JP7239377B2 - Machine learning device, failure prediction device, control device, and printed circuit board - Google Patents

Description

The present invention relates to a machine learning device, a failure prediction device, a control device, and a printed circuit board.

Components such as ICs (Integrated Circuits) and capacitors on printed boards are joined to PWBs (Printed Wiring Boards) by soldering.
A jointed portion (hereinafter referred to as a "solder joint") may crack or peel off a pad on the printed board side due to stress applied to the printed board or distortion due to vibration or the like. If this solder joint defect occurs in, for example, an important part of a printed circuit board of a control device that controls a machine tool, such as a CPU (Central Processing Unit), a CNC (Computerized Numerical Control) malfunctions. put away. Then, machining by the machine tool stops and downtime occurs. Factors of distortion include vibration during machining of the machine tool, impact when opening and closing the door of the machine tool, thermal deformation of the printed board, and the like.
Therefore, a technique is known for detecting the occurrence of defective solder joints in components on printed boards. See Patent Documents 1 and 2, for example.

Japanese Patent Application Laid-Open No. 2008-140800 JP-A-2005-265579

However, even if it is possible to detect the occurrence of defects in the solder joints from the amount of strain measured on the printed board, it is sometimes difficult to detect the signs of such defects.

Therefore, it is desired to detect a sign of a defective solder joint before the CNC malfunctions during processing due to the defective solder joint.

(1) One aspect of the machine learning device of the present disclosure is the above-described Using state data indicating the state of the environment in which the printed circuit board is mounted, each day included in a range having the date of failure as the upper limit and the day before the date of failure as the lower limit is set as a base date. Input data for extracting, as status data for each base date, a set of status data in a range that traces back a second period shorter than the first period, and generating status data for a plurality of the extracted base dates as input data. a generation unit, if at least one of the plurality of state data of the reference date includes at least one piece of state data included in a third period that is equal to or less than the length of the second period from the date of failure; If it is assumed that a failure will occur within the period, and the state data included in the third period from the date of the failure is not included, label generation that generates label data indicating that no failure will occur within the third period from the base date. a learning unit that performs supervised learning using a unit, input data generated by the input data generation unit, and label data generated by the label generation unit, and generates a trained model; Prepare.

(2) One aspect of the failure prediction device of the present disclosure includes a learned model acquired by learning by the machine learning device of (1), a storage unit that stores the state data, and Among the state data, the state data in the past range up to the second period before the present is input to the learned model, and a failure occurs in the printed circuit board within the third period. and a determination unit that determines whether or not.

(3) One aspect of the failure prediction device of (2) includes the machine learning device of (1).

(4) One aspect of the control device of the present disclosure includes the failure prediction device of (2) or (3).

(5) One aspect of the printed circuit board of the present disclosure includes the failure prediction device of (2) or (3).

According to one aspect, a sign of a defective solder joint can be detected before the defective solder joint causes the CNC to malfunction during processing.

1 is a functional block diagram showing a functional configuration example of a numerical control system according to a first embodiment; FIG. It is a figure which shows an example of the external appearance of the printed board mounted in a control apparatus. FIG. 2B is a diagram showing an example of a state in which the part of FIG. 2A is viewed from the side; FIG. 4 is a diagram illustrating an example of generation of input data using strain data of an arbitrary printed board; 2 is a diagram showing an example of a learned model constructed by the machine learning device of FIG. 1; FIG. 4 is a flowchart for explaining failure prediction processing of a failure prediction device in an operation phase; It is a functional block diagram showing a functional configuration example of a numerical control system according to a second embodiment. FIG. 7 is a diagram showing an arrangement example of strain gauges and temperature sensors on the printed board of FIG. 6; FIG. 4 is a diagram illustrating an example of generation of input data using strain data and temperature data of an arbitrary printed board; 7 is a diagram showing an example of a learned model constructed by the machine learning device of FIG. 6; FIG. 4 is a flowchart for explaining failure prediction processing of a failure prediction device in an operation phase; FIG. 11 is a functional block diagram showing a functional configuration example of a numerical control system according to a third embodiment; FIG. FIG. 4 is a diagram illustrating an example of generation of input data using strain data, temperature data, and integrated data of an arbitrary printed board; 12 is a diagram showing an example of a learned model constructed by the machine learning device of FIG. 11; FIG. 4 is a flowchart for explaining failure prediction processing of a failure prediction device in an operation phase; FIG. 11 is a functional block diagram showing a functional configuration example of a numerical control system according to a fourth embodiment; FIG. 16 is a diagram showing an example of a learned model constructed by the machine learning device of FIG. 15; FIG. 4 is a flowchart for explaining failure prediction processing of a failure prediction device in an operation phase;

A first embodiment will be described below with reference to the drawings.
In the first embodiment, strain data measured by a sensor unit on a printed board is exemplified as state data. Further, as a failure, a failure caused by a defective solder joint is exemplified.
Unless otherwise specified, the "predetermined period" shall be "1 year", the "first period" shall be "13 months", the "second period" shall be "1 month", and the "third period" shall be " 1 week." Note that the "predetermined period", "first period", "second period", and "third period" are not limited to these and can be determined as appropriate.
<First Embodiment>
FIG. 1 is a functional block diagram showing a functional configuration example of a numerical control system according to the first embodiment. As shown in FIG. 1, the numerical control system has a machine tool 100, a server 200 and a machine learning device 300. Machine tool 100 incorporates control device 110 on which printed circuit board 120 is mounted. As a result, the control device 110 and the printed board 120 are subjected to stress, vibration, and the like that accompany the operation of the machine tool 100 .
Machine tool 100 and server 200, and server 200 and machine learning device 300 may be connected to each other via a network (not shown) such as a LAN (Local Area Network) or the Internet. Machine tool 100 and machine learning device 300 may be directly connected to each other via a connection interface (not shown), or may be connected via a network (not shown).

The machine tool 100 is a machining machine that operates under numerical control of a control device 110 .
Note that the present embodiment is not limited to the machine tool 100, and can be widely applied to industrial machines in general. Industrial machinery includes various machines such as, for example, machine tools, industrial robots, service robots, forging machines and injection molding machines.
Controller 110 is a numerical controller known to those skilled in the art and controls the operation of machine tool 100 . If the machine tool 100 is a robot or the like, the control device 110 may be a robot control device or the like.

Control device 110 is included in failure prediction device 150 in the operation phase, as will be described later. Specifically, the failure prediction device 150 selects the state data measured by the n strain gauges 130(1) to 130(n) arranged on the printed circuit board 120, and uses the present as the base point to determine the state data before the second period. state data that goes back to the past is input to the learned model provided by the machine learning device 300, and a failure occurs in the printed circuit board 120 within the third period due to a defective solder joint. (n is an integer equal to or greater than 1).
In addition, hereinafter, the strain gauges 130(1) to 130(n) may be collectively referred to as "strain gauges 130" when there is no need to distinguish them individually.

Before describing the failure prediction device 150, machine learning for generating a trained model will be described.

<Print board 120>
As shown in FIG. 1, a strain gauge 130 is arranged on the printed board 120 and is configured to measure the amount of strain of the printed board 120 at each arranged position. Each of the strain gauges 130 outputs a signal containing the measured strain amount to the control device 110 as state data.
The sampling time during which the strain gauge 130 measures the amount of strain may be about 1 millisecond to several milliseconds, since the vibration generated by the operation of the machine tool 100 has a frequency of several kHz. However, the sampling time is not limited to this, and may be set as appropriate according to the operating state of machine tool 100 .
FIG. 2A is a diagram showing an example of the appearance of printed circuit board 120 mounted on control device 110 .
As shown in FIG. 2A, for example, a plurality of components 121 to 129 are arranged on the printed board 120 by soldering to a PWB.

Moreover, the strain gauge 130 is arranged on the back surface of the printed board 120 as shown in FIG. 2B described later.
Note that the components bonded to the printed board 120 include, for example, an IC including a CPU, a capacitor, a memory such as a ROM (Read Only Memory) or a RAM (Random Access Memory), an LED (Light Emitting Diode), and an external output terminal. , an input device (not shown) such as a keyboard and a touch panel included in the control device 110, and a connector for connecting to an output device (not shown) such as a liquid crystal display.

FIG. 2B is a diagram showing an example of a state in which the component 121 of FIG. 2A is viewed from the side. As shown in FIG. 2B, for example, component 121 is bonded onto printed circuit board 120 with a plurality of solder joints 40 . A strain gauge 130(2) is arranged on the back surface of the printed board 120 on which the component 121 is arranged. Note that FIG. 2B shows an example in which a crack 42 occurs in one solder joint 40 among a plurality of solder joints 40 and the solder joint 40 becomes defective.

<Control device 110 in the learning phase>
In the learning phase, the control device 110 collects teacher data necessary for performing machine learning. Therefore, as described above, the control device 110 stores the state data indicating the state of the environment in which the printed board 120 is mounted, which is measured by the n strain gauges 130 arranged on the printed board 120 to be mounted, as follows: Obtained from the printed board 120 . Control device 110 outputs the acquired state data to server 200 via a communication unit (not shown) of machine tool 100 .

The control device 110 in the learning phase includes an input section 151 and a storage section 153 . The control device 110 in the learning phase does not have to include the determination unit 152 and the trained model 1531 .

The input unit 151 inputs strain data measured by strain gauges 130 arranged on a printed board 120, which will be described later, via an input/output interface (not shown) of the control device 110, for example. The input unit 151 outputs the input strain data to the storage unit 153 as state data.

The storage unit 153 is, for example, a nonvolatile memory such as an EEPROM (Electrically Erasable Programmable Read Only Memory).
The storage unit 153 has a data area 1532 for storing strain data of the printed board 120 measured by each strain gauge 130 and acquired from the input unit 151 .

Note that the control device 110 may transmit the strain data stored in the data area 1532 to the server 200 when the amount of strain data stored in the data area 1532 reaches a predetermined amount, for example. Server 200 may store the received strain data measured by strain gauge 130 . As a result, control device 110 can overwrite strain data measured by strain gauge 130 to be newly acquired and store it in data area 1532 .

After that, in the learning phase, when a failure occurs in the printed board 120 due to, for example, a defect in the solder joint 40 and the printed board 120 becomes abnormal, the control device 110 controls the liquid crystal display of the control device 110. The failure is displayed on the output device (not shown).
After confirming the failure display on the output device (not shown), the user (service personnel) of the control device 110 confirms whether or not there is strain data that has not been transmitted to the server 200 in the data area 1532 of the storage unit 153. . If there is strain data that has not been sent to the server 200 and the analysis of the printed board 120 reveals that the failure of the printed board 120 is due to the defective solder joints 40 on the printed board 120, the user The transmitted strain data is transmitted to the server 200 . As a result, the server 200 can acquire all the strain data measured by the strain gauges 130 during the period from when the printed circuit board 120 starts working until it fails.

<Machine learning device 300>
The machine learning device 300 acquires from the server 200, for example, state data measured over 13 months, which is longer than one year, up to the date when an arbitrary printed circuit board fails. Using the acquired state data, the machine learning device 300 uses each date included in a range with the date of failure as the upper limit and the date one year before the date of the failure as the lower limit. A set of status data in the range going back to the past is extracted as status data for each base date, and the extracted status data for a plurality of base dates are generated as input data.
In addition, the machine learning device 300 generates, as a label (correct answer), data indicating whether or not a failure will occur within one week for each of the plurality of base date status data of the generated input data.
The machine learning device 300 performs supervised learning using the generated training data of input data and label pairs, and constructs a trained model described later.
By doing so, the machine learning device 300 can provide the built learned model 1531 to the failure prediction device 150 included in the control device 110 .
The machine learning device 300 will be specifically described.

The machine learning device 300 has an input data generation unit 301, a label generation unit 302, a learning unit 303, and a storage unit 304, as shown in FIG.

In the learning phase, the input data generation unit 301 selects strain data (state data) measured over a period of 13 months longer than one year up to the date of failure on an arbitrary printed circuit board, with the date of failure as the upper limit. A set of strain data in the range going back to the past one month with each day included in the range with the day one year before the failure date as the lower limit as the base date, as the strain data for each base date Extract. The input data generation unit 301 generates the extracted strain data of the plurality of base dates as input data. The input data generation unit 301 outputs the generated input data to the storage unit 304 .
More specifically, for example, the input data generator 301 may acquire from the server 200 strain data measured over a period of thirteen months longer than one year from the time of failure in any printed circuit board. The input data generation unit 301, from the acquired strain data, as shown in FIG. The date of failure is extracted as the strain data of the base date. Next, the input data generation unit 301 extracts the strain data included in the past range 55 up to one month before the day before the failure date as the base date, with the date one day before the failure date as the base date. Extract as strain data. Then, the input data generation unit 301 uses the day one year before the failure date, that is, the day 364 days before the failure date as the base date, and the strain data included in the past range 55 up to one month ago, This is repeated until 364 days before the failure date is extracted as the strain data of the base date. As a result, the input data generation unit 301 can generate distortion data of the base date of 365 patterns as input data.

FIG. 3 is a diagram illustrating an example of generation of input data using strain data for an arbitrary printed board.
The graph shown on the left side of FIG. 3 shows the time variation of the amount of strain measured over a period of 13 months longer than one year from the date d1 when an arbitrary printed board failed. A solid line indicates the amount of strain. Also, dashed lines indicate intervals of one month.
In the graph shown on the right side of FIG. 3, the data of the graph shown on the left side of FIG. It indicates that the date is traced one day at a time from the date of failure.
Note that, for example, when there are k arbitrary strain data of printed boards, the input data generation unit 301 can generate strain data of the base date of k×365 patterns as input data (k is 2 or more). integer).

The label generation unit 302 generates data indicating whether or not a failure will occur within one week for each of the distortion data of the 365 patterns of the input data, as label data (correct data). The label generation section 302 outputs the generated label data to the storage section 304 .
More specifically, the label generator 302 includes, for example, at least one strain data included in one week from the date of failure among the 365 patterns of the strain data of the base date generated by the input data generator 301. If the strain data included in one week from the date of the failure is not included, the label data is generated as "failure will not occur within one week". In this case, the strain data for 7 base-point days is labeled "failure will occur within 1 week", and the strain data for 358 base-point days is labeled "no failure will occur within 1 week". Become.
Note that the label is not limited to "failure will occur within one week" or not, and may be represented by binary values such as "1" and "0".

The learning unit 303 receives pairs of the above-mentioned input data and labels as training data from the storage unit 304, and uses the received training data to perform supervised learning so that the strain gauges 130 of the printed board 120 to be monitored Build a trained model 1531 that determines whether or not "a failure will occur within a week" based on the strain data in the past up to one month ago, starting from the present measured by do. In the following description, the strain data in the past one month from the present is also referred to as "past one-month strain data from the present".
Here, it is desirable to prepare a large number of training data for supervised learning. For example, strain data (state data) for any printed circuit board may be obtained in a laboratory vibration test. Alternatively, the training data may be acquired from each of the control devices 110 in various places that are actually in operation at the customer's factory or the like. This makes it possible to acquire training data under various circumstances.

The learning unit 303 then provides the constructed trained model 1531 to the control device 110 .
FIG. 4 is a diagram showing an example of a trained model 1531 constructed by the machine learning device 300 of FIG. Here, as shown in FIG. 4, the trained model 1531 indicates whether or not "failure will occur within one week" with strain data for the past one month starting from the present as an input layer. A multilayer neural network with data as an output layer is exemplified.
In this case, in the operation phase, the failure prediction device 150 sequentially inputs strain data for the past one month from the present measured by each strain gauge 130 to the trained model 1531. It is possible to determine whether or not "failure will occur within one week". Then, for example, the failure prediction device 150 determines that "a failure will occur within one week" with respect to the strain data for the past one month from the present measured by at least one strain gauge 130. In this case, it is possible to predict a sign that a defect will occur in the solder joint 40 on the printed circuit board 120 within one week.

Note that, when learning unit 303 acquires new teacher data after constructing trained model 1531, learning unit 303 further performs supervised learning on trained model 1531 to obtain trained model 1531 constructed once. may be updated.
Also, the trained model 1531 may be shared between the printed boards 120 of the same type. In other words, machine learning device 300 constructs trained model 1531 for each different type of printed circuit board 120 and provides failure prediction device 150 with trained model 1531 corresponding to the type of printed circuit board 120 . may
The supervised learning described above may be performed by online learning, batch learning, or mini-batch learning.
Online learning is a learning method in which supervised learning is performed immediately each time teacher data is created. In addition, batch learning is a learning method in which, while the creation of teacher data is repeated, multiple teacher data are collected according to the repetition, and supervised learning is performed using all the collected teacher data. is. Furthermore, mini-batch learning is an intermediate learning method between online learning and batch learning, in which supervised learning is performed each time a certain amount of teacher data is accumulated.
The storage unit 304 is a RAM or the like, and stores the input data generated by the input data generation unit 301, the label data generated by the label generation unit 302, the trained model 1531 constructed by the learning unit 303, and the like.
Machine learning for generating the trained model 1531 included in the failure prediction device 150 has been described above.
Next, the failure prediction device 150 will be described.

<Failure prediction device 150>
As shown in FIG. 1 , the failure prediction device 150 includes an input section 151 , a determination section 152 and a storage section 153 . Here, the input unit 151 is as described in the learning phase, and the storage unit 153 is also as described in the learning phase, except for the trained model 1531 .
As described in the learning phase, the input unit 151 inputs strain data measured by the strain gauges 130 of the printed board 120 via, for example, an input/output interface (not shown) of the control device 110 . The input unit 151 outputs the input strain data to the storage unit 153 .

Of the strain data stored in the storage unit 153, the determination unit 152 inputs each strain data of the strain gauge 130 in a past range up to one month ago with the present as the base point to the learned model 1531. to determine whether or not a failure will occur within one week. Then, when there is strain data determined that a failure will occur within one week, the determination unit 152 outputs a sign of failure of the solder joints 40 on the printed board 120 to an output device (not shown) of the control device 110. .
The storage unit 153 stores the learned model 1531 constructed by the machine learning device 300 and the like.

<Failure Prediction Processing of Failure Prediction Device 150 in Operation Phase>
Next, the operation related to the failure prediction process of the failure prediction device 150 will be described.
FIG. 5 is a flowchart for explaining failure prediction processing of the failure prediction device 150 in the operation phase.
Referring to FIG. 5 , in step S 11 , the input unit 151 inputs strain data on the printed board 120 measured by each of the strain gauges 130 and outputs the input strain data to the storage unit 153 .

In step S12, the determining unit 152 inputs strain data for the past one month from the current point of each strain gauge 130 out of the strain data stored in the storage unit 153 into the learned model 1531, A failure will occur within a week" or not.

In step S13, the determination unit 152 determines whether or not there is strain data for which it is determined that "failure will occur within one week" in the determination of step S12. If there is strain data determined that "a failure will occur within one week", the process proceeds to step S14, and if there is no strain data determined that "a failure will occur within one week", the process proceeds to step S11. move.

In step S<b>14 , determination unit 152 outputs a sign of failure of solder joints 40 on printed circuit board 120 to an output device (not shown) of control device 110 .

As described above, the failure prediction apparatus 150 according to the first embodiment uses the strain data (state data) previously measured over a period of 13 months longer than one year from the date when an arbitrary printed board failed. Then, the strain data up to one month prior to each date included in the range with the date of failure as the upper limit and the day one year before the date of failure as the lower limit as the strain data for each base date Training using strain data of a plurality of extracted base dates as input data, and data indicating whether or not a failure will occur within one week for each of the strain data of the plurality of base dates as label data. By inputting strain data for the past one month measured on the printed board 120 to be monitored from the present to the trained model 1531 acquired by performing supervised learning with data, within one week Determine whether a failure will occur.
As a result, the failure prediction device 150 can detect a sign of failure of the solder joints 40 before the failure of the solder joints 40 causes malfunction of the CNC during processing. Then, the failure prediction device 150 prevents the printed board 120 from being operated while the machine tool 100 is not in operation, before the control device 110 malfunctions during processing due to a defect in the solder joints 40 of the monitored printed board 120 . It can be replaced, and the downtime of the machine tool 100 can be reduced.
The first embodiment has been described above.

<Second embodiment>
In the second embodiment, as state data, strain data measured on a printed board by a sensor unit and temperature data measured on a printed board by a temperature sensor are exemplified. And this point is different from the first embodiment.

FIG. 6 is a functional block diagram showing a functional configuration example of a numerical control system according to the second embodiment. Note that functional blocks having functions similar to those in FIG.
As shown in FIG. 6, the machine tool 100 incorporates a control device 110A on which a printed board 120A is mounted.
Control device 110A is included in failure prediction device 150A in the operation phase, as will be described later.
Before describing failure prediction device 150A, machine learning for generating a trained model will be described.

<Printed board 120A>
As shown in FIG. 6, the printed circuit board 120A has n strain gauges 130 and m temperature sensors 135(1) to 135(m) (m is 1 or more). integer).
In the following description, the temperature sensors 135(1) to 135(m) are also collectively referred to as "temperature sensors 135" when there is no need to distinguish them individually.
The strain gauge 130 measures the strain amount of the printed board 120A at each arranged position, and the temperature sensor 135 measures the temperature of the printed board 120A at each arranged position. Then, strain gauge 130 outputs a signal including the measured strain amount to control device 110 as strain data. Temperature sensor 135 also outputs a signal including the measured temperature to control device 110 as temperature data.
FIG. 7 is a diagram showing an arrangement example of the strain gauges 130 and the temperature sensors 135 on the printed board 120A of FIG.
The strain gauge 130 and the temperature sensor 135 may be arranged adjacent to each other or may be arranged at separate positions.
Also, the sampling time during which the temperature sensor 135 measures the temperature may be about 1 millisecond to several milliseconds, similar to the strain gauge 130 . However, if the temperature of the printed circuit board 120A does not change significantly in milliseconds, the temperature data obtained from the temperature measured by the temperature sensor 135 is the average value for each hour or day, or the maximum value for each hour or day. etc.

<Control Device 110A in Learning Phase>
In the learning phase, the control device 110A collects teacher data necessary for machine learning in the same way as the control device 110 in FIG. Therefore, as described above, the control device 110 controls the temperature of the environment in which the printed board 120A is mounted, which is measured by the n strain gauges 130 and the m temperature sensors 135 arranged on the mounted printed board 120A. State data indicating the state is obtained from the printed board 120A. Control device 110A outputs the acquired state data to server 200 via a communication unit (not shown) of machine tool 100 .

110 A of control apparatuses in a learning phase are provided with 151 A of input parts, and the memory|storage part 153. FIG. 110 A of control apparatuses in a learning phase do not need to be equipped with the determination part 152A and the learned model 1531A.

The input unit 151A receives, for example, strain data measured by each strain gauge 130 arranged on a printed board 120A described below and temperature measured by each temperature sensor 135 via an input/output interface (not shown) of the control device 110. Enter data as state data. The input unit 151A outputs the input strain data and temperature data to the storage unit 153 .

Note that, in the control device 110A, for example, similarly to the control device 110 in FIG. Strain data and temperature data may be transmitted to server 200 . Then, the controller 110 may overwrite the newly obtained strain data and temperature data measured by the strain gauge 130 and the temperature sensor 135 and store them in the data area 1532 .
Also, in the learning phase, when the printed board 120A is in an abnormal state, the control device 110A has strain data and temperature data that have not been transmitted to the server 200 in the data area 1532 of the storage unit 153, and the failure of the printed board 120A is detected. is found to be due to a defect in the solder joint 40, the unsent strain data and temperature data may be sent to the server 200 based on the user's operation.

<Machine learning device 300A>
The machine learning device 300A has an input data generation unit 311, a label generation unit 312, a learning unit 313, and a storage unit 304, as shown in FIG.
The input data generator 311 has a strain data extractor 340 and a temperature data extractor 341 .
More specifically, in the learning phase, the input data generator 311 collects strain data and temperature data (state data) measured over a period of 13 months longer than one year from the failure of an arbitrary printed board. It may be acquired from the server 200 .
The strain data extracting unit 340 extracts the strain data included in the past range 55 from the obtained strain data, as shown in FIG. , the date of failure is extracted as the strain data of the reference date. Next, the strain data extracting unit 340 extracts the strain data included in the past range 55 up to one month before, with the date one day before the failure date as the base date, and extracts the strain data from the date one day before the failure date. Extract as strain data. Then, the strain data extraction unit 340 extracts the strain data included in the past range 55 up to one month ago with the day one year before the failure date, that is, the day 364 days before the failure date as the base date, This is repeated until 364 days before the failure date is extracted as the strain data of the base date.
Further, the temperature data extraction unit 341 extracts the temperature data included in the past range 55 from the acquired temperature data up to one month before, which is indicated by the thick-line rectangle, with the date of failure as the base date. Extract as the temperature data of the base date. Next, the temperature data extracting unit 341 extracts the temperature data included in the past range 55 up to one month before, with the date one day before the failure as the base date, and Extract as temperature data. Then, the temperature data extracting unit 341 extracts the temperature data included in the past range 55 up to one month before, with the day one year before the failure date, that is, the day 364 days before the failure date as the base date. This is repeated until 364 days before the failure date is extracted as the temperature data of the base date.
The input data generation unit 311 sets the strain data and temperature data of the same base date among the extracted strain data and temperature data of each base date, and sets the strain data and temperature data of the base date of 365 patterns as input data. Generate as The input data generation unit 311 outputs the generated input data to the storage unit 304 .
FIG. 8 is a diagram illustrating an example of generation of input data using strain data and temperature data for an arbitrary printed board.
The graph on the left side of FIG. 8 shows temporal changes in strain amount and temperature measured over a period of 13 months longer than one year from the date d1 when an arbitrary printed board failed. The solid line indicates the amount of strain, and the dashed line indicates the temperature.
In the graph shown on the right side of FIG. 8, the data of the graph shown on the left side of FIG. It indicates that the date is traced one day at a time from the date of failure.

The label generator 312 generates label data (correct data) indicating whether or not each of the strain data and temperature data of the 365 patterns of the input data will fail within one week. The label generation unit 312 outputs the generated label data to the storage unit 304 .
More specifically, for example, when the label generation unit 312 includes at least one strain data and temperature data included in one week from the date of failure among the strain data and temperature data of the base date of the 365 patterns, " If the strain data and temperature data included in one week from the date of the failure are not included, the label data that says "the failure will not occur within one week" is generated. In this case, the strain data and temperature data for 7 base-point days are labeled as "failure within 1 week", and the strain and temperature data for 358 base-point days are labeled as "failure within 1 week". It will be labeled as "does not occur".

Learning unit 313 accepts the pairs of input data and labels described above from storage unit 304 as training data, and performs supervised learning using the received training data to obtain strain gauge 130 of printed board 120A to be monitored. And a trained model 1531A that determines whether or not "failure will occur within one week" based on the strain data and temperature data for the past one month starting from the present measured by the temperature sensor 135 To construct.
Then, the learning unit 313 provides the built trained model 1531A to the failure prediction device 150A included in the control device 110A.

FIG. 9 is a diagram showing an example of a trained model 1531A constructed by machine learning device 300A of FIG. Here, as shown in FIG. 9, the trained model 1531A uses strain data and temperature data for the past one month with the present as the input layer, and determines whether "failure will occur within one week" or not. A multi-layer neural network whose output layer is data indicating whether or not is exemplified.
In this case, in the operation phase, the failure prediction device 150A applies strain data and temperature data for the past one month based on the combination of the strain gauge 130 and the temperature sensor 135 to the trained model 1531A. by sequentially inputting, it is possible to determine whether or not "failure will occur within one week". Then, the failure prediction device 150A, for example, for strain data and temperature data for the past one month from the present measured by at least one combination of the strain gauge 130 and the temperature sensor 135, "within one week If it is determined that a failure will occur in the printed circuit board 120A within one week, it is possible to predict a sign that a defect will occur in the solder joints 40 of the printed circuit board 120A.

Note that, when learning unit 313 acquires new teacher data after constructing trained model 1531A, learning unit 313 further performs supervised learning on trained model 1531A, thereby obtaining trained model 1531A constructed once. may be updated.
Also, the learned model 1531A may be shared among the same type of printed boards 120A. In other words, machine learning device 300A constructs learned model 1531A for each different type of printed board 120A, and provides failure prediction device 150A with learned model 1531A corresponding to the type of printed board 120A. may
The machine learning for generating the trained model 1531A included in the failure prediction device 150A has been described above.
Next, the failure prediction device 150A will be described.
<Failure prediction device 150A>
As shown in FIG. 6, the failure prediction device 150A includes an input section 151A, a determination section 152A, and a storage section 153. FIG. Here, the input unit 151A is as described in the learning phase. Also, the storage unit 153 is as described in the learning phase, except for the trained model 1531A.

Judgment unit 152A inputs strain data and temperature data for the past one month from the current time out of strain data and temperature data stored in storage unit 153 into learned model 1531A, and within one week Determine whether a failure will occur.
More specifically, the determination unit 152A sequentially obtains strain data and temperature data for the past one month based on the present based on the combination of the strain gauge 130 and the temperature sensor 135 for the learned model 1531A. By inputting it, it is determined whether or not "failure will occur within one week". Then, when determining that the combination of at least one strain gauge 130 and temperature sensor 135 "failure will occur within one week", the determining unit 152A controls the predictor of failure of the solder joints 40 on the printed board 120A. Output to an output device (not shown) of the device 110A.
The determination unit 152A may be a combination of the strain gauge 130 and the temperature sensor 135, for example, the strain gauge 130 and the temperature sensor 135 closest to the strain gauge 130.

<Failure Prediction Processing of Failure Prediction Device 150A in Operation Phase>
Next, the operation related to the failure prediction process of the failure prediction device 150A will be described.
FIG. 10 is a flowchart for explaining the failure prediction processing of the failure prediction device 150A in the operation phase.
In step S31, the input unit 151A inputs strain data and temperature data in the printed board 120A measured by the strain gauge 130 and the temperature sensor 135, respectively, and outputs the input strain data and temperature data to the storage unit 153. .

In step S32, the determination unit 152A selects the strain data and temperature data stored in the storage unit 153, based on the combination of the strain gauge 130 and the temperature sensor 135, based on the current strain data for the past one month. and temperature data are input to the learned model 1531A, and it is determined whether or not "failure will occur within one week".

In step S33, the determination unit 152A determines whether or not there is a combination of the strain gauge 130 and the temperature sensor 135 determined to "fail within one week" in the determination of step S32. If there is a combination determined that "a failure will occur within a week", the process proceeds to step S34, and if there is no combination determined that "a failure will occur within a week", the process proceeds to step S31.

In step S34, the determination unit 152A outputs a sign of failure of the solder joints 40 on the printed board 120A to an output device (not shown) of the control device 110A.

As described above, the failure predicting apparatus 150A according to the second embodiment provides strain data and temperature data (state data) measured in advance for a period of 13 months longer than one year from the date of failure in an arbitrary printed board. ), the strain data and temperature data up to one month before each date included in the range with the date of failure as the upper limit and the day one year before the date of failure as the lower limit, Extract strain data and temperature data for each base date, use the extracted strain data and temperature data for a plurality of base dates as input data, and each combination of strain data and temperature data for the plurality of base dates must be within one week. A trained model 1531A obtained by performing supervised learning using training data whose label data is data indicating whether or not a failure will occur in the monitored printed circuit board 120A. By inputting strain data and temperature data for the past one month, it is determined whether or not a failure will occur within one week.
As a result, the failure prediction device 150A can detect a sign of failure of the solder joints 40 before the failure of the solder joints 40 causes the CNC to malfunction during processing. Then, the failure prediction device 150A prevents the printed board 120A from being operated while the machine tool 100 is not in operation, before the control device 110A malfunctions during processing due to a defect in the solder joints 40 of the monitored printed board 120A. It can be replaced, and the downtime of the machine tool 100 can be reduced.
The second embodiment has been described above.

<Third Embodiment>
In the third embodiment, the state data includes strain data measured on the printed board by the sensor unit, temperature data measured on the printed board by the temperature sensor, and integration of the amount of strain after the printed board starts operating. Illustrate the data. And this point is different from the first embodiment.
A third embodiment will be described below.

FIG. 11 is a functional block diagram showing a functional configuration example of a numerical control system according to the third embodiment. Note that functional blocks having functions similar to those in FIG. 6 are assigned the same reference numerals, and descriptions thereof are omitted.
As shown in FIG. 11, the machine tool 100 incorporates a control device 110B on which a printed board 120A is mounted.
The control device 110B is included in the failure prediction device 150B in the operation phase, as will be described later.
Before describing failure prediction device 150B, machine learning for generating a trained model will be described.

<Control Device 110B in Learning Phase>
In the learning phase, the control device 110B collects teacher data necessary for performing machine learning in the same way as the control device 110 in FIG. Therefore, as described above, the control device 110B controls the temperature of the environment in which the printed circuit board 120A is mounted, which is measured by the n strain gauges 130 and the m temperature sensors 135 arranged on the mounted printed circuit board 120A. State data indicating the state is obtained from the printed board 120A. In addition, control device 110 acquires, as state data, integrated data of the amount of strain after printed board 120A starts operating. Control device 110B outputs the acquired state data to server 200 via a communication unit (not shown) of machine tool 100 .

The control device 110B in the learning phase includes an input section 151A, a storage section 153, and an integration section 154. The control device 110B in the learning phase does not have to include the determination section 152B and the trained model 1531B.
The integration unit 154 uses the strain data of each strain gauge 130 stored in the data area 1532 of the storage unit 153, for example, to calculate the integrated value of the amount of strain after the printed board 120A to be monitored starts operating. It is calculated for each strain gauge 130 . The integration unit 154 outputs the calculated integration value to the storage unit 153 as integration data (state data).

<Machine learning device 300B>
The machine learning device 300B has an input data generation unit 321, a label generation unit 322, a learning unit 323, and a storage unit 304, as shown in FIG.
The input data generator 321 has a strain data extractor 340 , a temperature data extractor 341 , and an integrated data extractor 342 .
More specifically, in the learning phase, the input data generator 321 collects strain data, temperature data, and integrated data (state data) may be obtained from the server 200 .
The integrated data extracting unit 342 extracts integrated data included in the past range 55 from the acquired integrated data up to one month ago, which is indicated by a thick-line rectangle, with the date of failure as the base date, as shown in FIG. 12 . , the date of failure is extracted as integrated data of the base date. Next, the integrated data extracting unit 342 extracts the integrated data included in the past range 55 up to one month before, with the day one day before the failure date as the base date. Extract as integrated data. Then, the integrated data extracting unit 342 extracts the integrated data included in the past range 55 up to one month ago with the day one year before the failure date, that is, the day 364 days before the failure date as the base date. This is repeated until 364 days before the failure date is extracted as the integrated data of the base date.
The input data generation unit 321 sets strain data, temperature data, and integrated data for the same base date among the extracted strain data, temperature data, and integrated data for each base date, and generates 365 patterns of strain data for the base date. , temperature data and integrated data are generated as input data. The input data generation unit 321 outputs the generated input data to the storage unit 304 .
FIG. 12 is a diagram illustrating an example of generation of input data using strain data, temperature data, and integrated data for an arbitrary printed board.
The graph on the left side of FIG. 12 shows temporal changes in the amount of strain, temperature, and integrated value measured over a period of 13 months longer than one year from the date d1 when an arbitrary printed board failed. The solid line indicates the amount of strain, the dashed line indicates the temperature, and the thick line indicates the integrated value.
The graph shown on the right side of FIG. 12 is based on the data of the graph shown on the left side of FIG. It indicates that the date is traced one day at a time from the date of failure.

The label generation unit 322 uses data indicating whether or not a failure will occur within one week for each of the strain data, temperature data, and integrated data of the 365 patterns of the input data as label data (correct data). Generate. The label generation section 322 outputs the generated label data to the storage section 304 .
More specifically, the label generation unit 322, for example, among the strain data, temperature data, and integrated data of the 365 patterns of the base date, at least the strain data, temperature data, and integrated data included in one week from the date of failure. If one is included, it is assumed that "failure will occur within one week." If strain data, temperature data, and integrated data included in one week from the date of failure are not included, it is assumed that "failure will not occur within one week." Generate label data for

The learning unit 323 accepts the pairs of the input data and the label from the storage unit 304 as training data, and performs supervised learning using the received training data to obtain the current point on the monitored printed board 120A. A trained model 1531B for judging whether or not "failure will occur within one week" is constructed based on strain data, temperature data, and integrated data for the past one month.
Then, the learning unit 323 provides the built trained model 1531B to the failure prediction device 150B included in the control device 110B.

FIG. 13 is a diagram showing an example of a learned model 1531B constructed by machine learning device 300B in FIG. Here, as shown in FIG. 13, the trained model 1531B uses strain data, temperature data, and integrated data for the past one month from the present as an input layer, and "failure will occur within one week." A multi-layer neural network whose output layer is data indicating whether or not is exemplified.
In this case, in the operation phase, the failure prediction device 150B applies strain data and temperature data for the past one month based on the combination of the strain gauge 130 and the temperature sensor 135 to the trained model 1531B. And by sequentially inputting the integrated data, it is possible to determine whether or not "a failure will occur within one week." Then, the failure prediction device 150B, for example, for strain data, temperature data, and integrated data for the past one month from the present of at least one combination of the strain gauge 130 and the temperature sensor 135, "within one week If it is determined that a failure will occur in the printed circuit board 120A within one week, it is possible to predict a sign that a defect will occur in the solder joints 40 of the printed circuit board 120A.

Note that, when learning unit 323 acquires new teacher data after constructing trained model 1531B, learning unit 323 further performs supervised learning on trained model 1531B to obtain trained model 1531B constructed once. may be updated.
Also, the learned model 1531B may be shared among the same type of printed boards 120A. In other words, machine learning device 300B constructs learned model 1531B for each different type of printed circuit board 120A, and provides failure prediction device 150B with learned model 1531B corresponding to the type of printed circuit board 120A. may
The machine learning for generating the learned model 1531B included in the failure prediction device 150B has been described above.
Next, the failure prediction device 150B will be described.

<Failure prediction device 150B>
As shown in FIG. 11, the failure prediction device 150B includes an input section 151A, a determination section 152B, a storage section 153, and an integration section 154. FIG. Here, the integrating section 154 is as described in the learning phase.
The determining unit 152B inputs strain data, temperature data, and integrated data for the past one month from the present among the strain data, temperature data, and integrated data stored in the storage unit 153 to the learned model 1531B. to determine whether or not a failure will occur within one week.
More specifically, the determining unit 152B obtains strain data, temperature data, and integrated data for the past one month based on the current time according to the combination of the strain gauge 130 and the temperature sensor 135 for the trained model 1531B. are entered in order, it is determined whether or not "failure will occur within one week." Then, when determining that "a failure will occur within one week" in at least one combination of the strain gauge 130 and the temperature sensor 135, the determination unit 152B controls the sign of failure of the solder joint 40 on the printed board 120A. Output to an output device (not shown) of the device 110B.
6, the determination unit 152B may be a combination of the strain gauge 130 and the temperature sensor 135, for example, the strain gauge 130 and the temperature sensor 135 closest to the strain gauge 130.

<Failure Prediction Processing of Failure Prediction Device 150B in Operation Phase>
Next, the operation related to the failure prediction process of the failure prediction device 150B will be described.
FIG. 14 is a flowchart for explaining failure prediction processing of the failure prediction device 150B in the operation phase. In the flowchart of FIG. 14, the same step numbers are assigned to the same steps as those shown in FIG. 10, and detailed description thereof will be omitted.
In step S31, the input unit 151A inputs strain data and temperature data in the printed board 120A measured by the strain gauge 130 and the temperature sensor 135, respectively, and outputs the input strain data and temperature data to the storage unit 153. .

In step S311, using the strain data of each strain gauge 130 stored in the storage unit 153, the integration unit 154 calculates the integrated data of the strain amount after the printed board 120A to be monitored starts operating. is calculated every time, and the calculated integration data is output to the storage unit 153 .

In step S321, the determination unit 152B selects the strain data, temperature data, and integrated data stored in the storage unit 153 for the past one month based on the combination of the strain gauge 130 and the temperature sensor 135. are input to the learned model 1531B, and it is determined whether or not "failure will occur within one week".

In step S33, the determination unit 152B determines whether or not there is a combination of the strain gauge 130 and the temperature sensor 135 for which it was determined in the determination of step S321 that "failure will occur within one week." If there is a combination determined that "a failure will occur within a week", the process proceeds to step S34, and if there is no combination determined that "a failure will occur within a week", the process proceeds to step S31.

In step S34, the determination unit 152B outputs a sign of failure of the solder joints 40 on the printed board 120A to an output device (not shown) of the control device 110B.

As described above, the failure prediction device 150B according to the third embodiment can obtain strain data, temperature data, and integrated data measured in advance for a period of 13 months longer than one year from the date of failure in an arbitrary printed board. Using (state data), strain data, temperature Data and integrated data are extracted as strain data, temperature data, and integrated data for each base date, and the extracted strain data, temperature data, and integrated data for a plurality of base dates are used as input data, and strain data for the plurality of base dates. , each combination of the temperature data and the integrated data, the trained model 1531B obtained by performing supervised learning with training data labeled with data indicating whether or not a failure will occur within one week, By inputting strain data, temperature data, and integrated data for the past one month from the current time acquired in the printed circuit board 120A to be monitored, it is determined whether or not a failure will occur within one week.
As a result, the failure prediction device 150B can detect a sign of failure of the solder joints 40 before the failure of the solder joints 40 causes the CNC to malfunction during processing. Then, the failure prediction device 150B prevents the printed board 120A from being operated while the machine tool 100 is not in operation, before the control device 110B malfunctions during processing due to a defect in the solder joints 40 of the monitored printed board 120A. It can be replaced, and the downtime of the machine tool 100 can be reduced.
The third embodiment has been described above.

<Fourth Embodiment>
In the fourth embodiment, as the state data, the strain data measured by the sensor unit on the printed board is used, and the integrated data of the amount of strain after the printed board starts operating, and the temperature indicating the temperature of the printed board Illustrate the data. And this point is different from the first embodiment.
A fourth embodiment will be described below.

FIG. 15 is a functional block diagram showing a functional configuration example of a numerical control system according to the fourth embodiment. Note that functional blocks having functions similar to those in FIG. 11 are denoted by the same reference numerals, and descriptions thereof are omitted.
As shown in FIG. 15, machine tool 100 incorporates control device 110C on which printed board 120A is mounted.
The control device 110C is included in the failure prediction device 150C in the operation phase, as will be described later.
Before describing the failure prediction device 150C, machine learning for generating a learned model will be described.

<Control Device 110C in Learning Phase>
In the learning phase, the control device 110C collects teacher data necessary for machine learning in the same way as the control device 110 in FIG. Therefore, as described above, the control device 110C controls the temperature of the environment in which the printed board 120A is mounted, which is measured by the n strain gauges 130 and the m temperature sensors 135 arranged on the mounted printed board 120A. State data indicating the state is obtained from the printed board 120A. In addition, the control device 110C acquires, as state data, integrated data of the amount of strain after the printed board 120A starts operating. The control device 110C outputs the acquired state data to the server 200 via a communication unit (not shown) of the machine tool 100C.
110 C of control apparatuses in a learning phase are provided with 151 A of input parts, the memory|storage part 153, and the integrating|accumulating part 154. FIG. The control device 110C in the learning phase does not have to include the determination unit 152B and the trained model 1531C.

<Machine learning device 300C>
300 C of machine learning apparatuses have the input data production|generation part 331, the label production|generation part 332, and the learning part 333, as shown in FIG.
The input data generator 331 has a temperature data extractor 341 and an integrated data extractor 342 .
More specifically, in the learning phase, the input data generation unit 331 acquires from the server 200 integrated data and temperature data measured over a period of 13 months longer than one year from the failure of an arbitrary printed board. You may
The input data generation unit 331 extracts the temperature data for each base date extracted by the temperature data extraction unit 341 and the integrated data for each base date extracted by the integration data extraction unit 342. Data is set as one set, and integrated data and temperature data of 365 patterns of base dates are generated as input data. The input data generation unit 331 outputs the generated input data to the storage unit 304 .

The label generator 332 generates label data (correct data) indicating whether or not a failure will occur within one week for each of the integrated data and temperature data of the 365 patterns of the input data. The label generation unit 332 outputs the generated label data to the storage unit 304 .
More specifically, for example, when the label generation unit 332 includes at least one integrated data and temperature data included in one week from the date of failure among the integrated data and temperature data of the 365 patterns of the base date, Label data indicating "Failure will not occur within one week" is generated if integrated data and temperature data included in one week from the date of failure are not included.

Learning unit 333 accepts the pairs of the above-mentioned input data and labels from storage unit 304 as training data, and performs supervised learning using the received training data to obtain the current point on printed board 120A to be monitored. A trained model 1531C for determining whether "a failure will occur within a week" is constructed based on the integrated data and temperature data for the past one month.
Then, the learning unit 333 provides the built trained model 1531C to the failure prediction device 150C included in the control device 110C.

FIG. 16 is a diagram showing an example of a learned model 1531C constructed by the machine learning device 300C of FIG. Here, as shown in FIG. 16, the trained model 1531C uses integrated data and temperature data for the past one month with the present as the input layer, and determines whether "failure will occur within one week" or not. A multi-layer neural network whose output layer is data indicating whether or not is exemplified.
In this case, in the operation phase, the failure prediction device 150C applies integrated data and temperature data for the past one month based on the combination of the strain gauge 130 and the temperature sensor 135 to the trained model 1531C. by sequentially inputting, it is possible to determine whether or not "failure will occur within one week". Then, for example, the failure prediction device 150C, for example, for the integrated data and the temperature data for the past one month from the present of at least one combination of the strain gauge 130 and the temperature sensor 135, "Failure will occur within one week. will occur", it is possible to predict a sign that a defect will occur in the solder joint 40 on the printed circuit board 120A within one week.

Note that, when learning unit 333 acquires new teacher data after constructing trained model 1531C, learning unit 333 further performs supervised learning on trained model 1531C to may be updated.
Also, the learned model 1531C may be shared among the same type of printed boards 120A. In other words, machine learning device 300C builds learned model 1531C for each different type of printed board 120A, and provides failure prediction device 150C with learned model 1531C corresponding to the type of printed board 120A. may
Machine learning for generating the learned model 1531C provided in the failure prediction device 150C has been described above.
Next, the failure prediction device 150C will be described.

<Failure prediction device 150C>
As shown in FIG. 15, the failure prediction device 150C includes an input section 151A, a determination section 152C, a storage section 153, and an integration section 154. FIG.
The determination unit 152C inputs the accumulated data and the temperature data for the past one month from the current point out of the accumulated data and the temperature data stored in the storage unit 153 into the learned model 1531C, and within one week Determine whether a failure will occur.
More specifically, the determination unit 152C sequentially obtains integrated data and temperature data for the past one month based on the current time according to the combination of the strain gauge 130 and the temperature sensor 135 for the trained model 1531C. By inputting it, it is determined whether or not "failure will occur within one week". Then, when determining that the combination of at least one strain gauge 130 and temperature sensor 135 will fail within one week, the determining unit 152C controls the predictor of failure of the solder joints 40 on the printed circuit board 120A. Output to an output device (not shown) of the device 110C.
6, the determination unit 152C may be a combination of the strain gauge 130 and the temperature sensor 135, for example, the strain gauge 130 and the temperature sensor 135 closest to the strain gauge 130.

<Failure prediction processing of the failure prediction device 150C in the operation phase>
Next, the operation related to the failure prediction process of the failure prediction device 150C will be described.
FIG. 17 is a flowchart for explaining failure prediction processing of the failure prediction device 150C in the operation phase. In the flowchart of FIG. 17, the same step numbers are assigned to the same steps as those shown in FIG. 14, and detailed description thereof will be omitted.
In step S31, the input unit 151A inputs strain data and temperature data in the printed board 120A measured by the strain gauge 130 and the temperature sensor 135, respectively, and outputs the input strain data and temperature data to the storage unit 153. .

In step S311, using the strain data of each strain gauge 130 stored in the storage unit 153, the integration unit 154 calculates the integrated data of the strain amount after the printed board 120A to be monitored starts operating. is calculated every time, and the calculated integration data is output to the storage unit 153 .

In step S322, the determination unit 152C selects integrated data for the past one month based on the combination of the strain gauge 130 and the temperature sensor 135, among the integrated data and temperature data stored in the storage unit 153. and temperature data are input to the trained model 1531C to determine whether or not "failure will occur within one week".

In step S33, the determination unit 152C determines whether or not there is a combination of the strain gauge 130 and the temperature sensor 135 determined to "fail within one week" in the determination of step S322. If there is a combination determined that "a failure will occur within a week", the process proceeds to step S34, and if there is no combination determined that "a failure will occur within a week", the process proceeds to step S31.

In step S34, the determination unit 152C outputs a sign of failure of the solder joints 40 on the printed board 120A to an output device (not shown) of the control device 110C.

As described above, the failure predicting device 150C according to the fourth embodiment has integrated data and temperature data (state data) measured in advance over a period of 13 months longer than one year from the date when an arbitrary printed board failed. ), the integrated data and temperature data up to one month before the day of failure is the base date, and the date included in the range is the day one year before the failure date. Extract as integrated data and temperature data for each base date, use the extracted integrated data and temperature data for a plurality of base dates as input data, and each combination of the plurality of integrated data and temperature data for one month is 1 A trained model 1531C obtained by performing supervised learning using training data labeled with data indicating whether or not a failure will occur within a week, and the current obtained in the monitored printed circuit board 120A is used as a base point. By inputting integrated data and temperature data for the past one month, it is determined whether or not a failure will occur within one week.
As a result, the failure prediction device 150C can detect a sign of failure of the solder joints 40 before the CNC malfunctions during processing due to the failure of the solder joints 40 . Then, the failure prediction device 150C prevents the printed board 120A from being operated while the machine tool 100 is not in operation, before the control device 110C malfunctions during processing due to a defect in the solder joints 40 of the monitored printed board 120A. It can be replaced, and the downtime of the machine tool 100 can be reduced.
The fourth embodiment has been described above.

Although the first to fourth embodiments have been described above, the failure prediction device 150 (150A, 150B, 150C) and the machine learning device 300 (300A, 300B, 300C) are limited to the above-described embodiments. It is not intended to be modified, but includes modifications and improvements within the scope that can achieve the purpose.

<Modification 1>
In the first to fourth embodiments described above, the machine learning device 300 (300A, 300B, 300C) was exemplified as a device different from the machine tool 100, but it is not limited to this. A part or all of the functions of the machine learning device 300 (300A, 300B, 300C) may be provided in the machine tool 100 or the control device 110 (110A, 110B, 110C).

<Modification 2>
In the first to fourth embodiments described above, failure prediction devices 150 (150A, 150B, 150C) were included in control devices 110 (110A, 110B, 110C), but are not limited to this. A part or all of the input unit 151 (151A), the determination unit 152 (152A, 152B, 152C), the storage unit 153 (153A, 153B, 153C), and the integration unit 154 of the failure prediction device 150, the machine tool 100, The printed board 120 (120A) may be provided.
Alternatively, the server 200 may include part or all of the input unit 151 (151A), the determination unit 152 (152A, 152B, 152C), the storage unit 153 (153A, 153B, 153C), and the integration unit 154, for example. may Also, each function of the failure prediction device may be implemented using a virtual server function or the like on the cloud.
Furthermore, the failure prediction device may be a distributed processing system in which each function of the failure prediction device is appropriately distributed to a plurality of servers.

<Modification 3>
Also, for example, in the above-described first to fourth embodiments, strain data, temperature data, and integrated data are transmitted to the server 200, but the present invention is not limited to this. The server 200 centrally manages the strain data, temperature data, and integrated data of the printed boards 120 (120A) mounted on the controllers 110 (110A, 110B, 110C) of the machine tools 100 connected via a network or the like. You may

<Modification 4>
Further, for example, in the first to fourth embodiments, the "predetermined period" is set to one year, the "first period" is set to "one year", the "second period" is set to "one month", And the "third period" is set to "one week", but it is not limited to this. The "predetermined period", the "first period", the "second period", and the "third period" may be appropriately set according to the environment of the factory where the machine tool 100 is installed.

Note that each function included in the failure prediction device 150 (150A, 150B, 150C) and the machine learning device 300 (300A, 300B, 300C) in the first to fourth embodiments is hardware, software Or each can be realized by a combination of these. Here, "implemented by software" means implemented by a computer reading and executing a program.

The program can be stored and delivered to the computer using various types of non-transitory computer readable media. Non-transitory computer-readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (eg, flexible discs, magnetic tapes, hard disk drives), magneto-optical recording media (eg, magneto-optical discs), CD-ROMs (Read Only Memory), CD- R, CD-R/W, semiconductor memory (e.g., mask ROM, PROM (Programmable ROM), EPROM (Erasable PROM), flash ROM, RAM). (transitory computer readable medium) Examples of transitory computer readable media include electrical signals, optical signals, and electromagnetic waves Transitory computer readable media include electrical wires and optical fibers, etc. The program can be delivered to the computer via a wired or wireless communication path.

It should be noted that the steps of writing a program recorded on a recording medium include not only processes that are executed chronologically in order, but also processes that are executed in parallel or individually, even if they are not necessarily processed chronologically. It also includes

In other words, the machine learning device, failure prediction device, control device, and printed circuit board of the present disclosure can take various embodiments having the following configurations.

(1) The machine learning device 300 of the present disclosure, in any printed circuit board mounted in the control device 110 that controls the industrial machine (machine tool 100), over a first period longer than a predetermined period until the date of failure Using the measured state data indicating the state of the environment in which the printed circuit board 120 is mounted, each of the days included in the range with the date of failure as the upper limit and the date before the predetermined period as the lower limit as the base date. input data for extracting, as status data for each base date, a set of status data in a range retroactively past a second period shorter than the first period, and generating status data for a plurality of extracted base dates as input data generating unit 301, when at least one of the plurality of base date status data includes at least one status data included in a third period equal to or less than the length of the second period from the date of failure, within the third period from the base date a label generation unit 302 for generating label data indicating that no failure will occur within the third period from the base date when the failure occurs and the status data included in the third period from the date of the failure is not included; A learning unit 303 that executes supervised learning using the input data generated by the generation unit 301 and the label data generated by the label generation unit 302 to generate a trained model 1531 .
According to this machine learning device 300, the occurrence of defects in the solder joints 40 of the printed board 120 is predicted before the controller 110 malfunctions during processing due to defects in the solder joints 40 of the printed board 120. A trained model 1531 can be built.

(2) The state data may include strain data measured by the sensor section (strain gauge 130) on the printed board.
By doing so, the machine learning device 300 can construct the trained model 1531 .

(3) The state data may include strain data measured by the sensor unit (strain gauge 130) on the printed board and temperature data measured by the temperature sensor 135 on the printed board.
By doing so, machine learning device 300A can construct learned model 1531A.

(4) State data includes strain data measured by the sensor unit (strain gauge 130) on the printed board, temperature data measured by the temperature sensor 135 on the printed board, and the amount of strain after the printed board started operating. may include integrated data of
By doing so, machine learning device 300B can construct learned model 1531B.

(5) The state data is the strain data measured by the sensor unit (strain gauge 130) on the printed board, and the integrated data of the amount of strain after the printed board starts operating, and the temperature sensor 135 on the printed board. It may also include measured temperature data.
By doing so, the machine learning device 300C can construct the learned model 1531C.

(6) State data indicating the state of the environment in which the printed circuit board 120 is mounted, measured over a first period longer than a predetermined period up to the date of failure, may be stored in the server 200 .
By doing so, it is possible to acquire all measured state data during the period from when the printed board 120 starts operating until it fails.

(7) The failure prediction device 150 (150A, 150B, 150C) of the present disclosure includes a trained model 1531 (1531A, 1531B, 1531C) acquired by learning by the machine learning device 300 (300A, 300B, 300C), A storage unit 153 for storing state data, and among the state data stored in the storage unit 153, state data in a past range from the present to the second period before the learned model 1531 (1531A, 1531B). , 1531C) and determines whether or not a failure will occur in the printed circuit board 120 within the third period.
According to this failure prediction apparatus, it is possible to predict the occurrence of defects in the solder joints 40 of the printed board 120 before the controller 110 malfunctions during processing due to defects in the solder joints 40 of the printed board 120 . can.

(8) The failure prediction device 150 (150A, 150B, 150C) may include the machine learning device 300 (300A, 300B, 300C).
By doing so, the same effects as those of (1) to (6) described above can be obtained.

(9) The control devices 110 (110A, 110B, 110C) of the present disclosure may include failure prediction devices 150 (150A, 150B, 150C).
According to this control device 110 (110A, 110B, 110C), the same effect as the above (7) or (8) can be obtained.

(10) The printed boards 120, 120A of the present disclosure may include failure prediction devices 150 (150A, 150B, 150C).
According to the printed boards 120 and 120A, the same effect as the above (7) or (8) can be obtained.

100 machine tool 110 control device 150 failure prediction device 200 server 300 machine learning device 120 printed board 151 input unit 152 determination unit 130 strain gauge 1531 learned model

Claims (10)

In any printed circuit board mounted on a control device that controls an industrial machine,
Using state data indicating the state of the environment in which the printed circuit board is mounted, which is measured over a first period longer than a predetermined period until the date of failure,
A state in which a second period, which is shorter than the first period, is traced back to the past, with each day included in the range having the date of failure as the upper limit and the day before the predetermined period from the date of failure as the lower limit as the base date. an input data generation unit that extracts a set of data as state data for each base date and generates the extracted plurality of base date state data as input data;
If at least one piece of status data included in a third period of time from the date of failure to the length of the second period or less is included among the plurality of status data of the base date, the failure occurs within the third period from the base date. a label generation unit for generating label data indicating that no failure will occur within a third period from the base date if the status data included in the third period from the date of the failure is not included;
a learning unit that performs supervised learning using the input data generated by the input data generation unit and the label data generated by the label generation unit to generate a trained model;
A machine learning device with The state data includes strain data measured by a sensor unit on the printed board,
The machine learning device according to claim 1. 2. The machine learning device according to claim 1, wherein said state data includes strain data measured on said printed board by a sensor unit and temperature data measured on said printed board by a temperature sensor. The state data includes strain data measured on the printed board by a sensor unit, temperature data measured on the printed board by a temperature sensor, and integrated data of the amount of strain after the printed board starts operating. , The machine learning device according to claim 1. The state data is integrated data of the amount of strain after the printed board starts operating using the strain data measured by the sensor unit on the printed board, and temperature data measured by the temperature sensor on the printed board. 2. The machine learning device of claim 1, comprising: state data indicating the state of the environment in which the printed circuit board is mounted, which is measured over a first period longer than a predetermined period up to the date of failure, is stored in a server;
The machine learning device according to any one of claims 1 to 5. A trained model acquired by learning by the machine learning device according to any one of claims 1 to 6;
a storage unit that stores the state data;
Among the state data stored in the storage unit, the state data in the past range up to the second period before the present is input to the learned model, and the third period is performed on the printed board. a determination unit that determines whether a failure will occur within the period;
A failure prediction device. The failure prediction device according to claim 7, comprising said machine learning device. A control device comprising the failure prediction device according to claim 7 or 8. A printed board comprising the failure prediction device according to claim 7 or 8.

Images