by the machine learning device 2 will be described. Fig. 5 is a diagram for explaining the input data and output data used for learning by the machine learning device according to the embodiment. The point in time at which day T ends in Fig. 5 is a specific point in time (the current point in time).

[0106] The input data input to the model generation unit 27 of the machine learning device 2 includes the machining conditions and calculation data (cumulative machining waste amount) from the (T-N)th day to the Tth day, and the machining conditions and calculation data (cumulative machining waste amount) from the end of the Tth day to the (T+1)th day. The period from the (T-N)th day to the Tth day refers to the period from the start of the (T-N)th day to the end of the Tth day. Furthermore, the period from the end of the Tth day to the (T+1)th day refers to the period from the end of the Tth day to the end of the (T+1)th day.

[0107] The input data input to the model generation unit 27 also includes the degree of deterioration of the consumables (pressure loss of the filtration filter) from the (T-N)th day to the Tth day. Note that the input data input to the model generation unit 27 may also include measurement data (such as temperature) from the (T-N)th day to the Tth day.

[0108] The model generation unit 27 also uses the deterioration level of the consumables as label data (actual measurement value). The model generation unit 27 compares output data (predicted value) output during inference with the label data (actual measurement value), and generates a learning model based on the input data and the label data so that the output data (predicted value) output during inference approaches the label data (actual measurement value).

[0109] This enables the learning model to infer the degree of deterioration of the consumables after the next processing is completed based on the processing conditions and calculation data from the past to a specific point in time that were set at the time of inference, the processing conditions and calculation data after the next processing is completed, and the degree of deterioration of the consumables from the past to a specific point in time.

[0110] Next, the inference process performed by the inference device 3 of the present embodiment and the inference process of a comparative example will be described. Fig. 6 is a diagram for explaining the inference process performed by the inference device according to the embodiment and the inference process of the comparative example.

[0111] 6, the first processing example of the inference processing is shown in the upper part, and the second processing example of the inference processing is shown in the lower part. The second processing example is an example of the inference processing by the inference device 3, and the first processing example is a comparative example of the inference processing.

[0112] In the first processing example, the input data used for inference is data at a specific point in time, and in the second processing example, the data used for inference is data from the past up to a specific point in time and future data.

[0113] In the first processing example, the degree of deterioration of a consumable item after one processing run (pressure loss of the filtration filter) is inferred based on the processing conditions for one processing run used for one processing run at a specific point in time, and output data (predicted value) that is the inference result is output.

[0114] In the second processing example, the inference device 3 infers the degree of deterioration of the consumables after the next processing (here, after the end of the (t+1) day) based on the processing conditions and calculation data (cumulative processing waste amount) from the (t-n) day to the t day, the processing conditions and calculation data (cumulative processing waste amount) from the end of the t day to the (t+1) day, and the degree of deterioration of the consumables (pressure loss of the filtration filter) from the (t-n) day to the t day, and outputs output data (predicted value) that is the inference result.

[0115] In addition, the inference device 3 may infer the degree of deterioration of the consumables after the next processing is completed using the next processing conditions and calculation data (cumulative processing waste amount) for one processing run after the end of the tth day, instead of the processing conditions and calculation data (cumulative processing waste amount) for the period from the end of the tth day to the (t+1)th day.

[0116] In this way, in the second processing example, the inference device 3 infers the degree of deterioration of the consumables after the next processing using data from the past up to a specific point in time and future data, thereby enabling accurate inference.

[0117] Furthermore, in the second processing example, the inference device 3 infers the degree of deterioration of the consumables after the next processing based on information on a physical quantity that has a significant impact on the degree of deterioration of the consumables (the cumulative amount of processing waste of the filtration filter), thereby enabling accurate inference.

[0118] 7 is a graph illustrating the relationship between the cumulative amount of processing waste learned by the machine learning device according to the embodiment and the pressure loss of the filtration filter, where the horizontal axis of the graph in FIG. 7 represents the cumulative amount of processing waste V and the vertical axis represents the pressure loss P of the filtration filter.

[0119] 7 shows a graph in which the relationship between the cumulative amount of processing waste V of the filters Fa, Fb, and Fc and the pressure loss P of the filtration filter is plotted for each unit time Δt (e.g., one day). The pressure loss P of the filtration filter increases relative to the cumulative amount of processing waste V, and the cumulative amount of processing waste V is a parameter highly correlated with the pressure loss P of the filtration filter. This relationship between the cumulative amount of processing waste V and the pressure loss P of the filtration filter differs for each of the filters Fa, Fb, and Fc. That is, the filters Fa, Fb, and Fc are each different in at least one of the processing machine in which they are installed, the processing conditions of the processing machine, the user who uses the processing machine, and the processing environment, and therefore each exhibits different changes.

[0120] As described above, the relationship between the cumulative waste amount V and the pressure loss P of the filtration filter differs for each of the filters Fa, Fb, and Fc. Therefore, even if the cumulative waste amount V is determined, the pressure loss P is not necessarily uniquely determined. In other words, even if a certain cumulative waste amount (here, cumulative waste amount V0) is the same, the pressure loss P may differ for each of the filters Fa, Fb, and Fc.

[0121] Furthermore, even if the cumulative amount of machining waste V0 is the same, if the pressure loss P differs for each of the filters Fa, Fb, and Fc, the rate of change of the pressure loss P with respect to the cumulative amount of machining waste V (ΔP / ΔV) also differs for each of the filters Fa, Fb, and Fc.

[0122] FIG. 7 shows a case where the pressure loss P differs for each of the filters Fa, Fb, and Fc even when the cumulative amount of machining waste V is the same, and the rate of change of the pressure loss P with respect to the cumulative amount of machining waste V (ΔP / ΔV) also differs for each of the filters Fa, Fb, and Fc.

[0123] The machine learning device 2 generates a learning model by learning the relationship between the accumulated amount of processing waste V and the pressure loss P of the filtration filter as shown in FIG.

[0124] Note that, once the cumulative machining waste amount V is determined, the pressure loss P may be uniquely determined. If the cumulative machining waste amount V is the same for each filter and the pressure loss P is also the same for each filter, the rate of change of the pressure loss P with respect to the cumulative machining waste amount V (ΔP / ΔV) will also be the same for each filter. In this case, the machine learning device 2 may learn the correspondence relationship between the cumulative machining waste amount V and the pressure loss P. In other words, the machine learning device 2 may generate a learning model without using time-series data.

[0125] In this case, the data acquisition unit 20 of the machine learning device 2 acquires learning data including the accumulated machining waste amount V and the pressure loss P at the time when the next machining is completed. The model generation unit 27 uses the learning data to generate a learning model for inferring the pressure loss P from the accumulated machining waste amount V. In other words, the learning model generated by the model generation unit 27 learns the correspondence relationship between the accumulated machining waste amount and the pressure loss of the filtration filter. In this case, the inference unit 36 ​​infers the pressure loss of the filtration filter from the accumulated machining waste amount by applying the accumulated machining waste amount to the learning model.

[0126] 8 is a graph illustrating the relationship between the accumulated amount of waste and the pressure loss of the filtration filter, as inferred by the inference device according to the embodiment. The horizontal axis of the graph shown in FIG. 8 represents the accumulated amount of waste V, and the vertical axis represents the pressure loss P of the filtration filter.

[0127] 8, similar to Fig. 7, shows a graph in which the relationship between the accumulated amount of processing waste V of the filters Fa, Fb, and Fc and the pressure loss P of the filtration filters is plotted for each unit time Δt (e.g., one day). The inference device 3 uses the relationship between the accumulated amount of processing waste V from the past up to a specific point in time and the pressure loss P of the filtration filters as input data for a learning model, and infers the pressure loss P of the filtration filters corresponding to the accumulated amount of processing waste V when the next processing is completed.

[0128] 8, the relationship between the accumulated machining waste amount V from the (t-n)th day to the tth day and the pressure loss P of the filtration filter, which is indicated by a filled marker, is input data to the inference unit 36. In this embodiment, not only information on the accumulated machining waste amount V on the tth day, which is a specific point in time, but also time-series data on the accumulated machining waste amount V from the specific point in time up to N unit times before is used as input data to the inference unit 36.

[0129] For example, the time series data of the accumulated machining waste amount V from the (t-n)th day to the (t+1)th day is V t-n , V t-n+1 , ..., V t-1 , V t , V t+1 In addition, the time series data of the pressure loss P from the (t-n)th day to the (t+1)th day is P t-n , P t-n+1 , ..., P t-1 , P t , P t+1 In addition, the time series data of the processing conditions from the (t-n)th day to the (t+1)th day is expressed as X t-n , X t-n+1 , ..., X t-1 , X t , X t+1 Let's say.

[0130] In this case, the inference unit 36 ​​receives as input data the time series data of the accumulated machining waste amount V from the (t-n)th day to the (t+1)th day, the time series data of the machining conditions from the (t-n)th day to the (t+1)th day, and the time series data of the pressure loss P from the (t-n)th day to the tth day, and infers the pressure loss P on the (t+1)th day, which is one unit of time into the future. t-n , V t-n+1 , ..., V t-1 , V t , V t+1 and X from the (t-n)th day to the (t+1)th day t-n , X t-n+1 , ..., X t-1 , X t , X t+1 and P from (t-n)th day to tth day t-n , P t-n+1 , ..., P t-1 , P t So, P on the (t+1)th day t+1 Infer that.

[0131] In this way, the learning model learns the pressure loss P of the filtration filter based on the most recent trend information (rising, falling, stagnation, etc.), and therefore the inference unit 36 ​​can predict the pressure loss P of the filtration filter with high accuracy based on the most recent trend information.

[0132] In this way, since the inference unit 36 ​​can predict the pressure loss P of the filtration filter with high accuracy, the user can arrange for a replacement filtration filter before the filtration filter reaches the end of its life, thereby reducing production loss due to waiting for the delivery of the filtration filter. Furthermore, since the inference unit 36 ​​can predict the pressure loss P of the filtration filter with high accuracy, the user can set an appropriate machining sequence according to the remaining life of the filtration filter. This allows the user to prevent the filtration filter from reaching the end of its life in the middle of machining, thereby optimizing the wire electric discharge machining process.

[0133] In this embodiment, a case has been described in which supervised learning is applied to the learning algorithm used by the model generation unit 27, but the learning algorithm is not limited to supervised learning. As for the learning algorithm, reinforcement learning, unsupervised learning, semi-supervised learning, or the like can also be applied in addition to supervised learning.

[0134] The model generation unit 27 may also learn the pressure loss of the filtration filter according to learning data created for multiple processing machines. The model generation unit 27 may acquire learning data from multiple processing machines used in the same area, or may learn the pressure loss of the filtration filter using learning data collected from multiple processing machines operating independently in different areas. It is also possible to add or remove processing machines from which learning data is collected during the process. Furthermore, a machine learning device that has learned the pressure loss of the filtration filter for a certain processing machine may be applied to a different processing machine, and the pressure loss of the filtration filter for the different processing machine may be re-learned and updated.

[0135] Furthermore, the learning algorithm used in the model generation unit 27 can be deep learning, which learns to extract the features themselves, or machine learning can be performed according to other known methods, such as genetic programming, functional logic programming, or support vector machines.

[0136] In the filter of a wire electric discharge machine, machining debris accumulates as it repeatedly adheres to and peels off the filter. In a wire electric discharge machine, as machining debris accumulates on the filter, the pressure loss of the filter increases. In this case, if the amount of machining debris in the machining fluid increases gradually over time, more machining debris peels off the filter, and less machining debris accumulates on the filter. On the other hand, if the amount of machining debris in the machining fluid increases rapidly over time, less machining debris peels off the filter, and more machining debris accumulates on the filter.

[0137] In this embodiment, the inference device 3 infers the next deterioration level using time-series data such as the planned machining condition data D11, the measurement result time-series data D13, the calculation result time-series data D12, and the specific point-in-time deterioration level data D14. This enables the inference device 3 to infer the next deterioration level according to the transition of conditions (such as machining conditions) that affect the deterioration level of the consumables. The inference device 3 can infer, for example, pressure loss according to the transition of the accumulated amount of machining waste and the accumulation of machining waste on the filtration filter, thereby enabling the accurate inference of the next deterioration level.

[0138] 9 is a diagram showing the configuration of a machining system according to an embodiment. The machining system 100 includes a wire electric discharge machine 10. An example of the wire electric discharge machine 10 is the wire electric discharge machine described above. The wire electric discharge machine 10 includes an inference device 3, a machine learning device 2, and a learning model storage unit 5.

[0139] In the wire electric discharge machine 10, the machine learning device 2 generates a learning model using information (learning data) acquired from the wire electric discharge machine 10. This learning model is stored in the learning model storage unit 5. The inference device 3 infers the pressure loss of the filtration filter when the next machining is completed, based on the information (inference data) acquired from the wire electric discharge machine 10 and the learning model read from the learning model storage unit 5.

[0140] The machine learning device 2 may be located outside the wire electric discharge machine 10. The machine learning device 2 may also generate a learning model using information acquired from a machine other than the wire electric discharge machine 10. The inference device 3 may also be located outside the wire electric discharge machine 10.

[0141] Furthermore, the learning model storage unit 5 may be located outside the wire electric discharge machine 10. The learning model stored in the learning model storage unit 5 may be a learning model generated using information acquired from a machine other than the wire electric discharge machine 10.

[0142] When the machine learning device 2 is located outside the wire electric discharge machine 10, the machine learning device 2 is connected to the wire electric discharge machine 10 via a network. When the inference device 3 is located outside the wire electric discharge machine 10, the inference device 3 is connected to the wire electric discharge machine 10 via a network. When the learning model storage unit 5 is located outside the wire electric discharge machine 10, the learning model storage unit 5 is connected to the machine learning device 2 and the inference device 3 via a network. At least one of the machine learning device 2, the inference device 3, and the learning model storage unit 5 may exist on a cloud server.

[0143] Next, we will explain the hardware configurations of the machine learning device 2 and the inference device 3. Note that the machine learning device 2 and the inference device 3 have similar hardware configurations, so here we will explain the hardware configuration of the inference device 3.

[0144] The inference device 3 is realized by a processing circuit. This processing circuit may be a processor and memory that executes a program stored in memory, or may be dedicated hardware. The processing circuit is also called a control circuit.

[0145] FIG. 10 is a diagram illustrating an example of the configuration of a processing circuit included in an inference device according to an embodiment, where the processing circuit is implemented using a processor and a memory. The processing circuit 90 illustrated in FIG. 10 is a control circuit and includes a processor 91 and a memory 92. When the processing circuit 90 is configured using the processor 91 and the memory 92, each function of the processing circuit 90 is implemented by software, firmware, or a combination of software and firmware. The software or firmware is written as a program and stored in the memory 92. The processor 91 reads and executes the program stored in the memory 92 to implement each function of the processing circuit 90. That is, the processing circuit 90 includes the memory 92 for storing an inference program that results in the processing of the inference device 3 being executed. This inference program can also be said to be a program that causes the inference device 3 to execute each function implemented by the processing circuit 90. This inference program may be provided by a storage medium on which the program is stored, or by other means such as a communication medium.

[0146] The inference device 3 is realized by the processor 91 executing an inference program stored in the memory 92. That is, the inference program executed by the inference device 3 has a modular configuration including a schedule creation unit 31, a calculation unit 32, a measurement unit 33, a deterioration level acquisition unit 34, a data acquisition unit 35, and an inference unit 36, which are loaded onto a main storage device and generated on the main storage device. The processing performed by at least one of the schedule creation unit 31, the calculation unit 32, the measurement unit 33, and the deterioration level acquisition unit 34 may be executed by a program separate from the inference program.

[0147] The learning program used by the machine learning device 2 has a modular configuration including a schedule creation unit 21, a calculation unit 22, a measurement unit 23, a degradation level acquisition unit 24, an input data acquisition unit 25, a label acquisition unit 26, and a model generation unit 27, which are loaded onto a main storage device and generated on the main storage device. The processing performed by at least one of the schedule creation unit 21, the calculation unit 22, the measurement unit 23, and the degradation level acquisition unit 24 may be executed by a program separate from the learning program.

[0148] Here, the processor 91 is, for example, a CPU (Central Processing Unit), a processing device, an arithmetic unit, a microprocessor, a microcomputer, or a DSP (Digital Signal Processor), etc. The memory 92 is, for example, a non-volatile or volatile semiconductor memory such as a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable ROM), or an EEPROM (Electrically EPROM), a magnetic disk, a flexible disk, an optical disk, a compact disk, a minidisk, or a DVD (Digital Versatile Disc).

[0149] 11 is a diagram illustrating an example of a processing circuit when the processing circuit included in the inference device according to the embodiment is configured with dedicated hardware. The processing circuit 93 illustrated in FIG. 11 corresponds to, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or a combination thereof.

[0150] The processing circuits 90 and 93 may be partially implemented by dedicated hardware and partially implemented by software or firmware. In this way, the processing circuits 90 and 93 can implement the above-mentioned functions by dedicated hardware, software, firmware, or a combination of these. The inference device 3 may be implemented by a single processing circuit or multiple processing circuits.

[0151] In this way, the inference device 3 of the embodiment uses a learning model for inferring next deterioration level data at the time when the next processing is completed from the processing condition time-series data and the specific time degradation level data D14 to infer next deterioration level data from the inference data (processing condition time-series data and specific time degradation level data D14) input from the data acquisition unit 35. This allows the inference device 3 to accurately infer the deterioration level of the consumables based on the processing condition time-series data.

[0152] Furthermore, the machine learning device 2 of the embodiment uses learning data including the processing condition time-series data, the specific time degradation level data D4, and the next degradation level data D5 to generate a learning model for inferring the next degradation level data D5 from the processing condition time-series data and the specific time degradation level data D4. This enables the machine learning device 2 to generate learning data that can accurately infer the degradation level of a consumable based on the processing condition time-series data.

[0153] Furthermore, since the machine learning device 2 generates a learning model based on information on physical quantities (such as the accumulated amount of processing waste) that have a significant impact on the degree of deterioration of consumables, it is possible to generate a learning model using a small amount of data in machine learning.

[0154] The configurations shown in the above embodiments are merely examples, and may be combined with other known technologies, and parts of the configurations may be omitted or modified without departing from the spirit of the invention.

[0155] 1 Learning inference system, 2 Machine learning device, 3 Inference device, 5 Learning model storage unit, 10 Wire electric discharge machine, 20, 35 Data acquisition unit, 21, 31 Schedule creation unit, 22, 32 Calculation unit, 23, 33 Measurement unit, 24, 34 Deterioration degree acquisition unit, 25 Input data acquisition unit, 26 Label acquisition unit, 27 Model generation unit, 36 Inference unit, 90, 93 Processing circuit, 91 Processor, 92 Memory, 100 Machining system, D1, D11 Planned machining condition data, D2, D12 Calculation result time series data, D3, D13 Measurement result time series data, D4, D14 Specific time point deterioration degree data, D5, D15 Next deterioration degree data, Fa, Fb, Fc Filter, P Pressure loss, V, V0 Cumulative machining waste amount, W1, W2 Weight, Δt Unit time.

Claims

1. An inference device comprising: a data acquisition unit that acquires inference data including processing condition time-series data that indicates the transition of processing conditions applied to a processing machine up to a specific point in time and processing conditions to be used in the next processing, and specific time degradation level data that indicates the transition of the degree of deterioration of a consumable used by the processing machine up to a specific point in time; and an inference unit that infers and outputs next time degradation level data from the inference data input from the data acquisition unit using a learning model for inferring next time degradation level data that indicates the degree of deterioration of the consumable at the time the next processing is completed from the processing condition time-series data and the specific time degradation level data.

2. The inference device described in claim 1, characterized in that the inference data further includes data on the physical quantities of the consumables calculated based on the processing conditions, the learning data used when generating the learning model includes the processing condition time series data, the specific point in time deterioration degree data, the next deterioration degree data, and the physical quantity data, and the inference unit infers the next deterioration degree data from the inference data input from the data acquisition unit using a learning model for inferring the next deterioration degree data from the processing condition time series data, the specific point in time deterioration degree data, and the physical quantity data.

3. The inference device according to claim 2, characterized in that the data on the physical quantities includes physical quantity time series data showing the trends of the physical quantities up to a specific point in time, and the physical quantities at the time when the next processing is completed.

4. The inference device according to claim 2 or 3, characterized in that the machining machine is a wire electric discharge machine, the consumable is a filter that removes machining debris mixed in the machining fluid, the physical quantity data is the cumulative amount of machining debris mixed in the machining fluid, and the degree of deterioration of the consumable is the pressure loss of the filter.

5. A machine learning device comprising: a data acquisition unit that acquires learning data including time series data showing the transition of processing conditions applied to a processing machine up to a specific point in time, processing condition time series data including processing conditions to be used in the next processing, specific time degradation level data showing the transition of the deterioration level of a consumable used by the processing machine up to a specific point in time, and next time degradation level data showing the deterioration level of the consumable at the time the next processing is completed; and a model generation unit that uses the learning data to generate a learning model for inferring the next time degradation level data from the processing condition time series data and the specific time degradation level data.

6. The machine learning device described in claim 5, characterized in that the learning data includes data on physical quantities of the consumables calculated based on the processing conditions, and the model generation unit uses the learning data to generate the learning model for inferring the next deterioration level data from the processing condition time series data, the physical quantity data, and the specific point-in-time deterioration level data.

7. The machine learning device according to claim 6, characterized in that the data on the physical quantities includes physical quantity time-series data showing the transition of the physical quantities up to a specific point in time, and the physical quantities at the time when the next processing is completed.

8. The machine learning device according to any one of claims 5 to 7, wherein the machining conditions used in the next machining are planned machining conditions.

9. The machine learning device described in any one of claims 5 to 8, characterized in that the time series data showing the changes in the machining conditions up to a specific point in time are planned machining conditions or measured actual machining conditions.

10. The machine learning device according to claim 6 or 7, characterized in that: the machining machine is a wire electric discharge machine; the consumable is a filter that removes machining debris mixed in the machining fluid; the physical quantity data is the cumulative amount of machining debris mixed in the machining fluid; and the degree of deterioration of the consumable is the pressure loss of the filter.

11. A processing system comprising: a processing machine; and an inference device that infers the degree of deterioration of consumables used by the processing machine, wherein the inference device comprises: a first data acquisition unit that acquires inference data including processing condition time-series data, which includes time-series data indicating the transition of processing conditions applied to the processing machine up to a specific point in time and processing conditions to be used in the next processing, and specific time degradation level data indicating the transition of the degree of deterioration of the consumables up to a specific point in time; and an inference unit that infers and outputs next degradation level data from the inference data input from the first data acquisition unit, using a learning model for inferring next degradation level data indicating the degree of deterioration of the consumables at the time the next processing is completed from the processing condition time-series data and the specific time degradation level data.

12. The processing system described in claim 11, further comprising a learning device that generates the learning model, wherein the learning device comprises: a second data acquisition unit that acquires learning data including the processing condition time series data, the specific point-in-time degradation level data, and the next degradation level data; and a model generation unit that generates the learning model using the learning data.

Images