JP7432161B2 - sensor system - Google Patents

Description

The present invention relates to a sensor system.

BACKGROUND ART Conventionally, in order to improve the accuracy of devices such as sensors, noise has been predicted using a learning model generated using machine learning techniques, and noise has been removed from data output by the device.

For example, Patent Document 1 discloses an expert diagnosis method in which a learning model is automatically replaced and a new learning training is performed to form a new judgment system. The expert diagnosis method includes the step of automatically replacing the learning model to update the expert system.

Further, Patent Document 2 discloses a power consumption estimation device that includes a learning/classifier 12 serving as a power consumption estimation device, and a data exchange unit 15 that exchanges a learning model with another learning model. The power consumption estimating device determines which learning pattern the network traffic pattern is closest to, and replaces it with an appropriate learning model.

Special Publication No. 2010-519691 JP2013-97713A

However, learning models are not always improved through learning. For example, learning may cause overlearning, which deteriorates performance. If a learning model with degraded performance is applied to a predetermined device such as a sensor, the entire system using the predetermined device will be adversely affected. Therefore, it may be necessary to evaluate the learning model.

The present invention has been made in view of these circumstances, and an object thereof is to provide a sensor system that can suppress deterioration in the performance of a learning model.

A sensor system according to one aspect of the present invention includes an acquisition unit that acquires sensing data output by a sensor, and a determination unit that performs a predetermined determination using an output value obtained by inputting the sensing data to a first learning model. a learning unit that generates a second learning model by machine learning using sensing data; a learning model determining unit that determines which of the performance of the first learning model and the performance of the second learning model is superior; and a substitute unit that substitutes the first learning model with the second learning model when the learning model determining unit determines that the performance of the second learning model is superior to the performance of the first learning model.

According to this aspect, the first learning model makes a determination based on the sensing data, and the second learning model is further generated using the sensing data. Then, only when it is determined that the performance of the second learning model is superior to the performance of the first learning model, the first learning model is replaced by the second learning model. Therefore, deterioration of the learning model due to overfitting or the like is suppressed.

According to the present invention, it is possible to provide a sensor system that can suppress deterioration in the performance of a learning model.

FIG. 1 is a diagram showing an overview of a sensor system 1 according to a first embodiment. FIG. 2 is a diagram showing functional blocks of a master unit 10 according to the first embodiment. FIG. 3 is a diagram for explaining learning model generation processing by the sensor system 1 according to the first embodiment. FIG. 3 is a diagram for explaining performance determination processing of a learning model by the sensor system 1 according to the first embodiment. FIG. 3 is a diagram for explaining performance determination processing of a learning model by the sensor system 1. FIG. It is a schematic diagram for explaining the composition of the biting detection part 91 concerning a 2nd embodiment. It is a figure showing the functional block of master unit 10 concerning a 2nd embodiment. FIG. 7 is an operation flow diagram for explaining learning model generation processing and performance determination processing by the sensor system 1 according to the second embodiment. FIG. 7 is an operation flow diagram for explaining the learning process in S207. FIG. 7 is an operation flow diagram for explaining learning model determination processing in S303. It is a figure showing an outline of sensor system 1 concerning a 3rd embodiment. It is a figure showing the functional block of master unit 10 concerning a 3rd embodiment. FIG. 7 is an operational flow diagram of learning model generation processing and the like by the sensor system 1 according to the third embodiment. This is an example of a learning progress display screen. This is an example of a display screen of a learning completion notification. It is a figure which shows another example of a screen of a learning progress display. FIG. 1 is a diagram showing an example of a system with a connected configuration using a plurality of sensors. FIG. 2 is a diagram for explaining a learning section in a system with a connected configuration using a plurality of sensors.

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment (hereinafter also referred to as "this embodiment") according to one aspect of the present invention will be described below based on the drawings. In addition, in each figure, those with the same reference numerals have the same or similar configurations.

[First embodiment]
<Sensor system 1>
FIG. 1 is a diagram showing an overview of a sensor system 1 according to a first embodiment. The sensor system 1 includes a master unit 10, a first slave unit 20a, and a second slave unit 20b.
, third slave unit 20c, first sensor 30a, second sensor 30b, third sensor 30
c and a PLC 40. First sensor 30a, second sensor 30b and third sensor 30c
is arranged along the line L, and measures data indicating the passage status of the workpiece conveyed on the line L. The first sensor 30a, the second sensor 30b, and the third sensor 30c are each
For example, it may be constituted by a fiber sensor. Alternatively, these sensors 30 may be transmissive or retroreflective photoelectric sensors. Alternatively, these sensors 30 may be displacement sensors that project a laser beam onto the object and obtain a signal value corresponding to the distance to the object based on the principle of triangulation. Alternatively, these sensors 30 may be distance measuring sensors that obtain signal values corresponding to the distance to the object based on the round trip time of light reflected by the object.

First slave unit 20a, second slave unit 20b, third slave unit 2
0c is connected to each of the plurality of sensors 30 and acquires data measured by the plurality of sensors 30. More specifically, the first slave unit 20a is connected to the first sensor 30a, the second slave unit 20b is connected to the second sensor 30b, and the third slave unit 20c is connected to the third sensor 30c.

PLC40 is a control device. The master unit 10 is connected to a plurality of slave units 20a and a control device. In this specification, the first slave unit 20a
, the second slave unit 20b, and the third slave unit 20c are collectively referred to as the slave unit 20, and the first sensor 30a, the second sensor 30b, and the third sensor 30c are collectively referred to as the sensor 30. Note that the configuration of the sensor system 1 according to this embodiment is an example, and the number of multiple sensors 30 and the number of multiple slave units 20a included in the sensor system 1 are arbitrary. Also,
The control device does not necessarily have to be the PLC 40.

The master unit 10 may be connected to the PLC 40 via a communication network such as a LAN (Local Area Network). Slave unit 20 is physically and electrically connected to master unit 10. In this embodiment, the master unit 10 stores information received from the slave unit 20 in the storage section, and transmits the stored information to the PLC 40. Therefore, the data acquired by the slave unit 20 is unified by the master unit 10 and transmitted to the PLC 40.

As an example, a determination signal and detection information are transmitted from the slave unit 20 to the master unit 10. The determination signal is a signal indicating a determination result regarding the workpiece determined by the slave unit 20 based on data measured by the sensor 30. For example, when the sensor 30 is a photoelectric sensor, the determination signal may be an on signal or an off signal obtained by comparing the amount of light received by the sensor 30 with a threshold value by the slave unit 20. The detection information is a detection value obtained by the detection operation of the slave unit 20. For example, when the sensor 30 is a photoelectric sensor, the detection operation may be an operation of emitting and receiving light, and the detection information may be the amount of light received.

The slave unit 20 may be attached to the side of the master unit 10. Parallel communication or serial communication may be used for communication between master unit 10 and slave unit 20. That is, the master unit 10 and the slave unit 20 may be physically connected through a serial transmission path and a parallel transmission path. For example, the determination signal may be transmitted from the slave unit 20 to the master unit 10 on a parallel transmission path, and the detection information may be transmitted from the slave unit 20 to the master unit 10 on a serial transmission path. Note that the master unit 10 and the slave unit 20 may be connected through either a serial transmission path or a parallel transmission path.

<Master unit 10>
FIG. 2 is a diagram showing functional blocks of the master unit 10 according to the first embodiment. The master unit 10 is an example of a control device, and includes a learning progress display section 11 and a learning model determination section 1.
2, a communication section 14, a determination section 15, a determination display section 16, an output section 17, and an input section 18. These may be realized by one processor included in the master unit 10, or may be realized by a plurality of processors.

The learning progress display unit 11 is a means for displaying the degree of learning progress of the learning model 2 and that the learning model 1 has been replaced, and is configured with, for example, an LED or a liquid crystal.

The learning model determination unit 12 compares the performance of the learning model 1 and the performance of the learning model 2, and determines which one has better performance. Any method may be used to determine the performance of the learning model. For example, the determination method may define that the smaller the variance value of the difference between sensing data and predicted data in a predetermined period is, the better the learning model is. Alternatively, the determination method may define that the learning model is better as the absolute values of the maximum and minimum values of the difference between sensing data and predicted data in a predetermined period are smaller. Alternatively, the determination method is
It may be defined that the smaller the average value of the difference between sensing data and predicted data in a predetermined period is, the better the learning model is.

Further, the learning model determination unit 12 may transmit the determination result of the performance of the learning model to a higher-level unit such as the PLC 40 via the learning progress display unit 11 or the communication unit 14. Further, the learning model determination unit 12 may determine the performance of the learning model at regular intervals, for example.

The learning unit 13 performs machine learning using sensing data as learning data between sensing, generates a learning model 2, and stores the learning model 2 in the storage unit 20. The communication unit 14 is an interface that communicates with the PLC 40. The communication unit 14 may communicate with a higher-level unit other than the PLC 40. The communication unit 14 may be a communication module capable of wired communication such as a field bus, or a communication module capable of wireless communication.

The determination unit 15 uses the learning model 1 to determine the workpiece based on the sensing data. The determination display section 16 displays the determination result. The mode of display is not particularly limited. The output unit 17 outputs the determination result to an external device such as the PLC 40. The output value may be a binary value of On/Off or an encoded multi-bit signal. The input unit 18 includes buttons and the like that accept user operations. The input unit 18 may accept input for rolling back learning data, for example.

The sensing section 19 includes a slave unit 20a, a processing section 19a, and an acquisition section 19b. The processing section 19a and the acquisition section 19b may be included in the master unit 10, or may not be included in the master unit 10, for example. Processing section 19a
processes the data received from the slave unit 20a. The mode of processing is not particularly limited. The acquisition unit 19b acquires the data processed by the processing unit 19a, and stores it in the storage unit 20 as sensing data.

The storage unit 20 is composed of one or more memories. The storage unit may have one or more access buses. The storage unit 20 stores, for example, sensing data, a data processing method, a learning model 1, a learning model 2, an old learning model, a learning determination method, and a model 1 determination result.

The sensing data may be data measured by the sensor 30 that is not processed, or may be data processed by the processing section 19a. The data processing method is a method that defines processing of processing sensing data into learning data. The processing method is not particularly limited, and may include a case where "no processing" is performed (data measured by the sensor 30 is used as is). Learning model 1 (first learning model) is a learning model used for determining the workpiece. Learning model 2 (second learning model) is a learning model generated by the learning unit 13. The old learning model is a learning model related to learning model 1 before learning model 1 was replaced by learning model 2. The learning determination method is data that defines a determination method for determining the performance of a learning model. The model 1 determination result is data indicating the determination result regarding the performance of the learning model 1.

Note that the master unit 10 may have a different pair of learning models for each slave unit. In this case, for example, learning model 1 and learning model 2 correspond to the first slave unit 20a, and learning model 3 (third learning model) and learning model 4 (fourth learning model) correspond to the second slave unit 20b. handle.

<Learning model generation process>
FIG. 3 is a diagram for explaining the learning model generation process by the sensor system 1 according to the first embodiment.

First, the learning unit 13 reads the data processing method stored in the storage unit 20. The data processing method is not particularly limited, but may include, for example, averaging processing, FFT (fast Fourier transform) processing, smoothing processing, and the like. Next, the learning unit 13 obtains the sensing data stored in the storage unit 20 and generates learning data by processing the sensing data using the data processing method described above. Next, the learning unit 13 performs machine learning using the generated learning data to generate a learning model. Next, the learning unit 13 stores the generated learning model as the learning model 2 in the storage unit 20. This completes the learning model generation process.

<Learning model performance judgment process>
4A and 4B are diagrams for explaining performance determination processing of a learning model by the sensor system 1 according to the first embodiment.

First, the learning model determination unit 12 acquires sensing data from the storage unit 20, and post-processes the sensing data using the post-processing unit 1, post-processing unit 2, etc. based on the learning determination method. The content of the post-processing is not particularly limited, but may include, for example, averaging processing, FFT processing, subtraction/addition processing, and the like. Next, the learning model determining unit 12 performs prediction processing based on the learning model 2. Next, the learning model determination unit 12 calculates a determination result using a determination result calculation unit. The content of the method for calculating the determination result is not particularly limited, but for example, as described above, the determination result may be calculated using the variance value, maximum value/minimum value, average value, etc. Next, the learning model determination unit 1
2, the determination result comparison/reflection unit compares the determination results. At this time, if a determination result is calculated that the performance of the learning model 2 is superior to the performance of the learning model 1, the learning model 2 is substituted for the learning model 1. With this, the learning model performance determination process is completed.

[Second embodiment]
The second embodiment will be described below. The second embodiment is an example in which the present invention is applied to a biting detection section 91 of a packaging device. In the following, portions of the second embodiment that are different from the first embodiment will be described, and descriptions of the same configurations as the first embodiment will be omitted as appropriate.

<Bite detection section 91>
FIG. 5 is a schematic diagram for explaining the configuration of the biting detection section 91 according to the second embodiment. The seal parts 61 and 62 are provided with two sets of welding and cutting parts, A1 and B1 and A2 and B2, and a sampling period is set for each set of these welding and cutting parts according to the timing at which the welding and cutting process is performed. Ru. When welding the film material 101, the heater parts 81a, 81b, 82a,
If a foreign object gets caught between the seal parts 82b and the cutter parts 83, 84, the rotation axes of the seal parts 61, 62 are displaced in a direction perpendicular to the axes. Since the metal plate 90 is displaced in accordance with the deviation of the rotation axis, by measuring the displacement with the proximity sensor 91, the biting can be detected.

FIG. 6 is a diagram showing functional blocks of the master unit 10 according to the second embodiment. Second
The master unit 10 according to the embodiment acquires sensing data from the biting detection section 91 instead of the plurality of slave units 20a.

<Learning model generation and learning model performance judgment processing>
FIG. 7 is an operation flow diagram for explaining learning model generation processing and performance determination processing by the sensor system 1 according to the second embodiment.

First, the data processing method is uploaded to the master unit 10 by the upper unit (
S201). The data processing method is not particularly limited, but for example, "data is not processed, t-
n, t-n+1, ..., t (t: measurement time, n: integer) n-dimensional data is input to the learning model" may be uploaded as a data processing method. Next, the learning determination method is uploaded to the master unit 10 by the upper unit (S202). The learning determination method is not particularly limited, but for example, “variance value Σ|f(t)−p|2(f
(t): predicted value by the learning model at time t; p: measured value) is smaller, the performance of the learning model is better.'' may be uploaded as a learning determination method.

Next, the upper unit sets the model 1 judgment result to the worst value "0%" (S
203). Next, the operation of the biting detection section 91 is started (S204). Next, the upper unit starts learning the master unit 10 (S205). Next, set the end flag to 0,
Each time is set to 1 (S206). Next, a learning process is executed (S206)
. The execution of the learning process (S206) is repeated until the end flag becomes 1 (S208).

FIG. 8 is an operation flow diagram for explaining the learning process in S207. First, it is determined whether a period defined as a predetermined learning interval has elapsed (S301). in particular,
Determine whether the remainder obtained by dividing the time by the learning interval is 0, and if the determination result is Yes.
Repeat this until. If the determination result in S301 is Yes, the learning model 2 is generated/substituted according to the data processing method stored in the storage unit 20. Next, learning model determination processing is performed (S303).

FIG. 9 is an operation flow diagram for explaining the learning model determination process in S303. first,
It is determined whether a period defined as a predetermined determination interval has elapsed (S401). Specifically, it is determined whether the remainder obtained by dividing the time by the determination interval is 0 or not, and this is repeated until the determination result becomes Yes. If the determination result in S401 is Yes, the performance of the learning model 2 is evaluated according to the learning determination method stored in the storage unit 20. Next, it is determined whether the performance of learning model 2 is superior to that of learning model 1 (S403).
If the determination result is No, the process ends. On the other hand, if the determination result is Yes, the learning model 2 is copied to the learning model 1 stored in the storage unit 20 (S404).
Next, the result of learning model 2 is stored in the model 1 determination result (S405). This completes the learning model generation process and performance determination process.

[Third embodiment]
<Sensor system 1>
FIG. 10 is a diagram showing an overview of the sensor system 1 according to the third embodiment. Below, portions of the third embodiment that are different from the first embodiment will be described, and descriptions of portions common to both embodiments will be omitted as appropriate. The sensor system 1 includes a master unit 10, a first slave unit 20a, a second slave unit 20b, a third slave unit 20c, a first sensor 30a, a second sensor 30b, a third sensor 30c, a PLC 40, and an external storage medium 50. . The external storage medium 50 may be, for example, a PC, a USB memory, or the like. The sensor system 1 can use previous learning models and sample data stored in the external storage medium 50. Furthermore, the sensor system 1 can copy and use data acquired by another system. The master unit 10 includes a display section 60. Display section 60
For example, displays the current learning progress and learning completion determination. This allows the user to check the current learning progress and learning completion determination.

<Master unit 10>
FIG. 11 is a diagram showing functional blocks of the master unit 10 according to the third embodiment. The master unit 10 is an example of a control device, and includes a communication section 14, a determination section 15, and an acquisition section 19b.
, a correction section 19c, a timer 21, and a display section 60. These may be realized by one processor included in the master unit 10, or may be realized by a plurality of processors. Further, the storage unit 20 included in the master unit 10 stores, for example, sensing data, measurement timing, and a learned model.

FIG. 12 is an operational flow diagram of learning model generation processing and the like by the sensor system 1 according to the third embodiment. First, the master unit 10 executes a predetermined setting process related to the start of learning (S501). Next, the master unit 10 receives input of sample data (
S502). Next, the master unit 10 displays a learning progress display screen via the display unit 60 (S503). Next, when the learning is completed, the master unit 10 displays a learning completion notification screen via the display unit 60 (S504).

FIG. 13 is an example of a learning progress display screen. The learning progress display screen is a display screen that shows the ratio of the current model determination accuracy to the target accuracy (target accuracy). The user can grasp the learning progress status through the learning progress display screen. The sensor system 1 displays the determination accuracy of the model and the learning status (progress) by inputting sample data. For the judgment accuracy of the model, the correct answer rate for test data for judgment rate analysis may be used. Specifically, the cross-validation results (hold-out
(leave-one-out method, leave-one-out method, etc.) or the correct answer rate for specific data specified by the user. Further, the correct answer rate (target accuracy) of the target learning model may be set arbitrarily by the user. Alternatively, the number of samples required for prediction may be calculated by determining a regression curve from the past number of learning samples and changes in model determination accuracy.

FIG. 14 is an example of a learning completion notification display screen. The learning completion notification display screen is a display screen that indicates that the accuracy of model determination has reached the target accuracy. After learning is completed, it may be possible to select whether to update or cancel the learning model. Specifically, as shown in FIG. 14, a selection section for selecting between updating the learning model and canceling is displayed on the learning completion notification display screen. When learning model update is selected, the old learning model is updated with the model for which learning has been completed. If cancel is selected, the update of the learning model is canceled.
Further, the sensor system 1 may compare the determination accuracy of past learning models with the determination accuracy of the current learning model, and automatically update when the performance improves. Further, in the sensor system 1, the learning model may be backed up every time the model is updated, and the backed up learning model may be restored at an arbitrary timing.

FIG. 15 is a diagram showing another example of the learning progress display screen. As shown in FIG. 15, the learning progress display screen may include numerical values of target discrimination accuracy, model discrimination accuracy, learning progress rate, and a display bar showing the learning progress rate.

<Learning status display on connected amplifier>
FIG. 16 is a diagram showing an example of a system with a connected configuration using a plurality of sensors. In this system, the output signals of some slave units are used as trigger input sensors, and the output signals of other slave units are used as data acquisition sensors. Specifically, the system includes a plurality of slave units 20. Each slave unit 20 includes a sensor 1 for acquiring data for learning, a trigger sensor 1 for specifying a start, a trigger sensor 2 for specifying an end, a sensor 2 for acquiring data for learning, and a trigger sensor for specifying a start. 3 and a trigger sensor 4 that designates termination. Each slave unit 20 has an indicator light. The indicator light can indicate states such as learning in progress, trigger, learning complete, and AI disabled, depending on the display mode shown in FIG. 16, for example. With this configuration, by displaying the sensor status individually using indicator light colors, etc., it is possible to understand at a glance the role (status) of each sensor and whether learning has been completed. In addition to displaying the sensor status on the slave unit using the color of the indicator light, it may also be displayed in a form that allows each status to be identified using a blinking pattern of the indicator light, a 7-segment display, or a dot matrix display. Note that an indicator light may be provided in the master unit instead of or together with the slave unit.

FIG. 17 is a diagram for explaining a learning section in a system with a connected configuration using a plurality of sensors. As shown in FIG. 17, the data acquisition sensor 1 starts learning when the signal from the trigger sensor 1 becomes High, and ends the learning when the signal from the trigger sensor 2 becomes High. Similarly, the data acquisition sensor 2 detects that the signal of the trigger sensor 3 is High.
Learning starts when the signal becomes High, and ends when the signal from the trigger sensor 4 becomes High. In this way, when the slave unit is used as a trigger sensor, the learning section of the signal can be specified by specifying the data acquisition sensor associated with the signal.

As explained above, the above-described embodiment has particularly advantageous effects in the following two cases.
(1) Initial Learning At the time of initial learning, the learning model is updated at any time using sensing data as learning data, so in the conventional technology, the performance of the learning model may deteriorate due to overfitting. However, in the embodiment described above, only the learning model 2 is updated at any time, and the learning model 1 is updated only when the performance of the learning model 2 is determined to be "better" than the performance of the learning model 1. . Therefore, the learning model 1 of the master unit 10 always gives the best results.

(2) Real-time learning (semi-unsupervised learning)
Since there are two learning models, the learning model can be updated at any time while operating as a sensor (
real-time learning) becomes possible. As the cutter portion of the packaging device wears out and its condition changes over time, the measurement waveform that can be obtained by the proximity sensor also changes. However, this change is not sudden but gradual. At this time, when the variance value of the data predicted by learning model 1 and the measured value over a certain period of time is within the judgment result of model 1, that is, the data when it is judged that there is no bite is used as learning data and the learning model is updated. let As a result, learning model 2 is updated to a learning model that can respond to environmental changes such as the cutter part even while making measurements. Furthermore, since learning model 1 is updated by determining whether the learning model is good or bad using a learning determination method, it will not become worse. Therefore, by having two learning models, it is possible to evolve into a learning model that is compatible with environmental changes while making measurements.

1 Sensor system 10 Master unit 11 Learning progress display section 12 Learning model judgment section 13 Learning section 14 Communication section 15 Judgment section 16 Judgment display section 17 Output section 18 Input section 19 Sensing section 19a Processing section 19b Acquisition section 20 Storage section 20a First Slave unit 20b Second slave unit 20c Third slave unit 30 Sensor 30a First sensor 30b Second sensor 30c Third sensor 40 PLC
91 Biting detection part

Claims (9)

First sensing data outputted by a first sensor among a plurality of sensors that sequentially sense a target object, second sensing data outputted by a second sensor among the plurality of sensors, and among the plurality of sensors. an acquisition unit that acquires third sensing data output by the third sensor;
a determination unit that performs a predetermined determination using an output value obtained by inputting arbitrary sensing data output by the plurality of sensors into a first learning model;
a learning unit that generates a second learning model by machine learning using the first sensing data acquired after the second sensing data is acquired and before the third sensing data is acquired;
a learning model determination unit that determines which of the performance of the first learning model and the performance of the second learning model is superior;
a substitution unit that substitutes the first learning model with the second learning model when the learning model determining unit determines that the performance of the second learning model is superior to the performance of the first learning model;
A control device comprising: The learning model determination unit calculates a first difference between the arbitrary sensing data and an output value obtained by inputting the arbitrary sensing data into the first learning model, and the arbitrary sensing data, Which of the performance of the first learning model and the performance of the second learning model is superior based on the second difference between the arbitrary sensing data and the output value obtained by inputting the arbitrary sensing data to the second learning model. The control device according to claim 1, which determines whether the control device The learning model determination unit determines the first learning model using at least one of a variance value, a maximum value, a minimum value, and an average value of each of the first difference and the second difference in a predetermined period. and determining which of the performance of the second learning model is superior,
The control device according to claim 2. The control device according to any one of claims 1 to 3, wherein the plurality of sensors are proximity sensors or bite detection units. The control device according to any one of claims 1 to 4, further comprising a processing section that processes the arbitrary sensing data output by the plurality of sensors. The control device according to any one of claims 1 to 5, further comprising a learning progress display unit that displays the progress of determination by the learning model determining unit. The control device according to any one of claims 1 to 6, further comprising a display unit that displays a result of determination by the learning model determination unit. The control device according to claim 7, wherein the display unit displays the result of the determination in a manner according to a first evaluation method and/or in a manner according to a second evaluation method. further comprising an operation unit for generating alternative instructions for the learning model;
The alternative unit is a case where the learning model determination unit determines that the performance of the second learning model is superior to the performance of the first learning model, and the instruction is generated by the operation unit. The control device according to any one of claims 1 to 8, wherein the first learning model is replaced by the second learning model.