JP7572865B2 - Learning device and program, and die abnormality prediction device - Google Patents

Description

The present invention relates to a learning device and program, and a die abnormality prediction device, and in particular to a learning device and program that use machine learning to generate a prediction model for predicting the presence or absence of an abnormality in a die mounted on a press machine, and a die abnormality prediction device that predicts the presence or absence of an abnormality in a die based on a prediction model generated by machine learning.

A forging press machine presses a workpiece by rotating a crankshaft and raising and lowering a slide. As described in JP 2019-13947 A (Patent Document 1), such a forging press machine has traditionally monitored the load applied to the workpiece during processing and detected abnormalities based on the peak value of the load. This is because it is known that the peak value of the load is highly correlated with the quality of the product.

In addition, as described in JP 2019-13976 A (Patent Document 2), a technology has been proposed that predicts press failures by using multiple sensors that detect the operating status of the press.

JP 2019-13947 A JP 2019-13976 A

In Patent Document 1, it is possible to detect defective products based on the peak value of the load, but it is not possible to identify the cause (cause of anomaly).

In Patent Document 2, changes from the initial normal state in the measurement results obtained from multiple sensors are checked on a monitoring PC and past empirical rules are also taken into account to predict press failures, making it difficult for inexperienced operators to accurately predict failures.

In particular, press machine dies tend to break down more frequently than other parts, so there was a demand for technology that could detect die anomalies (breakdowns, damage, etc.) without relying on the experience of the worker.

The present invention has been made to solve the above problems, and its purpose is to provide a learning device and program for enabling accurate detection of abnormalities in dies mounted on a press machine, as well as a die abnormality prediction device.

A learning device according to one aspect of the present invention is a learning device that generates a prediction model by machine learning to predict the presence or absence of an abnormality in a die mounted on a press machine, and includes a data acquisition means, a feature calculation means, a generation means, and a model storage means. The data acquisition means acquires measurement data for each press cycle, including load data applied to a workpiece during press operation by the press machine and vibration data of the frame of the press machine. The feature calculation means calculates multiple types of feature amounts, including load feature amounts and vibration feature amounts, for each press cycle based on the measurement data acquired by the data acquisition means. The generation means performs machine learning on the correlation between the multiple types of feature amounts calculated by the feature calculation means and the die state, and generates a prediction model for predicting the presence or absence of an abnormality in the die for each press cycle. The model storage means stores the prediction model generated by the generation means.

Preferably, the generating means performs machine learning to learn the correlation between multiple types of feature values and the mold state for each mold pattern, and the model storing means stores the prediction model in association with mold identification information that identifies the mold pattern.

When the press is a multi-stage press equipped with multiple dies, it is desirable for the generation means to generate a prediction model that takes multiple types of feature quantities as input and outputs the presence or absence of anomalies for each die. In a broad sense, each die means a die for each process, without distinguishing between upper and lower dies, and in a narrow sense, means the upper and lower dies for each process.

In a configuration in which the frame has first and second supports provided on the upstream and downstream sides, respectively, of the multiple molds, it is preferable that the feature calculation means calculates a first vibration feature based on vibration data from a vibration meter attached to the first support, and calculates a second vibration feature based on vibration data from a vibration meter attached to the second support. Each vibration feature includes at least one of a wave height vector, a time vector, and a frequency vector, and it is preferable that it includes at least a frequency vector. More preferably, it includes a frequency vector, and a wave height vector or a time vector.

In addition to the load feature and vibration feature, it is preferable that the feature calculation means further calculates at least one of the die contact time, die height value, die temperature, die lubricant amount, and die shot number. The types of feature preferably include die temperature and die lubricant amount. By adding these to the explanatory variables of the prediction model, the magnitude of the load on the die can be detected with high accuracy. If the press is a multi-stage press equipped with multiple dies, it is preferable that the die temperature, die lubricant amount, and die shot number are measured for each die (at least for each process).

A die abnormality prediction device according to another aspect of the present invention is a die abnormality prediction device for predicting the presence or absence of an abnormality in a die mounted on a press machine, and includes a storage means for storing a prediction model generated by any of the learning devices described above. It also includes a data acquisition means for acquiring measurement data for each press cycle, including load data applied to a workpiece during press operation by the press machine and vibration data of the frame of the press machine, a feature calculation means for calculating multiple types of feature amounts, including load feature amounts and vibration feature amounts, for each press cycle based on the measurement data acquired by the data acquisition means, a prediction means for inputting the multiple types of feature amounts calculated by the feature calculation means into a prediction model stored in the storage means to predict the presence or absence of an abnormality in the die for each press cycle, and an output means for outputting the prediction result by the prediction means.

Preferably, the storage means stores a plurality of prediction models in association with die identification information that identifies the die pattern. In this case, it is preferable that the die abnormality prediction device further includes a model discrimination means that discriminates a prediction model to be used for die abnormality prediction from among the plurality of prediction models, based on the die pattern corresponding to the current press operation, before the start of press operation by the press machine.

A learning program according to yet another aspect of the present invention is a learning program that generates a prediction model by machine learning to predict the presence or absence of an abnormality in a die mounted on a press machine, and causes a computer to execute the steps of: calculating multiple types of feature quantities, including load feature quantities and vibration feature quantities, for each press cycle based on measurement data including load waveform data applied to a workpiece during press operation by the press machine and vibration waveform data of the frame of the press machine; learning the correlation between the multiple types of calculated feature quantities and the die state by machine learning to generate a prediction model for predicting the presence or absence of an abnormality in the die for each press cycle; and storing the generated prediction model in a model storage means.

According to the present invention, a predictive model can be used to accurately detect abnormalities in a mold.

1 is a diagram showing a schematic configuration of a forging press apparatus according to an embodiment of the present invention; 1 is a diagram showing a schematic diagram of a press machine and sensors provided in a forging press apparatus according to an embodiment of the present invention. FIG. FIG. 2 is a diagram showing a schematic diagram of the peripheral structure of a press machine according to an embodiment of the present invention, and operating status information detected by sensors. FIG. 2 is a diagram showing a schematic example of an arrangement of dies provided in a press machine according to an embodiment of the present invention. 1 is a block diagram showing a functional configuration of a learning device according to an embodiment of the present invention; 1 is a flowchart showing a method for generating a prediction model by a learning device according to an embodiment of the present invention. 13A to 13C are graphs for explaining a method for calculating a feature amount of frame vibration. 13A to 13D are graphs for explaining a method for calculating a feature amount of frame vibration. 1 is a block diagram showing a functional configuration of a mold abnormality prediction device according to an embodiment of the present invention. 4 is a flowchart showing a mold abnormality prediction method by the mold abnormality prediction device according to the embodiment of the present invention.

The embodiment of the present invention will be described in detail with reference to the drawings. Note that the same or corresponding parts in the drawings are given the same reference numerals and their description will not be repeated.

In this embodiment, the forging press machine is described as having the function of detecting abnormalities in the die based on the learning results of the learning device.

<Outline of the forging press device>
First, a schematic configuration of a forging press apparatus 10 will be described with reference to Figures 1 to 3. As shown in Figure 1, the forging press apparatus 10 includes a press machine 11 that presses a workpiece, and a PLC (programmable logic controller) 40 that controls the operation of the press machine 11.

As shown in FIG. 2, the press 11 has a slide 1 and a bolster 2 arranged to face each other vertically within a frame 3. The frame 3 includes a pair of supports 3a, 3b arranged on the front side of the press 11, and a pair of supports 3c, 3d arranged on the rear side of the press 11. The press 11 is configured so that the slide 1 connected to the connecting rod 5 moves up and down by rotating the crankshaft 4.

A flywheel 8, which is driven to rotate by a belt 7 by a main motor 6, is connected to one end of the crankshaft 4 via a clutch 9. A brake device 12 is attached to the other end of the crankshaft 4, which stops the rotation of the crankshaft 4 and stops the slide 1. The press 11 drives the flywheel 8 to rotate at a constant speed by the main motor 6, and with the clutch 9 engaged, transmits the rotation of the flywheel 8 to the crankshaft 4, raising and lowering the slide 1, thereby forging the workpiece placed on the bolster 2. The workpiece is processed into a size and shape according to the pattern of a die, which includes a lower die (not shown) installed on the bolster 2 and an upper die (not shown) installed at the lower end of the slide 1, to become the final product. "Die" means a die used for forging. The die is typically made of special steel.

The forging press 10 in this embodiment is a device that processes a workpiece by hot forging. The forging press 10 is equipped with a plurality of sensors as described below in order to detect the operating status of the press 11. As will be described later, the press 11 may be a multi-stage press that processes a workpiece through a plurality of steps. In other words, the press 11 may be equipped with a plurality of dies.
Load sensor 20: Detects the load applied to the workpiece. For example, it is configured with a strain gauge that detects the distortion of the frame 3.
Angle sensor 21: Detects the rotation angle (press angle) of the crankshaft 4 to detect the position of the slide 1 and the speed (stroke) of the slide 1.
Brake release pressure sensor 22: Detects the release pressure of the braking device 12.
Clutch pressure sensor 23: Detects the pressure of the clutch 9.
Clutch tank pressure sensor 24: Detects the pressure of the hydraulic oil that drives the clutch 9.
Brake tank pressure sensor 25: Detects the pressure of the cooling water for the brake device 12.
Coolant flow rate sensor 26: Detects the flow rate of the cooling water for the brake device 12.
BKO position displacement sensor 27: Detects the position of the lower knockout device of the press 11 (BKO position).
SKO position displacement sensor 28: Detects the position of the upper knockout device (SKO position).
Resistance temperature detector 29: Detects the temperature of each part of the press machine 11.
Lubricant flow rate sensor 30: Detects the flow rate of the lubricant that lubricates the die.
Air blow pressure sensor 31: Detects the pressure of the high-pressure air supplied to the press machine 11.
Material temperature sensor 32: Detects the temperature of the material supplied to the press machine 11.
Die height displacement sensor 33: Detects the die height of the press machine 11 (the distance between the bolster 2 and the bottom dead center position of the slide 1).
Mold temperature sensor 34: Detects the temperature of the mold.

As shown in FIG. 3, the forging press 10 further includes a conveying device 13 and a heater 14. The conveying device 13 is, for example, a belt conveyor, and conveys the workpiece W from the previous process to the press 11 (position of the bolster 2). The heater 14 is disposed upstream of the press 11 and is provided to prevent hardening of the workpiece W during transport by the conveying device 13. The heater 14 is, for example, a tunnel furnace.

The material temperature sensor 32 detects the temperature of the workpiece W immediately after it passes through the heater 14 (the workpiece W near the exit of the heater 14). The die temperature sensor 34 detects the temperature of the die mounted on the press machine 11.

In addition to the above-mentioned sensors, the forging press device 10 may further include a timer (not shown) that detects the time required for the workpiece W to reach the press position from the exit of the heater 14 (heater exit - press time). In addition, a workpiece presence/absence sensor (not shown) that detects the presence or absence of the workpiece W near the exit of the heater 14 may also be provided.

The press machine 11 in this embodiment is a multi-stage press machine that uses multiple dies to shape the workpiece W in stages. An example of the arrangement of dies is shown in FIG. 4. In the example of FIG. 4, the press machine 11 has three dies 51 to 53 arranged along the conveying direction A1 of the workpiece W. In this case, the workpiece W is formed in stages through a first process (process I) using die 51, a second process (process II) using die 52, and a third process (process III) using die 53, in that order.

The upper dies 51a, 52a, 53a of each of the dies 51 to 53 are fixed to the lower end of the slide 1 in the frame 3, and move up and down at the same time. The lower dies 51b, 52b, 53b of each of the dies 51 to 53 are fixed to the upper end of the bolster 2 in the frame 3. In the following description, when it is not necessary to distinguish between the three pairs of upper and lower dies 51a, 51b, 52a, 52b, 53a, 53b, they will be referred to as "dies 50".

The dies 51-53 of the press 11 are each lubricated by lubricating liquid supplied from a die lubrication means 60 (shown in phantom lines). In this case, a lubricating liquid flow sensor 30 and a die temperature sensor 34 are typically provided for each die 50.

The load sensor 20 is attached to one of the pillars of the frame 3 (for example, pillar 3c). A vibration meter 35 is attached to each of the left and right pillars 3c and 3d of the frame 3. One of the pillars 3c and 3d (pillar 3c) is located upstream in the conveying direction from the die 51 of step I, and the other of the pillars 3c and 3d (pillar 3d) is located downstream in the conveying direction from the die 53 of step III. Each vibration meter 35 is, for example, an acceleration sensor, and detects vibrations of the frame 3. Specifically, each vibration meter 35 detects vibrations in the approximately horizontal direction (front-back or left-right direction) of the pillars 3c and 3d to which it is attached.

Referring again to FIG. 1, the forging press device 10 is equipped with a die abnormality prediction device (hereinafter abbreviated as "abnormality prediction device") 200 that predicts the presence or absence of abnormalities in the dies 51 to 53. The abnormality prediction device 200 predicts the presence or absence of abnormalities in each die 50 based on the prediction model generated by the learning device 100. The learning device 100 and the abnormality prediction device 200 constitute a die abnormality prediction system SYS.

The learning device 100 and the abnormality prediction device 200 acquire detection signals from a plurality of sensors through the PLC 40 or without the PLC 40. The PLC 40 and these devices 100, 200 are connected by wire or wirelessly. The learning device 100 and the abnormality prediction device 200 are realized by a computer including a processor such as a CPU (Central Processing Unit) and a memory. Note that, although an example in which the PLC 40 and the abnormality prediction device 200 are provided separately is shown in this embodiment, this is not limiting, and the PLC 40 may have the functions of the abnormality prediction device 200 as described below.

<Learning device>
(Functional configuration)
5 is a block diagram showing the functional configuration of a learning device 100 according to this embodiment. The learning device 100 mainly includes a data acquisition unit 101, a format conversion unit 102, a result acquisition unit 103, a learning unit 104, and a model storage unit 114. The model storage unit 114 stores prediction models M1, M2, ... for each die pattern.

In this embodiment, the "die pattern" is specified by the product number (model number) corresponding to the shape of the dies 51 to 53, and the work pattern indicating the arrangement pattern of the work W on the dies 51 to 53. When the work W is continuously pressed by the forging press device 10, the product number of these work W is common. The arrangement pattern of the work W on the dies 51 to 53 includes a pattern in which the work W is arranged on all of the dies 51 to 53, a pattern in which the work W is arranged only on the first die 51, a pattern in which the work W is arranged only on the last die 53, and the like. Although the arrangement mode of the work W corresponding to the product number is determined one-to-one, since the work W is transported sequentially, at the start and end of the press operation, the work W is arranged only on some of the dies 51 to 53 in a pattern different from the arrangement mode of the work W. The arrangement mode of the work W corresponding to the product number includes a pattern in which the work W is arranged sequentially on all of the dies 51 to 53, a pattern in which the work W is arranged on the dies 51 to 53 every other die, and the like. The information for identifying the mold pattern may further include the projected area of the workpiece W.

The data acquisition unit 101 acquires measurement data from sensors equipped in the forging press device 10 when the press machine 11 is in operation (when the forging press device 10 is in press operation). In this embodiment, the data acquisition unit 101 acquires measurement data (hereinafter referred to as "cycle measurement data") and die pattern data for each press cycle.

The cycle measurement data is primary data and includes at least the load data obtained from the load sensor 20 and the vibration data obtained from the vibration meter 35. The cycle measurement data preferably further includes at least one of the temperature data of each mold 50 obtained from the mold temperature sensor 34, the lubricant data of each mold 50 obtained from the lubricant flow rate sensor 30, and the shot number data of each mold 50 counted for each press cycle.

The cycle measurement data acquired by the data acquisition unit 101 is stored in chronological order in the primary data storage unit 111, with a uniquely determined identification number (hereinafter referred to as "cycle No.") as a heading. Each cycle measurement data DT1 is stored in the primary data storage unit 111 together with the corresponding die pattern data. As described above, the die pattern data includes the product number and work pattern of the target workpiece W. The product number of the workpiece W is input by a user (operator or manager) via an input unit (not shown), for example, before the start of press operation by the forging press device 10. The work pattern is detected by a work pattern detection means (not shown) provided in the forging press device 10.

The format conversion unit 102 converts the format of the primary data into secondary data in a format suitable for machine learning. Specifically, the format conversion unit 102 functions as a feature calculation means, and calculates multiple types of feature amounts that become explanatory variables of the prediction model based on the cycle measurement data DT1. More specifically, multiple types of feature amounts including at least load feature amounts and vibration feature amounts are calculated for each press cycle.

Details of the types of features (secondary data) calculated by the format conversion unit 102 are shown in Table 1 below. Table 1 lists the types of features and the data items for each type. It also lists the type of each data item (scalar or vector).

The "load feature amount" is an index that can be calculated based on the load data obtained from the load sensor 20, and includes the load peak value or the load increase slope. The load increase slope is calculated as the slope of a linear expression that passes through two points, the rising position and the peak position of the molding load curve. In this embodiment, the load feature amount includes both the load peak value and the load increase slope. This makes it possible to accurately predict the presence or absence of an abnormality in each mold 50 that is directly related to the quality of the product.

The "mold contact time" is also an index that can be calculated based on the load data obtained from the load sensor 20, and the time between the rising point and the peak point of the molding load curve is calculated as the mold contact time. In this way, the mold contact time is calculated as a single characteristic value, not for each of the molds 51 to 53.

The feature value of the "die height value" is calculated from the output value (die height data) from the die height displacement sensor 33. By taking into account not only the load feature value but also the die height value, it becomes possible to more accurately predict the presence or absence of wear of the dies 51 to 53.

The feature value of "mold temperature" is calculated based on temperature data obtained from the mold temperature sensor 34 attached to each mold 50. For example, the mold temperature sensor 34 may be provided only on the lower molds 51b, 52b, 53b of each of processes I to III, and only the feature value of the temperature of the lower molds 51b, 52b, 53b may be calculated. The feature value of the temperature of each mold 50 (at least the lower molds 51b, 52b, 53b) is calculated as the average temperature of a specified area immediately before pressing (during workpiece transport), for example. Note that the feature value of the mold temperature may be the maximum temperature, minimum temperature, etc. in addition to/instead of the average temperature.

The characteristic quantity of "die lubricant quantity" is calculated based on lubricant quantity data obtained from a lubricant flow sensor 30 provided in the lubricant passage to each die 50. The characteristic quantity of the lubricant quantity for each die 50 is, for example, the lubricant quantity (total quantity) for each die 50 immediately before pressing (during workpiece transport).

The feature quantity of "frame vibration" is calculated based on vibration data obtained from vibration meters 35 provided on the left and right supports 3c, 3d of the frame 3. The feature quantities of the left and right frame vibrations include, for example, at least one of the wave height vector, time vector, and frequency vector in a specific section of one cycle. As shown in Table 1, it is preferable that the feature quantity of the frame vibration includes two items, the wave height vector and the frequency vector, but it may also include two items, the time vector and the frequency vector. In this way, it is preferable that it includes at least the frequency vector. A specific method for calculating the feature quantity of the frame vibration will be described later in detail.

The feature value of "number of mold shots" is a value common to the primary data, and includes the number of shots (cumulative) for each mold 50.

The cycle characteristic data DT2, which includes multiple types of characteristic quantities calculated by the format conversion unit 102, is stored as secondary data in the secondary data storage unit 112. Each cycle characteristic data DT2 is stored in the secondary data storage unit 112, for example, in association with the same cycle number and mold pattern data as the original cycle measurement data DT1.

The result acquisition unit 103 acquires die abnormality information input by the user via an input means such as the operation unit 120. The die abnormality information is information indicating, for example, which die 50 an abnormality occurred in which press cycle. The result acquisition unit 103 generates result data indicating the die state for each press cycle (hereinafter referred to as "cycle result data") based on the acquired die abnormality information. The cycle result data DT3 generated by the result acquisition unit 103 is stored in the result data storage unit 113, for example, in association with the cycle number. The "die state" represents the state (abnormal/absent) of each die 50. Note that each cycle result data DT3 is not limited to being stored in association with the cycle number as long as it is possible to identify which cycle measurement data DT1 or cycle characteristic data DT2 it corresponds to.

The learning unit 104 uses a large amount of cycle characteristic data DT2 stored in the secondary data storage unit 112 and a large amount of cycle result data DT3 stored in the result data storage unit 113 as training data to machine-learn the correlation between multiple types of feature values for each press cycle and the die state. In other words, it generates a prediction model for predicting the presence or absence of die abnormalities for each press cycle. As the machine learning algorithm for the prediction model, a well-known algorithm, for example, isolation forest, is used. Isolation forest is a general-purpose machine learning algorithm used for anomaly detection.

In this embodiment, the learning unit 104 performs machine learning to learn the correlation between multiple types of feature values and the mold state for each "mold pattern," and generates prediction models M1, M2, ... for the number of mold patterns (i.e., the number of combinations of part numbers and work patterns). The multiple prediction models M1, M2, ... generated by the learning unit 104 are stored in the model storage unit 114. In other words, the model storage unit 114 stores each prediction model in association with the mold identification information that identifies the mold pattern.

The functions of the data acquisition unit 101, format conversion unit 102, result acquisition unit 103, and learning unit 104 described above are realized by the processor executing software. The operation unit 120 is configured with a user interface such as a keyboard, mouse, or touch panel. The primary data storage unit 111, secondary data storage unit 112, result data storage unit 113, and model storage unit 114 are typically configured with non-volatile storage devices provided in the computer.

(About operation)
Fig. 6 is a flowchart showing a method for generating a prediction model by the learning device 100 according to this embodiment. The model generation process shown in Fig. 6 is realized by the processor of the learning device 100 executing a learning program stored in advance in the memory. This process is started in response to a user inputting an instruction to start learning via the operation unit 120. In this embodiment, it is assumed that before this process is started, a plurality of cycle measurement data DT1 are stored in the primary data storage unit 111 in chronological order in association with cycle numbers and mold pattern data. It is also assumed that a plurality of cycle result data DT3 are stored in the result data storage unit 113 in association with cycle numbers.

Referring to FIG. 6, the learning device 100 identifies one die pattern (in a predetermined order) from the combination of the part number of the workpiece W and the workpiece pattern (step S2), and searches the primary data storage unit 111 for cycle measurement data DT1 associated with that die pattern (step S4).

Then, the format conversion unit 102 reads all of the searched (target) cycle measurement data DT1 from the primary data storage unit 111 (step S6), and converts each of the read cycle measurement data DT1 into cycle characteristic data DT2 (step S8). In other words, the format conversion unit 102 calculates the multiple types of characteristic amounts shown in Table 1 above based on each data included in the cycle measurement data DT1, and records them in the secondary data storage unit 112 as cycle characteristic data DT2. Here, an example of calculating the characteristic amount for "frame vibration" will be described with reference to Figures 7 and 8.

FIG. 7(A) is a graph showing the stroke of the slide 1 along the time axis. FIG. 7(B) and (C) are graphs showing the transition of the magnitude of the vibration (pulse) along the same time axis as FIG. 7(A), and the unit of the vertical axis is, for example, "m/ ^s2 ". FIG. 8(B) is a graph showing the stroke of the slide 1, as in FIG. 7(A). FIG. 8(A) is a graph showing the brake release signal along the same time axis. FIG. 8(C) is a graph showing the transition of the magnitude of the vibration (pulse), as in FIG. 7(B) and (C). FIG. 8(D) is a graph showing the transition of the power spectrum density with the frequency on the horizontal axis (unit: Hz). Note that, in this example, it is assumed that the vibration is detected by a vibration meter 35 attached to the rear left support 3c of the frame 3.

A method for calculating the pulse height vector will be described with reference to Fig. 7(B). The format conversion unit 102 extracts pulses in a section TS that exceed a baseline (predetermined value) BL, and divides the extracted section TS into d rectangular pieces along the time axis. The format conversion unit 102 then processes the pulse data into a d-dimensional vector (p = [ _h1 , _h2 , ..., _hj , ..., _hd ]) and calculates the pulse height vector. Here, " _hj " is the average value within each rectangular piece.

A method of calculating the time vector will be described with reference to FIG. 7C. The format conversion unit 102 divides the pulse in the section TS into _dw in the crest direction, and obtains _2dw times that intersect with this division line (horizontal line). Then, the time data is processed into a 2dw _- dimensional vector (q=[ _w1 , _w2 , ..., _wj , ..., _w2d ]) to calculate the time vector. However, for " _wj ", two outer intersections with respect to the pulse apex position are selected. For example, when focusing on the division line Lx, the times t1 and t4 of the outer intersections are selected, and times t2 and t3 are not selected. Time t1 corresponds to the time when the upper mold of the molds 51 to 53 starts to contact the workpiece W, and time t4 corresponds to the time when the upper mold of the molds 51 to 53 starts to separate from the workpiece W.

The method of calculating the frequency vector will be described with reference to FIG. 8. The format conversion unit 102 extracts three types of analysis intervals from the vibration waveform shown in FIG. 8(C), namely, the "pre-stroke interval," the "pre-die contact interval," and the "die contact interval." The "pre-stroke interval" is the interval before time t11 when the brake release signal shown in FIG. 8(A) turns ON. The "pre-die contact interval" is the interval from time t11 to time t12 when the stroke position shown in FIG. 8(B) becomes the die contact position. The "die contact interval" is the interval from time t12 to time t13 when the brake release signal shown in FIG. 8(A) turns OFF.

The format conversion unit 102 performs a Fourier transform on the vibration signal for each analysis interval and calculates the power spectral density PSD. Figure 8 (D) shows the waveform of the power spectral density from 0 to fs/2 Hz. "fs" is the sampling frequency, which in this example is fs/2 = 50.

The format conversion unit 102, for example, sets the lower limit frequency _fmin = 0 and the upper limit frequency _fmax = fs/2, divides this section FS into d rectangular pieces (for example, 20 pieces), and processes the power spectral density PSD within the section FS into a d-dimensional vector (p = [ _h1 , _h2 , ..., _hj , ..., _hd ], where " _hj " is the average value within each rectangular piece) in the same manner as the wave-height vector. This allows the frequency vector for each analysis section to be obtained.

Of the three types of analysis intervals, only the frequency vectors of the "mold contact interval" that is strongly related to the dies 51 to 53 may be calculated. In other words, the frequency vectors of the "pre-stroke interval" and the "pre-mold contact interval" do not need to be included in the cycle characteristic data DT2.

The analysis interval used to calculate the frequency vector may also be used to calculate the wave height vector described above. That is, as shown in FIG. 8(C) as a schematic, each analysis interval may be divided into d rectangular pieces (e.g., 20 pieces), and the pulse data may be processed into a d-dimensional vector for each analysis interval to calculate the wave height vector. In this case, the format conversion unit 102 may calculate only the wave height vector for the "mold contact interval".

The cycle characteristic data DT2, which includes at least one of the vibration wave height vector, vibration time vector, and vibration frequency vector calculated by the above-mentioned method, is recorded in the secondary data storage unit 112 in association with the same cycle number as the target cycle measurement data DT1. As shown in FIG. 5, the cycle characteristic data DT2 may also be stored in association with the target mold pattern data.

Referring again to FIG. 6, next, the learning unit 104 reads all of the cycle result data DT3 corresponding to the cycle measurement data DT1 read in step S6 from the result data storage unit 113 (step S10). Specifically, the cycle result data DT3 associated with the same cycle number as the target cycle measurement data DT1 is read. This allows the result data corresponding to each cycle characteristic data DT2 obtained in step S8 (i.e., data indicating the presence or absence of anomalies in each type 50) to be extracted.

The learning unit 104 performs machine learning to learn the correlation between the cycle feature data DT2 obtained in step S8 and the cycle result data DT3 read out in step S10 (step S12). That is, the learning unit 104 generates a prediction model that takes multiple types of feature values as input and outputs the presence or absence of anomalies for each die 50 (individual die), i.e., the die state. After generating the prediction model, the learning unit 104 stores it in the model storage unit 114 in association with die identification information that identifies the die pattern (product number, work pattern) identified in step S2 (step S14).

If there are any unprocessed die patterns among the combinations of the part number of the workpiece W and the work pattern (YES in step S16), the process returns to step S2 and repeats the above process. When processing for all die patterns is completed (NO in step S16), the series of model generation processes ends. As a result, prediction models M1, M2, ... are stored in the model storage unit 114 for the number of combinations of the part number of the workpiece W and the work pattern. The example in Figure 5 shows that prediction model M1 is a trained model for part number "30" and work pattern No. "15", and prediction model M2 is a trained model for part number "80" and work pattern No. "15".

As described above, according to this embodiment, in addition to the load feature including the load peak value or load increase slope, at least the vibration feature of the left (upstream) support 3c and the vibration feature of the right (downstream) support 3d are used as explanatory variables of the prediction model. When an abnormality occurs in any of the multiple dies 51 to 53, depending on the position (process), a difference occurs in the left and right vibration feature, or even if there is almost no difference, both vibration feature values become abnormal values. Therefore, by including the left and right vibration feature in the explanatory variables of the prediction model, it is possible to predict which of the dies 51 to 53 has an abnormality. In addition, in this embodiment, by adding at least one of the temperature data, lubricant volume data, and shot number data for each die 50 to the explanatory variables of the prediction model, it is possible to accurately predict which of the upper and lower dies of the dies 51 to 53 has an abnormality.

<Abnormality prediction device>
(Functional configuration)
9 is a block diagram showing the functional configuration of an abnormality prediction device 200 according to this embodiment. The abnormality prediction device 200 mainly includes a model storage unit 213, a data acquisition unit 202, a format conversion unit 203, a prediction unit 204, and an output unit 230.

The model storage unit 213 prestores multiple prediction models M1, M2, ... generated by the learning device 100. In the model storage unit 213, each prediction model is also stored in association with the mold identification information. The learning device 100 and the abnormality prediction device 200 are connected, for example, via a network, and a communication unit (not shown) of the abnormality prediction device 200 receives the prediction models M1, M2, ... stored in the model storage unit 114 from a communication unit (not shown) of the learning device 100, and stores the received prediction models M1, M2, ... in the model storage unit 213. Alternatively, the prediction models M1, M2, ... stored in the model storage unit 114 of the learning device 100 may be stored in the model storage unit 213 using a removable recording medium.

After the press machine 11 starts operating, the data acquisition unit 202 acquires measurement data (cycle measurement data) DT11, which is the same type as the cycle measurement data DT1 in the learning device 100, together with the die pattern data for each press cycle.

The format conversion unit 203 functions as a feature calculation means, similar to the format conversion unit 102 of the learning device 100. That is, based on the cycle measurement data DT11, it calculates the multiple types of feature amounts shown in Table 1 above, and generates cycle feature data DT12 that includes these feature amounts.

The prediction unit 204 first determines the prediction model to be used for predicting anomalies in the dies 51-53 for each cycle based on the acquired die pattern data. Specifically, it determines the prediction model that corresponds to the die pattern in the current cycle from among the multiple prediction models M1, M2, ... stored in the model storage unit 213. It then inputs multiple types of feature amounts included in the cycle characteristic data DT12 into the determined prediction model to predict the presence or absence of anomalies in the dies 51-53. In other words, the "mold state" output by the prediction model becomes the predicted result.

The output unit 230 outputs the prediction results by the prediction unit 204. The output unit 230 outputs the presence or absence of an abnormality in each mold 50 for each press cycle. In addition, the output unit 230 may further output the importance of the data items listed in Table 1 (degree of influence on the prediction result) and the confidence level of the prediction result.

The functions of the data acquisition unit 202, format conversion unit 203, and prediction unit 204 described above are realized by the processor executing software. The model storage unit 213 is typically configured with a non-volatile storage device provided in the computer. The output unit 230 is typically configured with a display unit that displays the prediction results. Note that the output unit 230 may be configured with a communication unit that outputs the prediction results to the monitoring PC 42 shown in FIG. 1 via the PLC 40 or without the PLC 40.

(About operation)
Fig. 10 is a flowchart showing a method for predicting an abnormality of the dies 51 to 53 by the abnormality prediction device 200 according to the present embodiment. The abnormality prediction process shown in Fig. 10 is realized by the processor of the abnormality prediction device 200 executing an abnormality prediction program stored in advance in the memory.

Referring to FIG. 10, when the operation of the press machine 11 is started, the data acquisition unit 202 acquires cycle measurement data DT11, which includes output data from predetermined sensors equipped in the forging press device 10, together with die pattern data (step S26).

Then, the format conversion unit 203 converts the cycle measurement data DT11 acquired by the data acquisition unit 202 in step S26 into cycle characteristic data DT12 (step S28). That is, the format conversion unit 203 calculates the multiple types of characteristic amounts shown in Table 1 above based on each data included in the cycle measurement data DT11.

The prediction unit 204 also determines which prediction model to use this time from among the multiple prediction models M1, M2, ... stored in the model storage unit 213 based on the mold pattern data acquired by the data acquisition unit 202 in step S26 (step S29). The process of determining the prediction model may be performed in parallel with the data format conversion process.

The prediction unit 204 inputs the cycle characteristic data DT12 obtained in step S28 (i.e., multiple types of characteristic quantities for one cycle) into the prediction model determined in step S29, and predicts the presence or absence of anomalies in each mold 50 for each press cycle (step S30).

The output unit 230 outputs the prediction results, which are the output of the prediction model, i.e., the presence or absence of an abnormality for each mold 50 (step S32). As an example, as shown in FIG. 9, the presence or absence of an abnormality is displayed for each of the upper mold (mold I above) of mold 51, the lower mold (mold I below) of mold 51, the upper mold (mold II above) of mold 52, ..., and the lower mold (mold III below) of mold 53. This makes it easy to identify molds 50 that may be abnormal. Note that the abnormality prediction results for molds 51 to 53 may be output without distinguishing between upper and lower molds.

The output unit 230 also outputs, for example, a graph showing the degree of influence of the data items listed in Table 1 on the prediction results (output of the prediction model) in step S30, and outputs the confidence of the prediction results in step S30 as a numerical value (%) (step S33). The confidence may be output for the prediction results (abnormality present/abnormality not present) for each type 50.

In this manner, in this embodiment, i) the abnormality prediction results for the molds 51 to 53, ii) the importance of each data item relative to the prediction results, and iii) the confidence level of the prediction results are output. This makes it easy to take measures against the prediction results.

The above steps S26 to S33 are repeated until the press operation is completed (NO in step S34).

As described above, the mold abnormality prediction system SYS according to this embodiment includes the learning device 100. Therefore, by using the prediction model generated by the learning device 100, the abnormality prediction device 200 can detect (predict) abnormalities in the molds 51 to 53 without relying on the experience of the worker. Since the molds 51 to 53 have a higher frequency of failure than other parts, this embodiment can reduce the burden on the worker.

In addition, in this embodiment, the presence or absence of an abnormality in the dies 51 to 53 can be predicted in real time for each press cycle, so that abnormalities in the dies 51 to 53 can be discovered early. This makes it possible to prevent or suppress mass production of defective products by the press machine 11. Note that if an abnormality is predicted in any of the dies 50 by the prediction model, the operation of the press machine 11 may be automatically stopped.

(Modification)
In the present embodiment, the vibrometer 35 is attached to the rear pillars 3c and 3d of the frame 3, and the vibration feature values of the rear pillars 3c and 3d are used for abnormality prediction. However, this is not limiting, and the vibrometer 35 may be attached to the front pillars 3a and 3d, and the vibration feature values of the rear pillars 3c and 3d may be used for abnormality prediction. Alternatively, the vibrometer 35 may be attached to all four pillars 3a to 3d, and the vibration feature values of the four pillars 3a to 3d may be used for abnormality prediction.

In addition, in this embodiment, the learning unit 104 generates a prediction model that outputs the presence or absence of an abnormality in each die 50, but it is also possible to generate a prediction model that outputs the presence or absence of an abnormality in the dies 51 to 53 for each process equipped in the press machine 11, regardless of whether they are upper or lower dies.

In addition, in this embodiment, an example has been described in which the press machine 11 is equipped with dies 51 to 53, but the number of dies is not limited to three. Also, the number of dies is not limited to multiple dies, and may be one.

In addition, the prediction models M1, M2, ... are stored in advance in the model storage unit 213 of the anomaly prediction device 200, but the prediction models M1, M2, ... in the model storage unit 213 may be updated in response to machine learning by the learning device 100. Alternatively, the anomaly prediction device 200 may have the functions of the learning device 100.

In addition, in this embodiment, the forging press device 10 is equipped with the abnormality prediction device 200, but this is not limiting, and the abnormality prediction device 200 may predict the presence or absence of abnormalities in the dies 51 to 53 offline.

The learning method executed by the learning device 100 can also be provided as a program. Similarly, the abnormality prediction method for the molds 51 to 53 executed by the abnormality prediction device 200 can also be provided as a program. Such a program can be provided by recording it on an optical medium such as a CD-ROM (Compact Disc-ROM) or a computer-readable non-transitory recording medium such as a memory card. The program can also be provided by downloading it over a network.

The program according to the present invention may execute processing by calling necessary modules in a specified sequence at a specified timing among program modules provided as part of a computer's operating system (OS). In such a case, the program itself does not include the above modules, and executes processing in cooperation with the OS. Programs that do not include such modules may also be included in the program according to the present invention.

The program according to the present invention may also be provided as being incorporated into a part of another program. In this case, the program itself does not include modules included in the other program, and the processing is carried out in cooperation with the other program. Such a program incorporated into another program may also be included in the program according to the present invention.

The embodiments disclosed herein should be considered to be illustrative and not restrictive in all respects. The scope of the present invention is indicated by the claims, not the above description, and is intended to include all modifications within the meaning and scope of the claims.

3 Frame, 3a, 3b, 3c, 3d Support, 10 Forging press device, 11 Press machine, 20 Load sensor, 35 Vibration meter, 51, 52, 53 Die, 100 Learning device, 101, 202 Data acquisition unit, 102, 203 Format conversion unit, 103 Result acquisition unit, 104 Learning unit, 111 Primary data storage unit, 112 Secondary data storage unit, 113 Result data storage unit, 114, 213 Model storage unit, 120 Operation unit, 200 Abnormality prediction device, 204 Prediction unit, 230 Output unit, DT1, DT11 Cycle measurement data, DT2, DT12 Cycle characteristic data, DT3 Cycle result data, M1, M2 Prediction model, SYS Die abnormality prediction system, W Work.

Claims (10)

A learning device that generates a prediction model by machine learning to predict the presence or absence of an abnormality in a die mounted on a press machine,
a data acquisition means for acquiring measurement data for each press cycle, the measurement data including load data applied to a workpiece during press operation by the press machine and vibration data in a substantially horizontal direction detected by a vibrometer attached to a support column of a frame of the press machine;
a feature calculation means for calculating a plurality of types of feature amounts including a load feature amount and a vibration feature amount for each press cycle based on the measurement data acquired by the data acquisition means;
a generation means for performing machine learning on the correlation between the plurality of types of feature amounts calculated by the feature amount calculation means and a die state, and generating a prediction model for predicting the presence or absence of an abnormality in the die for each press cycle;
A learning device comprising: a model storage means for storing the predictive model generated by the generation means. A learning device that generates a prediction model by machine learning to predict the presence or absence of an abnormality in a die mounted on a press machine,
a data acquisition means for acquiring measurement data including load data applied to a workpiece during press operation by the press machine and vibration data of a frame of the press machine for each press cycle;
a feature calculation means for calculating a plurality of types of feature amounts including a load feature amount and a vibration feature amount for each press cycle based on the measurement data acquired by the data acquisition means;
a generation means for performing machine learning on the correlation between the plurality of types of feature amounts calculated by the feature amount calculation means and a die state, and generating a prediction model for predicting the presence or absence of an abnormality in the die for each press cycle;
a model storage means for storing the prediction model generated by the generation means,
The vibration feature includes at least one of a wave height vector, a time vector, and a frequency vector. The generating means performs machine learning on correlations between a plurality of types of feature amounts and a mold state for each mold pattern,
3. The learning device according to claim 1, wherein the model storage means stores the prediction model in association with die identification information for identifying a die pattern. The press machine is a multi-stage press machine equipped with a plurality of dies,
The learning device according to claim 1 , wherein the generating means generates a prediction model that receives a plurality of types of feature quantities as input and outputs the presence or absence of an abnormality for each individual die. the frame has first and second support columns provided on the upstream and downstream sides of the plurality of dies, respectively;
5. The learning device according to claim 4, wherein the feature calculation means calculates a first vibration feature based on vibration data from a vibrometer attached to the first support, and calculates a second vibration feature based on vibration data from a vibrometer attached to the second support. The learning device according to any one of claims 1 to 5, wherein the feature calculation means further calculates at least one of a die contact time, a die height value, a die temperature, a die lubricant amount, and a die shot count in addition to the load feature and the vibration feature. A die abnormality prediction device for predicting the presence or absence of an abnormality in a die mounted on a press machine,
A storage means for storing a prediction model generated by the learning device according to any one of claims 1 to 6 ;
a data acquisition means for acquiring measurement data including load data applied to a workpiece during press operation by the press machine and vibration data of a frame of the press machine for each press cycle;
a feature calculation means for calculating a plurality of types of feature amounts including a load feature amount and a vibration feature amount for each press cycle based on the measurement data acquired by the data acquisition means;
a prediction means for inputting the plurality of types of feature quantities calculated by the feature quantity calculation means into a prediction model stored in the storage means, and predicting the presence or absence of an abnormality in the die for each press cycle;
An output means for outputting a prediction result by the prediction means. The storage means stores a plurality of prediction models in association with die identification information for identifying a die pattern,
The die abnormality prediction device according to claim 7, further comprising a model discrimination means for discriminating a prediction model to be used for die abnormality prediction from among the plurality of prediction models, based on a die pattern corresponding to a current press operation, before the press operation is started by the press machine . A learning program that generates a predictive model by machine learning to predict the presence or absence of an abnormality in a die mounted on a press machine,
calculating a plurality of types of feature quantities including a load feature quantity and a vibration feature quantity for each press cycle based on measurement data including load data applied to a workpiece during press operation by the press machine and vibration data in a substantially horizontal direction detected by a vibrometer attached to a support column of a frame of the press machine ;
A step of performing machine learning on the correlation between the calculated multiple types of feature amounts and the die state to generate a prediction model for predicting the presence or absence of an abnormality in the die for each press cycle;
and storing the generated prediction model in a model storage means. A learning program that generates a predictive model by machine learning to predict the presence or absence of an abnormality in a die mounted on a press machine,
calculating a plurality of types of feature quantities, including a load feature quantity and a vibration feature quantity including at least one of a wave height vector, a time vector, and a frequency vector, for each press cycle, based on measurement data including load data applied to a workpiece during a press operation by the press machine and vibration data of a frame of the press machine;
A step of performing machine learning on the correlation between the calculated multiple types of feature amounts and the die state to generate a prediction model for predicting the presence or absence of an abnormality in the die for each press cycle;
and storing the generated prediction model in a model storage means.

Images