JP2021146510A - Learning method, control method and printer - Google Patents

Abstract

To provide a technique capable of estimating an amount of deviation of a discharge position of ink by machine learning while reducing the number of measurement items to be input.SOLUTION: A learning method includes: generating a first learned model M1 with input of measured values D2 of a plurality of measurement items obtained from a plurality of sensors, and teacher data of an actual value D1 of a deviation amount; calculating an influence degree in the first learned model M1 with respect to each of the plurality of measurement items; selecting a part of measurement items among the plurality of measurement items based on the calculated influence degree; and generating a second learned model M2 with input of a measured value D2 of the selected measurement item, and teacher data of an actual value D1 of a deviation amount. With this, the second learned model M2 with high estimation accuracy is generated while reducing the number of the measurement items to be input.SELECTED DRAWING: Figure 4

Description

The present invention relates to machine learning for estimating the amount of deviation of the ink ejection position from the measured values of a plurality of sensors provided in the printing apparatus.

Conventionally, there is known an inkjet printing device that prints an image on a base material by ejecting ink from a plurality of heads while transporting a long strip-shaped base material in the longitudinal direction. The inkjet printing apparatus ejects inks of different colors from a plurality of heads. Then, a multicolor image is printed on the surface of the base material by superimposing the monochromatic images formed by the inks of each color. A conventional printing apparatus is described in, for example, Patent Document 1.

JP-A-2018-16412

In this type of printing apparatus, a slight misalignment (so-called “misplacement”) may occur between the plurality of monochromatic images described above. Misregistration occurs due to various factors such as rotation error of the roller that conveys the base material and expansion and contraction of the base material. In the conventional printing apparatus, the amount of misregistration is measured based on the photographed image of the base material after printing. Then, the ink ejection timing from the head is corrected based on the measured misregistration amount.

However, the transport state of the base material changes from moment to moment. Therefore, even if the amount of misregistration is measured based on the captured image of the base material after printing, the ink ejection timing for the base material during printing may not be appropriately corrected.

In order to grasp the amount of misregistration without depending on the photographed image of the base material after printing, for example, it is conceivable to use machine learning. Specifically, it is conceivable to input the measured values of various sensors mounted on the printing apparatus and generate a trained model for estimating the amount of misregistration from those measured values. However, the printing apparatus is equipped with a large number of sensors. Therefore, if all the measured values of these sensors are input to machine learning, the amount of calculation of the learning process and the estimation process using the trained model increases.

The present invention has been made in view of such circumstances, and an object of the present invention is to provide a technique capable of estimating the amount of deviation of the ink ejection position by machine learning while reducing the number of input measurement items. ..

In order to solve the above problems, the first invention of the present application is a printing apparatus that ejects ink from a plurality of heads to the surface of the base material while transporting the long strip-shaped base material in the longitudinal direction along a predetermined transport path. In a learning method for estimating the amount of mutual deviation of the ink ejection positions by the plurality of heads from the measured values of the plurality of sensors provided in the printing apparatus, a) obtained from the plurality of sensors. A first trained model that inputs the measured values of a plurality of measurement items to be input, uses the measured value of the deviation amount as teacher data, and outputs an estimated value of the deviation amount corresponding to the measured value by a supervised learning algorithm. A step of generating, b) a step of calculating the degree of influence in the first trained model for each of the plurality of measurement items, and c) a part of the plurality of measurement items based on the degree of influence. The step of selecting the measurement item of the above and d) the measurement value of the measurement item selected in the step c) are input, the measured value of the deviation amount is used as the teacher data, and the measured value is converted to the measurement value by the supervised learning algorithm. It includes a step of generating a second trained model that outputs a corresponding estimated value of the deviation amount.

The second invention of the present application is the learning method of the first invention, and in the step c), some of the measurement items having a high degree of influence are selected from the plurality of measurement items.

The third invention of the present application is the learning method of the first invention or the second invention, and in the step c), the computer that executes the steps a), b), and d) is described according to the degree of influence. Select some measurement items from multiple measurement items.

The fourth invention of the present application is a learning method of any one invention from the first invention to the third invention, and the measured value of the deviation amount has periodicity, and in the steps a) and d) , The actually measured value acquired in the time of one cycle or more of the fluctuation of the deviation amount is used as the teacher data.

The fifth invention of the present application is a control method for controlling the printing apparatus by using the second learned model generated by the learning method of any one of the first to fourth inventions, e. ) The step of inputting the measured values of the measurement items selected in the step c) into the second trained model and obtaining the estimated value of the deviation amount output from the second trained model, and f) the step. The step e) includes a step of correcting the ejection position of the ink with respect to the base material based on the estimated value of the deviation amount obtained in the step e).

The sixth invention of the present application is the control method of the fifth invention, in which g) the measured value of the measurement item selected in the step c) is input, and the measured value of the deviation amount is used as supervised data. The learning algorithm further includes a step of updating the second trained model.

The seventh invention of the present application is the control method of the sixth invention, and in the step g), when the type of the base material is changed in the printing apparatus, or the same base material is printed. It is executed in the middle of doing.

The eighth invention of the present invention includes a transport mechanism for transporting a long strip-shaped base material in the longitudinal direction along a predetermined transport path, a plurality of heads for ejecting ink to the surface of the base material, and the first to fourth inventions. Using the learning unit that generates the second trained model by the learning method of any one invention up to the present invention and the second learned model generated by the learning unit, the transport mechanism and the plurality of heads are used. A control unit that controls while correcting at least one of the above is provided.

According to the first to eighth inventions of the present application, the first trained model is once generated by learning by inputting the measured values of a plurality of measurement items, and then the degree of influence on the first trained model is determined. A second trained model is generated by performing re-learning focusing on some measurement items selected in consideration. As a result, it is possible to generate a second trained model with high estimation accuracy while reducing the number of input measurement items.

In particular, according to the third invention of the present application, the computer automatically selects the measured value according to the degree of influence. As a result, the workload of the user can be reduced.

In particular, according to the fourth invention of the present application, the relationship between the measured value and the amount of deviation having periodicity can be sufficiently learned.

In particular, according to the fifth invention of the present application, it is possible to suppress the deviation of the ink ejection position by the plurality of heads and obtain a high-quality print result. Further, since the number of measurement items to be input to the second learned data is small, the calculation load of the control unit can be reduced.

In particular, according to the sixth invention of the present application, the misregistration amount can be estimated more accurately according to the change in the usage state of the printing apparatus. Therefore, the ink ejection position with respect to the base material can be corrected more appropriately. Further, since the number of measurement items to be input in the update process is small, the calculation load of the update process can be reduced.

In particular, according to the seventh invention of the present application, the second trained model can be updated when the type of the base material is changed or while the printing process is performed on the same base material.

In particular, according to the eighth invention of the present application, it is possible to suppress the deviation of the ink ejection position by the plurality of heads and obtain a high-quality print result. Further, since the number of measurement items to be input to the second learned data is small, the calculation load of the control unit can be reduced.

It is a figure which showed the structure of the printing apparatus. It is a partial top view of the printing apparatus in the vicinity of a printing part. It is a block diagram which showed the connection between each part of a printing apparatus and a computer. It is a block diagram which conceptually showed the function of a computer. It is a flowchart which showed the flow of a learning process and a control process.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

<1. Printing device configuration>
FIG. 1 is a diagram showing a configuration of a printing apparatus 1 according to an embodiment of the present invention. The printing device 1 prints an image on the surface of the base material 9 by ejecting ink droplets from a plurality of heads 21 to 24 toward the base material 9 while transporting the long strip-shaped base material 9. It is a device to print. The base material 9 may be printing paper or a resin film. Further, the base material 9 may be a metal foil or a base material made of glass. As shown in FIG. 1, the printing device 1 includes a transport mechanism 10, a printing unit 20, a plurality of sensors 30, an image acquisition unit 40, and a computer 50.

The transport mechanism 10 is a mechanism for transporting the base material 9 in the transport direction along the longitudinal direction thereof. The transport mechanism 10 of the present embodiment includes an unwinding portion 11, a plurality of transport rollers 12, and a take-up portion 13. The base material 9 is unwound from the unwinding portion 11 and is conveyed along a transfer path composed of a plurality of transfer rollers 12. Each transfer roller 12 guides the base material 9 to the downstream side of the transfer path by rotating about the horizontal axis. The base material 9 is hung on a plurality of transport rollers 12 in a tensioned state. As a result, slack and wrinkles of the base material 9 during transportation are suppressed. The base material 9 after transportation is collected by the winding unit 13.

As shown in FIG. 1, the base material 9 moves below the plurality of heads 21 to 24 substantially parallel to the arrangement direction of the plurality of heads 21 to 24. At this time, the printed surface of the base material 9 is directed upward (heads 21 to 24). Hereinafter, among the plurality of transfer rollers 12, the four rollers located below the printing unit 20 will be referred to as the first roller 121, the second roller 122, the third roller 123, and the fourth roller 124, respectively. The first roller 121, the second roller 122, the third roller 123, and the fourth roller 124 are arranged in this order along the transport direction of the base material 9.

The printing unit 20 is a processing unit that ejects ink droplets (hereinafter referred to as “ink droplets”) to the base material 9 conveyed by the conveying mechanism 10. The printing unit 20 of the present embodiment has a first head 21, a second head 22, a third head 23, and a fourth head 24. The first head 21 is arranged above the first roller 121. The second head 22 is arranged above the second roller 122. The third head 23 is arranged above the third roller 123. The fourth head 24 is arranged above the fourth roller 124.

FIG. 2 is a partial top view of the printing apparatus 1 in the vicinity of the printing unit 20. As shown by the broken lines in FIG. 2, a plurality of nozzles 201 arranged in parallel with the width direction of the base material 9 are provided on the lower surface of each of the heads 21 to 24. Each of the heads 21 to 24 is directed from the plurality of nozzles 201 toward the upper surface of the base material 9, and each color of C (cyan), M (magenta), Y (yellow), and K (black), which are color components of the multicolor image. Ink droplets are ejected respectively.

That is, the first head 21 ejects C-color ink droplets onto the upper surface of the base material 9 at the first printing position P1 on the transport path. The second head 22 ejects M-color ink droplets onto the upper surface of the base material 9 at the second printing position P2 on the downstream side of the first printing position P1. The third head 23 ejects Y-color ink droplets onto the upper surface of the base material 9 at the third printing position P3 on the downstream side of the second printing position P2. The fourth head 24 ejects K-color ink droplets onto the upper surface of the base material 9 at the fourth printing position P4 on the downstream side of the third printing position P3.

In the present embodiment, the first printing position P1 is a position where the base material 9 comes into contact with the first roller 121. The second printing position P2 is a position where the base material 9 comes into contact with the second roller 122. The third printing position P3 is a position where the base material 9 comes into contact with the third roller 123. The fourth printing position P4 is a position where the base material 9 comes into contact with the fourth roller 124. The first printing position P1, the second printing position P2, the third printing position P3, and the fourth printing position P4 are arranged at intervals along the transport direction of the base material 9.

The four heads 21 to 24 print a single color image on the upper surface of the base material 9 by ejecting ink droplets. Then, a multicolor image is formed on the upper surface of the base material 9 by superimposing the four monochromatic images. Therefore, if the positions of the ink droplets ejected from the four heads 21 to 24 on the base material 9 in the transport direction are deviated from each other, the image quality of the printed matter deteriorates. It is important for improving the print quality of the printing apparatus 1 to suppress the magnitude of the mutual misalignment of the monochromatic images on the base material 9 (hereinafter referred to as "misregistration amount") within an allowable range. It becomes an element.

A drying processing unit for drying the ink discharged on the printing surface of the base material 9 may be further provided on the downstream side of the heads 21 to 24 in the transport direction. The drying treatment unit dries the ink by, for example, blowing a heated gas toward the base material 9 to vaporize the solvent in the ink adhering to the base material 9. However, the drying processing unit may dry the ink by another method such as light irradiation.

The plurality of sensors 30 are measuring instruments for measuring the conveyed state of the base material 9. The plurality of sensors 30 acquire measured values of items different from each other in each part of the transport mechanism 10. The measurement items of the sensor 30 include, for example, the rotation speed of the motor that operates the transfer mechanism 10, the rotation speed of some of the transfer rollers 12, the tension of the base material 9, and the vertical movement of the base material 9 (relative to the base material 9). The amount of variation in the vertical direction), the position of the edge of the base material 9 in the width direction, and the like can be included. Further, the sensors 30 for measuring the same item may be arranged at a plurality of positions on the transport path. The plurality of sensors 30 acquire the measured values of each measurement item in each part of the transport mechanism 10, and transmit a signal indicating the obtained measured values to the computer 50.

The image acquisition unit 40 is a camera that photographs the upper surface of the base material 9 that has passed through the printing unit 20. The image acquisition unit 40 is arranged so as to face the printing surface of the base material 9 at the photographing position P5 on the downstream side of the transport path from the printing unit 20. For the image acquisition unit 40, for example, a line sensor in which a plurality of image pickup elements such as CCD and CMOS are arranged in the width direction is used. The image acquisition unit 40 acquires image data of the printed base material 9 by photographing the printed surface of the base material 9. Then, the image acquisition unit 40 transmits the obtained image data to the computer 50.

The computer 50 is an information processing device for controlling the operation of each part in the printing device 1 and performing learning processing described later. FIG. 3 is a block diagram showing the connection between the computer 50 and each part of the printing device 1. As conceptually shown in FIG. 3, the computer 50 includes a processor 501 such as a CPU, a memory 502 such as a RAM, and a storage unit 503 such as a hard disk drive. A computer program CP for executing a printing process and a learning process described later is stored in the storage unit 503. Further, the computer 50 is communicably connected to the above-mentioned transport mechanism 10, four heads 21 to 24, a plurality of sensors 30, and an image acquisition unit 40, respectively. The computer 50 controls the operation of each of these parts according to the computer program CP and the learned model described later.

<2. About computers>
FIG. 4 is a block diagram conceptually showing the functions of the computer 50 described above. As shown in FIG. 4, the computer 50 includes a deviation amount measuring unit 51, a learning unit 52, and a control unit 53. Each function of the deviation amount measurement unit 51, the learning unit 52, and the control unit 53 is realized by operating the processor 501 of the computer 50 according to the computer program CP.

The deviation amount measurement unit 51 is a processing unit for measuring the above-mentioned misregistration amount based on the image data I input from the image acquisition unit 40. The deviation amount measurement unit 51 extracts an image of a specific pattern such as a register mark from the image data I, and recognizes the position of each color pattern in the extracted image by image processing. Then, the deviation amount measurement unit 51 calculates the mutual deviation amount of the positions of the patterns of each color as the actual measurement value D1 of the register deviation amount. The deviation amount measurement unit 51 acquires a large number of actual measurement values D1 by sequentially measuring such a misregistration amount with respect to the continuously input image data I. The acquired actual measurement value D1 of the misregistration amount is sent from the deviation amount measurement unit 51 to the learning unit 52.

The learning unit 52 receives the measured value D2 from the plurality of sensors 30. Further, the learning unit 52 acquires the measured value D1 of the misregistration amount from the deviation amount measuring unit 51. The learning unit 52 receives the measured value D2 as an input, uses the measured value D1 of the misplaced amount as the teacher data, and performs learning processing for estimating the misplaced amount from the measured value D2 by the supervised learning algorithm. As the supervised learning algorithm, for example, a support vector machine, a neural network, a linear model, a gradient boosting, or the like can be used. Specifically, while repeating the process of inputting the measured value D2 into the supervised learning model and outputting the estimated value of the misplaced amount from the model, the estimated value of the misplaced amount approaches the measured value D1. To adjust the model parameters. This will generate a trained model with adjusted parameters.

More specifically, as shown in FIG. 4, the learning unit 52 includes a first learning unit 521, an influence degree calculation unit 522, a measurement item selection unit 523, and a second learning unit 524. The learning unit 52 performs a two-step learning process by the first learning unit 521, the influence degree calculation unit 522, the measurement item selection unit 523, and the second learning unit 524. The details of the two-step learning process will be described later.

The control unit 53 is a processing unit for controlling the operations of the transport mechanism 10 and the four heads 21 to 24 described above. The control unit 53 uses the trained model generated by the learning unit 52 (the second trained model M2 described later) to control the movements of the transport mechanism 10 and the four heads 21 to 24 while correcting them.

<3. Flow of learning process and control process>
Subsequently, the flow of the learning process by the learning unit 52 and the control process by the control unit 53 will be described in detail. FIG. 5 is a flowchart showing the flow of the learning process and the control process.

As shown in FIG. 5, the learning unit 52 first receives the measured value D2 from the plurality of sensors 30, and also acquires the measured value D1 of the misregistration amount from the deviation amount measuring unit 51 (step S1). The measured values D2 of the plurality of sensors 30 include the rotational speed of the motor, the rotational speed of the transport roller 12, the tension of the base material 9, the vertical movement of the base material 9, the position of the edge of the base material 9 in the width direction, and the like. The measured values of the measurement items of are included. Further, the plurality of sensors 30 input a large number of measured values D2 obtained by continuously or intermittently measuring to the learning unit 52. Further, the deviation amount measurement unit 51 also inputs the actual measurement values D1 of a large number of misregistration amounts obtained by continuously or intermittently measuring to the learning unit 52.

The measured value D2 of the sensor 30 and the measured value D1 of the misregistration amount are acquired in the same time range. However, in order to learn the relationship between each measured value and the amount of misregistration at the printing positions P1 to P4, the time for acquiring the measured value D2 of each sensor 30 and the time for acquiring the measured value D1 for the amount of misregistration are used. May be shifted. For example, when the measured value D2 of the sensor 30 includes the rotation speed of a certain transport roller 12, the time to acquire the measured value D2 and the time to acquire the measured value D1 of the misregistration amount are printed on the transport roller 12. It may be shifted based on the positional relationship between the positions P1 to P4 and the photographing position P5.

The measured value D1 of the misregistration amount often has periodicity. Therefore, it is preferable that the time for acquiring the measured value D1 of the misregistration amount is at least one cycle of the fluctuation of the misregistration amount. Thereby, in the processing of the next steps S2 to S7, the relationship between the measured value D2 of the sensor 30 and the measured value D1 having periodicity can be sufficiently learned.

When the measured value D2 of a plurality of measurement items and the measured value D1 of the misplaced amount are accumulated in an amount sufficient for learning, the first learning unit 521 inputs the measured value D2 and the misplaced amount. The relationship between the measured value D2 and the measured value D1 of the misregistration amount is learned by the supervised learning algorithm using the measured value D1 of the above as the teacher data (step S2, first learning process). Specifically, the measured value D2 is input to the model for supervised learning, the estimated value of the misplaced amount is output, and the parameters of the model are set so that the estimated value of the misplaced amount approaches the measured value D1. adjust.

The first learning unit 521 repeats the learning process of step S2 above until a predetermined first end condition is satisfied (step S3: no). The first termination condition may be, for example, that the difference between the estimated value of the misregistration amount output from the model and the actually measured value D1 of the misplacement amount converges within a preset allowable range. Further, the first end condition may be that the number of times the learning process of step S2 is repeatedly executed reaches a preset number of times.

When the first end condition is satisfied, the learning unit 52 ends the learning process by the first learning unit 521 (step S3: yes). As a result, the first trained model M1 capable of outputting the estimated value of the misregistration amount is generated from the measured values D2 of the plurality of measurement items obtained from the plurality of sensors 30.

When the first trained model M1 is generated, the influence degree calculation unit 522 then calculates the influence degree in the first trained model M1 for each of the plurality of measurement items described above (step S4). The degree of influence is used when calculating an estimated value (objective variable) of the amount of misregistration from the measured values D2 (explanatory variable) of a plurality of measurement items in the model for supervised learning used in the first learning unit 521. , A value indicating the weighting of each measurement item. The degree of influence can be calculated by a predetermined formula from the trained parameters of the first trained model. Further, the influence degree calculation unit 522 normalizes and calculates the influence degree of each measurement item in a comparable form.

Subsequently, the measurement item selection unit 523 selects some measurement items among the plurality of measurement items based on the degree of influence calculated in step S4 (step S5). Specifically, first, the measurement item selection unit 523 arranges a plurality of measurement items in descending order of influence. Then, one or more measurement items having a high degree of influence are selected based on the preset selection criteria. For example, the measurement item selection unit 523 may select a preset number of measurement items in order from the one having the highest degree of influence. Alternatively, the measurement item selection unit 523 may select a measurement item having an influence degree larger than a preset threshold value.

After that, the second learning unit 524 inputs the measured value D2 of some of the measurement items selected in step S5, uses the measured value D1 of the misregistration amount as the teacher data, and uses the supervised learning algorithm to obtain the measured value D2. , The relationship between the misplaced amount and the measured value D1 is relearned (step S6, second learning process). Specifically, the measured value D2 of some of the measurement items selected in step S5 is input to the model for supervised learning, the estimated value of the misplaced amount is output, and the estimated value of the misplaced amount is calculated. Adjust the model parameters so that they approach the measured value D1.

As the measured value D2 and the measured value D1, the values acquired in step S1 may be used. Further, as the model for supervised learning, the same model as that used in step S2 may be used.

The second learning unit 524 repeats the learning process of step S6 above until a predetermined second end condition is satisfied (step S7: no). The second end condition may be, for example, that the difference between the estimated value of the misregistration amount output from the model and the actually measured value D1 of the misplacement amount converges within a preset allowable range. Further, the second end condition may be that the number of times the learning process of step S6 is repeatedly executed reaches a preset number of times.

When the second end condition is satisfied, the learning unit 52 ends the learning process by the second learning unit 524 (step S7: yes). As a result, the second trained model M2 capable of outputting the estimated value of the misregistration amount from the measured values D2 of some of the measurement items selected in step S5 is generated.

As described above, the learning unit 52 of the printing device 1 once generates the first trained model M1 by performing machine learning with the measured values D2 of a plurality of measurement items as inputs. Then, from a plurality of measurement items, some measurement items are selected in consideration of the degree of influence in the first trained model M1. After that, the second trained model M2 is generated by performing machine learning again with some of the selected measurement items as inputs. As a result, it is possible to generate the second trained model M2 with high estimation accuracy while reducing the number of input measurement items.

After the second trained model M2 is generated, the printing apparatus 1 executes a printing process on the base material 9 to be a product (step S8). At this time, the control unit 53 of the computer 50 uses the second learned model M2 to control the operations of the transfer mechanism 10 and the four heads 21 to 24 while correcting them.

Specifically, the control unit 53 acquires the measured value D2 from the sensor 30 that measures the measurement item selected in step S5 described above among the plurality of sensors 30 while operating the transport mechanism 10. Then, the acquired measured value D2 is input to the second trained model M2. Then, the estimated value of the misregistration amount is output from the second trained model M2. The control unit 53 corrects the operation of at least one of the transfer mechanism 10 and the four heads 21 to 24 based on the estimated value of the misregistration amount.

For example, when a certain amount of misregistration is estimated between the C color and the K color, the control unit 53 sets the ink ejection timing from the first head 21 or the fourth head 24 in a direction for canceling the misregistration. Shift to. In this way, the timing of ejecting ink droplets from the heads 21 to 24 is shifted in consideration of the amount of misregistration between each of the two colors. Alternatively, the rotation speed of the motor of the transport mechanism 10 is finely adjusted. As a result, the ink ejection position with respect to the base material 9 is corrected. As a result, it is possible to suppress misregistration between each of the two colors and obtain high-quality print results.

In particular, the second trained model M2 used in step S8 has a smaller number of measurement items to be input than the first trained model M1. By reducing the number of measurement items in this way, the calculation load of the control unit 53 at the time of executing the printing process can be reduced.

The printing device 1 of the present embodiment includes an image acquisition unit 40 for actually measuring the amount of misregistration. However, the image captured by the image acquisition unit 40 is an image at the imaging position P5 on the downstream side of the printing positions P1 to P4. Therefore, only the image acquisition unit 40 can grasp only the amount of misregistration after printing. However, if the second trained model M2 is used, the amount of misregistration generated at the print positions P1 to P4 can be estimated in advance based on the measured value D2 of each sensor 30.

Further, if the second trained model M2 is created, the amount of misregistration can be grasped without depending only on the image acquisition unit 40. Therefore, for example, even if the image acquisition unit 40 fails, the misregistration correction process can be performed based on the estimated value of the misregistration amount output from the second learned model M2.

Further, the image acquisition unit 40 may be installed only when the learning process of steps S1 to S7 described above is performed. Therefore, for example, the image acquisition unit 40 is installed and the learning process is performed only when the printing device 1 is started up or during maintenance, and after the learning process is completed, the image acquisition unit 40 is removed and the printing process is performed. You may.

<4. Modification example>
Although one embodiment of the present invention has been described above, the present invention is not limited to the above embodiment. Hereinafter, various modifications will be described focusing on the differences from the above-described embodiments.

<4-1. First modification>
The second trained model M2 may be updated after the printing process of step S8 described above is started. When updating the second trained model M2, the learning unit 52 acquires the measured value D2 from the sensor 30 that measures the measurement item selected in step S5 described above among the plurality of sensors 30. Further, the learning unit 52 acquires the measured value D1 of the misregistration amount from the deviation amount measuring unit 51. Then, the acquired measured value D2 is input, and the measured value D1 of the misregistration amount is used as the teacher data, and additional machine learning is performed by the supervised learning algorithm. As a result, the parameters of the second trained model M2 are updated.

If the second trained model M2 is updated, the misregistration amount can be estimated more accurately according to the change in the usage status of the printing apparatus 1. Therefore, the ink ejection position with respect to the base material 9 can be corrected more appropriately. In particular, in machine learning during the update process, only the measurement item selected in step S5 described above may be input from the plurality of measurement items. Therefore, the calculation load of the learning unit 52 can be reduced as compared with the case where all the measurement items are input.

The update process of the second trained model M2 is performed in parallel with, for example, the print process of step S8. In that case, the deviation amount measurement unit 51 may calculate the actual measurement value D1 of the misregistration amount based on the register mark printed on the margin portion of the base material 9. Further, the update process of the second trained model M2 may be performed at the time of maintenance between the print processes. In particular, when the type of the base material 9 to be conveyed in the printing apparatus 1 is changed, the relationship between the measured value D2 of each sensor and the misregistration amount is likely to change. It is preferable to do so. However, the update process of the second trained model M2 may be executed during the printing process on the same base material 9.

<4-2. Second modification>
In the above embodiment, the computer 50 executes a process (step S5) of selecting some measurement items based on the degree of influence. However, this selection process may be performed by the user of the printing device 1. In that case, the computer 50 displays the degree of influence of each measurement item calculated in step S4 on the screen of the display. The user may select some measurement items and input them to the computer 50 while referring to this screen.

However, as in the above embodiment, the computer 50 automatically selects the measurement item, so that the work load on the user can be further reduced.

<4-3. Other variants>
In the above embodiment, the measured value D2 of the sensor 30 is directly input to the model for supervised learning. Further, in the above embodiment, the measured value D1 of the misregistration amount is also used as it is as the teacher data. However, the measured value D2 of the sensor 30 may be processed into a value suitable for machine learning by performing a predetermined calculation or filter processing, and then input to the model. Further, the measured value D1 of the misregistration amount may also be used as teacher data after being processed into a value suitable for machine learning by performing a predetermined calculation or filter processing.

Further, in step S5 of the above embodiment, some measurement items having a high degree of influence were selected from the plurality of measurement items. However, the measurement items do not necessarily have to be selected in descending order of influence. For example, a measurement item that is known to have an effect in a specific situation may be included in the measurement items to be selected even if the degree of influence calculated in step S4 is low.

Further, in the above embodiment, the learning unit 52 and the control unit 53 are realized by the same computer 50. However, the learning unit 52 and the control unit 53 may be realized by separate computers.

Further, in the above embodiment, as shown in FIG. 2, the nozzles 201 are arranged in a row in the width direction in each of the heads 21 to 24. However, the nozzles 201 may be arranged in two or more rows in each of the heads 21 to 24.

Further, the printing apparatus 1 of the above embodiment includes four heads 21 to 24. However, the number of heads included in the printing apparatus 1 may be 2, 3, or 5 or more. For example, the printing apparatus 1 may include a head that ejects a special color ink in addition to each of the colors C, M, Y, and K.

Further, the elements appearing in the above-described embodiments and modifications may be appropriately combined as long as there is no contradiction.

1 Printing device 9 Base material 10 Conveying mechanism 11 Unwinding part 12 Conveying roller 13 Winding part 20 Printing part 21 1st head 22 2nd head 23 3rd head 24 4th head 30 Sensor 40 Image acquisition part 50 Computer 51 Misalignment amount Actual measurement unit 52 Learning unit 53 Control unit 521 1st learning unit 522 Impact degree calculation unit 523 Measurement item selection unit 524 2nd learning unit CP computer program D1 Actual measurement value of misregistration amount D2 Sensor measurement value I Image data M1 1st learning Completed model M2 2nd learned model P1 1st printing position P2 2nd printing position P3 3rd printing position P4 4th printing position P5 Shooting position

Claims (8)

In a printing device that ejects ink from a plurality of heads to the surface of the base material while transporting the long strip-shaped base material in the longitudinal direction along a predetermined transport path, measurement by a plurality of sensors provided in the printing device. It is a learning method for estimating the amount of mutual deviation of the ink ejection positions by the plurality of heads from the values.
a) The measured values of a plurality of measurement items obtained from the plurality of sensors are input, the measured values of the deviation amount are used as teacher data, and the estimated value of the deviation amount corresponding to the measured values is calculated by a supervised learning algorithm. The process of generating the first trained model to be output and
b) A step of calculating the degree of influence in the first trained model for each of the plurality of measurement items, and
c) A step of selecting a part of the measurement items from the plurality of measurement items based on the degree of influence, and
d) The measured value of the measurement item selected in the step c) is input, the measured value of the deviation amount is used as the teacher data, and the estimated value of the deviation amount corresponding to the measured value is obtained by the supervised learning algorithm. The process of generating the second trained model to be output and
Learning method with. The learning method according to claim 1.
In the step c), a learning method in which a part of the measurement items having a high degree of influence is selected from the plurality of measurement items. The learning method according to claim 1 or 2.
In the step c), a learning method in which a computer that executes the steps a), b), and d) selects a part of the measurement items from the plurality of measurement items according to the degree of influence. The learning method according to any one of claims 1 to 3.
The measured value of the deviation amount has periodicity and has periodicity.
In the step a) and the step d), a learning method in which the actually measured values acquired in a time of one cycle or more of the fluctuation of the deviation amount are used as teacher data. A control method for controlling the printing apparatus by using the second trained model generated by the learning method according to any one of claims 1 to 4.
e) A step of inputting the measured values of the measurement items selected in the step c) into the second trained model and obtaining an estimated value of the deviation amount output from the second trained model.
f) A step of correcting the ink ejection position with respect to the base material based on the estimated value of the deviation amount obtained in the step e).
A control method having. The control method according to claim 5.
g) Further having a step of updating the second trained model by the supervised learning algorithm using the measured values of the measurement items selected in the step c) as input and the measured values of the deviation amount as teacher data. , Control method. The control method according to claim 6.
The control method g) is executed in the printing apparatus when the type of the base material is changed or during the printing process on the same base material. A transport mechanism that transports a long strip-shaped base material in the longitudinal direction along a predetermined transport path,
Multiple heads that eject ink to the surface of the base material,
A learning unit that generates the second trained model by the learning method according to any one of claims 1 to 4.
Using the second trained model generated by the learning unit, a control unit that controls while correcting at least one of the transfer mechanism and the plurality of heads.
A printing device equipped with.

Images