Method for creating drive waveform, information processor, and program

Abstract

To provide a method for creating a drive waveform, an information processor, and a program which enable even an engineer having no expertise to efficiently create a drive waveform suitable for discharging a used liquid.SOLUTION: A method for creating a drive waveform to be used for driving a piezoelectric element of a liquid discharge head including the piezoelectric element includes, via one or more processors, predicting flight of liquid to be discharged by the liquid discharge head in a case of inputting an unknown drive waveform using a machine learning model that is trained through machine learning using data related to an actual flight shape of the liquid in a case where each of a plurality of drive waveforms is applied to the piezoelectric element using the liquid and the liquid discharge head, and determining a drive waveform suitable for discharging the liquid based on the prediction of the flight.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to a drive waveform creating method, an information processing device, and a program, and in particular to a technology for creating a drive waveform applied to a liquid ejection head that ejects liquid by driving a piezoelectric element, and an information processing technology for executing the processing. [Background technology]

[0002] In inkjet printing, when the ink used is different, even a slight change in the physical properties will change the flight shape of the ink ejected from the inkjet head, so obtaining good ejection characteristics has been a major challenge. The ejection characteristics can include, for example, landing position accuracy, the presence or absence of satellite droplets, droplet speed, droplet volume, and stability. In inkjet heads that eject ink by driving a piezoelectric element, there is a degree of freedom in the drive waveform, so developers often optimize the drive waveform for each ink used.

[0003] Patent Document 1 describes a system having a device that ejects liquid material using an inkjet head, the ejecting device including a unit that acquires identification information of the inkjet head, a unit that supplies a drive pulse to an actuator of the inkjet head for ejecting the liquid material, and an inspection unit that detects the state of droplets ejected from the inkjet head, and the system further includes a database in which the ejection characteristics of each inkjet head are associated with the identification information of each inkjet head, and an optimization unit that provides first optimization information for generating an optimized drive pulse based on the ejection characteristics of the inkjet head obtained from the identification information for a hypothetical attribute assumed for the liquid material to be ejected by the ejecting device, the optimization unit including a dynamic optimization unit that detects the state of the droplets ejected by the drive pulse generated based on the first optimization information using the inspection unit, assumes actual attributes regarding the ejection of the liquid material to be ejected based on the ejection characteristics of the inkjet head obtained from the identification information, and provides second optimization information that dynamically optimizes the drive pulse for the assumed actual attribute. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2021-160314 A Summary of the Invention [Problem to be solved by the invention]

[0005] In order to optimize the drive waveform for the ink being used, it has been common to predict the flight shape of the ink in response to the input of the drive waveform using an equivalent circuit model or a physical simulation method such as Computational Fluid Dynamics (CFD). However, with this method, it is difficult to build a model to use for prediction without advanced knowledge and experience in fluid dynamics and numerical calculations.

[0006] In addition, a common method for optimizing a drive waveform is to select from a group of drive waveforms prepared in advance, evaluate the characteristics, and determine the optimal drive waveform that satisfies the conditions of the desired characteristics. However, optimization involving trial and error requires a lot of time. Although attempts have been made to shorten the time required for optimizing the drive waveform, the conventional general method only provides a limited group of drive waveforms, making it impossible to search for a completely unknown drive waveform.

[0007] The above-mentioned problem is not limited to inkjet devices for printing purposes, but is a common problem to devices that use liquid ejection heads that eject various functional liquids.

[0008] The present disclosure has been made in consideration of the above circumstances, and aims to provide a drive waveform creation method, information processing device, and program that enable even an engineer who does not have advanced knowledge and experience in creating drive waveforms to efficiently create a drive waveform suitable for ejecting the ink to be used. [Means for solving the problem]

[0009] A drive waveform creation method according to a first aspect of the present disclosure is a method for creating a drive waveform used to drive a piezoelectric element of a liquid ejection head equipped with a piezoelectric element, and includes having one or more processors use a machine learning model trained by machine learning using data relating to the liquid to be ejected from the liquid ejection head and the actual flight shape of the liquid when each of a plurality of drive waveforms is applied to the piezoelectric element using the liquid ejection head to predict the flight of the liquid when an unknown drive waveform is input, and determining a drive waveform suitable for ejecting the liquid based on the flight prediction.

[0010] According to the first aspect, by using a trained machine learning model that has learned the relationship between drive waveforms and flight shapes through machine learning using data on actual flight shapes, it is possible to predict the flight shape of an unknown drive waveform, and a drive waveform suitable for ejecting liquid can be efficiently found based on this flight prediction.

[0011] A drive waveform generating method according to a second aspect may be configured in the drive waveform generating method according to the first aspect, wherein the parameters of the drive waveform include at least one of a pulse width, a slope, a pulse height, and a pulse interval.

[0012] A drive waveform creation method according to a third aspect may be configured such that, in the drive waveform creation method according to the first or second aspect, the learning phase of the machine learning model includes a step of compressing each of the multiple drive waveforms into a latent space of a dimension smaller than the dimension of the drive waveform.

[0013] A drive waveform generating method according to a fourth aspect may be configured such that, in the drive waveform generating method according to the third aspect, the drive waveform is input to an autoencoder and converted into coordinates in a latent space.

[0014] A drive waveform creating method according to a fifth aspect may be configured such that, in the drive waveform creating method according to the third or fourth aspect, the machine learning model is trained in a learning phase to predict an evaluation value when a drive waveform is applied, using a correspondence between coordinates in a latent space of each of a plurality of drive waveforms and an evaluation value based on an actual flight shape.

[0015] A drive waveform creation method according to a sixth aspect may be configured such that, in the drive waveform creation method according to the fifth aspect, the data relating to the actual flight shape includes an evaluation value indicating a characteristic extracted from an image in which the actual flight shape is captured.

[0016] A drive waveform creation method according to a seventh aspect may be configured in such a way that, in the drive waveform creation method according to the fifth or sixth aspect, the evaluation value includes at least one of values ​​indicating the droplet speed, droplet volume, and the presence or absence of satellite droplets of the liquid ejected from the liquid ejection head.

[0017] A drive waveform creation method according to an eighth aspect may be configured such that, in the drive waveform creation method according to any one of the fifth to seventh aspects, the flight prediction includes predicting an evaluation value, and one or more processors generate one or more unknown drive waveforms different from the multiple drive waveforms, calculate coordinates in a latent space from the unknown drive waveform, calculate an evaluation value predicted from the coordinates in the latent space of the unknown drive waveform using a machine learning model, and compare the evaluation value calculated using the machine learning model with a target value to determine a drive waveform that satisfies the target value.

[0018] A drive waveform creation method according to a ninth aspect may be a drive waveform creation method according to any one of the fifth to eighth aspects, wherein the machine learning model is a model that outputs an average value and a standard deviation of an evaluation value predicted from coordinates in a latent space.

[0019] A drive waveform creation method according to a tenth aspect may be configured in such a way that, in the drive waveform creation method according to the ninth aspect, one or more processors generate one or more unknown drive waveforms different from the multiple drive waveforms, calculate coordinates in a latent space from the unknown drive waveform, calculate an average value and a standard deviation of evaluation values ​​predicted from the coordinates in the latent space using a machine learning model, calculate a probability that the evaluation value will exceed a target value from the average value and standard deviation of the evaluation values ​​calculated using the machine learning model, and determine the drive waveform with the highest probability of exceeding the target value as the appropriate drive waveform.

[0020] A drive waveform creation method according to an eleventh aspect may be configured such that, in the drive waveform creation method according to any one of the eighth to tenth aspects, one or more processors calculate coordinates in a latent space from an unknown drive waveform using an autoencoder.

[0021] A drive waveform creation method according to a 12th aspect may be configured in such a way that, in the drive waveform creation method according to any one of the first to 11th aspects, one or more processors randomly extract parameter values ​​of the drive waveform based on a uniform distribution to generate a plurality of unknown drive waveforms different from the plurality of drive waveforms, and perform flight prediction for each drive waveform using a machine learning model.

[0022] A drive waveform creation method according to a thirteenth aspect may be configured such that, in the drive waveform creation method according to any one of the fifth to eleventh aspects, when one or more processors generate a plurality of unknown drive waveforms different from the plurality of drive waveforms by randomly extracting parameter values ​​of the drive waveform based on a uniform distribution, the relationship between the distance in the latent space and the variance of the evaluation value is clarified in advance by a variogram analysis, and the step size for searching the drive waveform is set based on the variogram analysis.

[0023] A drive waveform creation method according to a fourteenth aspect may be configured such that, in the drive waveform creation method according to the thirteenth aspect, the search step size is set to a distance equal to or greater than a distance at which the distance in the latent space and the variance of the evaluation value are uncorrelated, based on variogram analysis.

[0024] An information processing device according to a 15th aspect of the present disclosure is an information processing device that executes a drive waveform creation method according to any one of the 1st to 14th aspects, and includes one or more processors and one or more storage devices in which a machine learning model is stored.

[0025] A program according to a sixteenth aspect of the present disclosure is a program for causing a computer to execute the drive waveform generating method according to any one of the first to fourteenth aspects. Effect of the Invention

[0026] According to the present disclosure, even an engineer who does not have expertise in creating drive waveforms used in liquid ejection heads equipped with piezoelectric elements can efficiently create drive waveforms suitable for ejecting the liquid to be used. [Brief description of the drawings]

[0027] [Figure 1] FIG. 1 is a flowchart showing a processing procedure of a driving waveform generating method according to an embodiment. [Diagram 2] FIG. 2 is a waveform diagram showing an example of a driving waveform. [Diagram 3] FIG. 3 is an example of an image in which the flight shape is captured. [Figure 4] FIG. 4 is an image diagram of droplets ejected from an inkjet head. [Diagram 5] FIG. 5 is an explanatory diagram illustrating an example of the configuration of an autoencoder. [Figure 6] FIG. 6 is a diagram showing an example of mapping characteristic values ​​onto a latent space. [Figure 7] FIG. 7 is a graph showing an example of a variogram analysis of droplet velocity. [Figure 8] FIG. 8 is a graph showing an example of a variogram analysis of drop volume. [Figure 9] FIG. 9 is a graph showing an example of a variogram analysis regarding the number of drops. [Figure 10] FIG. 10 is a block diagram illustrating an example of a hardware configuration of an information processing device. [Figure 11] FIG. 11 is an explanatory diagram that shows a schematic configuration example of an inkjet device used in a discharge experiment for obtaining data used in learning. [Figure 12] FIG. 12 is a block diagram illustrating a schematic functional configuration of an information processing device that executes an autoencoder creation process. [Figure 13] FIG. 13 is a block diagram showing a schematic functional configuration of an information processing device that executes processing for creating a prediction model. [Figure 14] FIG. 14 is a block diagram showing a schematic functional configuration of an information processing device that executes a process of creating a learning dataset. [Figure 15] FIG. 15 is an explanatory diagram showing an example of a data set. [Figure 16]FIG. 16 is a block diagram showing a schematic functional configuration of an information processing device that executes processing for creating a prediction model. [Figure 17] FIG. 17 is a block diagram showing a schematic functional configuration of an information processing device that executes a process of searching for a promising driving waveform using a trained autoencoder and a prediction model constructed according to this embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0028] Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

[0029] Overview of the driving waveform generating method according to the embodiment In this embodiment, an example of a method and device for creating a machine learning model that predicts the behavior of an inkjet head equipped with a piezoelectric element, and a method and device for searching for a drive waveform that can achieve desired characteristics using the trained machine learning model are described.

[0030] Fig. 1 is a flowchart showing a processing procedure of a drive waveform creating method according to an embodiment. Each step from step S1 to step S8 shown in Fig. 1 is executed by one or more processors. Steps S1 to S5 are processing steps for creating a prediction model (machine learning model) using data on the actual flight shape when ink is ejected by applying each of a plurality of drive waveforms to a piezoelectric element through an ejection experiment using a combination of the ink and inkjet head to be used. Steps S6 to S8 are processing steps for searching for an appropriate drive waveform using a trained prediction model. Steps S1 to S5 correspond to the learning phase, and step S6 corresponds to the inference phase.

[0031] Here, an example will be described in which a first processor executes steps S1 to S5, and then a second processor different from the first processor executes steps S6 to S8, but, for example, the first processor may execute steps S6 to S8 instead of the second processor. Also, a third processor different from the second processor may execute step S3 instead of the first processor. Each of steps S1 to S8 will be described in detail below.

[0032] [Step S1: Acquiring an image showing the actual flight shape] In step S1, the first processor acquires an image (hereinafter referred to as a "flying shape image") of the flying shape when each of a plurality of driving waveforms is applied to a piezoelectric element using the ink and inkjet head to be used. In this embodiment, a discharge experiment is performed by actually applying a plurality of driving waveforms using a combination of the ink and inkjet head to be used, and a large amount of data on the correspondence between the driving waveform, which is the input, and the actual flying shape of the ink, which is the output, is collected.

[0033] [Example of driving waveform] FIG. 2 is a waveform diagram showing an example of a drive waveform. The horizontal axis represents time, and the vertical axis represents potential. The drive waveform 20 shown in FIG. 2 includes a preliminary vibration pulse 22, an ejection pulse 24, and a reverberation suppression pulse 26, and the parameters of the drive waveform are the width, slope, pulse height, and pulse interval of each pulse. Here, an example in which the drive waveform is expressed by 12 parameters will be described. In the example of the drive waveform 20 shown in FIG. 2, there are 12 parameters, namely, times t1 to t9 that define the pulse width, slope, and pulse interval, and potential differences E1 to E3 that define the pulse height. A plurality of drive waveforms with different combinations of the values ​​of these parameters are applied to a piezoelectric element of an inkjet head filled with the ink to be used, and the flight shape of the ejected ink is used as learning data.

[0034] The parameters of the drive waveform are not limited to the types (12 types) shown in Fig. 2. For example, the potential of the drive waveform may be changed in a curved manner with time, and the shape of the curve may be included in the parameters. The types of drive waveforms used for learning may be, for example, 100 types.

[0035] [Example of flight shape image] Figure 3 shows an example of an image of the flight shape of ink ejected from an inkjet head. Figure 3 shows the flight shape at each time as understood from a time-series image group obtained by continuously photographing ink ejected from an inkjet head by application of a drive waveform at regular time intervals. Figure 3 shows an example of images taken at intervals of 1 μs (microseconds). Note that 1 μs is an example of a regular time interval.

[0036] It is desirable to set an area sufficient to obtain the ink flight characteristics from the nozzle, which is the ink ejection port, as the shooting area. To grasp the flight state in the chronological direction (time axis direction), images are taken at regular time intervals, with the number of steps (number of images taken) sufficient to show the ink droplets outside the screen. Therefore, for one drive waveform, images for the number of time series are obtained. Figure 3 shows an example in which the area of ​​interest is cropped from images for the number of time series and then arranged in chronological order.

[0037] Considering subsequent image processing, it is preferable that the color contrast between the ink area and the background area is as clear as possible. Also, it is preferable that the area that separates the ink droplets from the background area has a sharp resolution.

[0038] As shown in FIG. 3, ink begins to be ejected from the nozzle of the inkjet head, forming a liquid column, which then separates from the nozzle and transforms into a droplet as it flies.

[0039] [Step S2: Extraction of characteristics from flight shape images] In step S2 of Fig. 1, the first processor extracts characteristics from the acquired flight shape image by image processing. The "characteristics" here are, for example, the droplet volume, droplet speed, and the presence or absence of satellite droplets, but may also include other characteristics. The droplet volume, droplet speed, and the presence or absence of satellite droplets are examples of evaluation values ​​(evaluation indexes) calculated based on the flight shape.

[0040] Droplet speed is the speed of the ink droplet, and is calculated by extracting the ink area through image processing and determining how much the ink area changes per unit time. Droplet volume is the volume of the ink droplet, and is calculated by converting the area of ​​the ink area extracted through image processing into a volume equivalent. In this case, only the droplet volume that actually separates from the nozzle and flies is added, and ink that does not separate from the nozzle and returns to the nozzle is not added. The presence or absence of satellite droplets refers to the difference between the state in which an ink droplet normally transforms into a single sphere after being ejected and flies (no satellite droplets), and the state in which an ink droplet splits into two or more spheres during the transformation and flies (satellite droplets present) (see F4B on the right of Figure 4).

[0041] An image of the droplet after ejection is shown in Figure 4. The left image F4A in Figure 4 shows the state where the liquid column extends from the ejection port immediately after ejection starts. The right image F4B shows the state where the main droplet and satellite droplets then split into spheres and fly away (with satellite droplets).

[0042] The presence or absence of satellite droplets is also judged based on whether the area is divided into multiple parts when ink droplets are extracted from the flight shape image by image processing. Furthermore, the presence or absence of satellite droplets is considered only for droplets that actually separate from the nozzle and fly, and ink that does not separate from the nozzle and returns to the nozzle is not considered in the droplet volume. Furthermore, the presence or absence of satellite droplets is a binary judgment result, but the distance of the final droplet when the first droplet (main droplet) reaches a certain distance (if there is no satellite droplet, the distance is set to 0), which is a numerical value indicating the distance.

[0043] [Step S3: Creating an autoencoder] In step S3 of Fig. 1, the first processor generates various possible forms of driving waveforms, compresses the driving waveforms into a latent space using an autoencoder, and optimizes the parameters of the autoencoder so as to reconstruct the input driving waveform from the compressed information. Note that "optimization" means approaching an optimal state, and is not limited to reaching a truly optimal state.

[0044] This step S3 is a process step of creating an autoencoder that compresses high-dimensional data of the driving waveform into a lower-dimensional latent space as a preliminary step for creating a prediction model. Step S3 may be executed as a process independent of steps S1 and S2, or may be executed prior to steps S1 and S2.

[0045] The process of step S3 will be described using an example in which the drive waveform is expressed by 12 parameters as shown in FIG. 2. The first processor randomly generates these parameters to generate various drive waveforms. Since the parameters E1, E2, and E3 shown in FIG. 2 represent potentials, the first processor extracts random real numbers from the range of potentials that can be input to the inkjet head. For the random real numbers, a uniform distribution may be assumed, or a normal distribution or other probability distribution may be assumed. However, the interval width is set to about the potential resolution of the input. Similarly, an appropriate range is set for t1 to t9, and a value corresponding to time is randomly extracted. Again, the interval width is set to about the time resolution of the input. In addition, inputs that are obviously not appropriate may be excluded in advance.

[0046] Any number of drive waveforms generated by random real numbers can be generated as long as time and memory capacity allow, and the more the number, the higher the learning effect, but for example, about 100,000 to 1 million points. The one-dimensional potential of the drive waveform is the number of steps of the time resolution, making it very high-dimensional. For example, when the drive waveform length is 30 μs and the time resolution is 0.01 μs, the number of steps expressing the potential of the drive waveform is 3000 points, and the drive waveform becomes a 3000-dimensional vector.

[0047] We consider mapping such high-dimensional vectors into a relatively low-dimensional space, the latent space. In recent years, a method called an autoencoder has been proposed as a method for mapping into the latent space. Autoencoders are a machine learning method that is widely used in fields such as artificial intelligence.

[0048] In step S3 of Figure 1, the first processor randomly generates a wide variety of drive waveforms, performs unsupervised learning using these many drive waveforms, and optimizes the parameters of the autoencoder so as to reconstruct and output a waveform identical to the input drive waveform.

[0049] As long as there is no change in the vectorization format (time resolution, etc.) that represents the drive waveform, once the parameters of the autoencoder are optimized, the optimized (trained) autoencoder can be reused.

[0050] Therefore, if a trained autoencoder is available, step S3 can be omitted.

[0051] [Autoencoder example] FIG. 5 shows an example of an autoencoder 50. The autoencoder 50 downsamples the input high-dimensional vector WFi multiple times in a convolutional layer, etc., outputs it as a low-dimensional latent space, and then upsamples it in a full-connection layer or a convolutional layer to output a reconstructed vector RWFi with the same dimensions as the input. At this time, a low-dimensional latent space is established by learning so that the input and output become the same vector. Note that the numbers in parentheses ( ) shown in FIG. 5 represent the order of the input tensor in the layer. For example, the input drive waveform (100, 400, 1) represents a tensor of 100 × 400 × 1. The dimensions of this input drive waveform correspond to the following, respectively.

[0052] [Table 1] In a low-dimensional latent space, vectors corresponding to waveforms that are similar as driving waveforms are arranged close to each other (close in distance defined by Euclidean distance, etc.). Similar driving waveforms can be considered to have similar characteristics.

[0053] The autoencoder 50 can change the intermediate layer to various ones depending on the purpose and application, but since the input vector is a time-series signal, it may be realized by a method that can handle time-series data, such as a recurrent neural network (RNN) or a long short-term memory (LSTM).

[0054] This method can compress, for example, an input drive waveform of about 3000 dimensions into a latent space of about 10 to 20 dimensions. For example, the autoencoder 50 can compress a 3000-dimensional vector into a vector in a 16-dimensional latent space.

[0055] [Step S4: Making latent variables of the driving waveform used in the ejection experiment] In step S4 of FIG. 1, the first processor inputs each of the multiple drive waveforms for which the ejection experiment was performed in step S1 into the trained autoencoder to obtain a vector in the latent space corresponding to each drive waveform. The vector in the latent space may be understood as a coordinate indicating a position in the latent space. The multidimensional drive waveform data is converted into a vector in a lower dimensional latent space by the autoencoder. Step S4 is an example of a process of compressing the drive waveform into a latent space of a dimension smaller than that of the drive waveform.

[0056] For example, by inputting the 100 types of drive waveforms obtained in step S1 into the trained autoencoder obtained in step S3, each drive waveform is mapped into a 16-dimensional latent space. It can be confirmed that similar drive waveforms are placed at positions close to each other in the latent space.

[0057] [Step S5: Creating a predictive model] In step S5, the first processor performs machine learning using data in which the vectors in the latent space obtained in step S4 are linked to the characteristics obtained in step S2, learns the correspondence between the vectors in the latent space and the characteristics, and creates a prediction model so that predicted values ​​of the characteristics can be output for any drive waveform.

[0058] The 100 types of driving waveforms obtained in step S1 correspond to the characteristics obtained in step S2. In other words, a space can be created in which the quantities (y values) of each characteristic, such as droplet volume, droplet speed, and the presence or absence of satellite droplets, correspond to vectors (x coordinates) in the 16-dimensional latent space (see Figure 6).

[0059] Figure 6 shows an example of mapping characteristic values ​​onto a latent space. For convenience of illustration, the latent space is two-dimensional in Figure 6, and the number of drops is shown as an example of a characteristic value. The number of drops can be an indicator of the presence or absence of satellite drops.

[0060] In this way, when there are multiple pieces of y-value data corresponding to an x-coordinate, a machine learning model is known that predicts a y-value at an unknown x-coordinate using multiple correspondences between known x-coordinates and y-values. For example, Gaussian process regression uses multiple combinations of x-coordinates and y-values ​​as training data to output the average value and standard deviation of y-values ​​for unknown x-coordinates. For similar purposes, a general linear regression model, a generalized linear model, or a support vector machine can also be used. In step S5, the parameters of the machine learning model are optimized, and a prediction model that predicts a y-value at an unknown x-coordinate is created.

[0061] [Step S6: Predicting flight for new drive waveforms using a prediction model] In step S6 of FIG. 1, the second processor generates a new driving waveform that was not implemented in step S1, inputs it to the autoencoder, obtains a corresponding vector in the latent space, and predicts its characteristics using a prediction model.

[0062] That is, in step S6, the driving waveforms that were not used in the ejection experiment in step S1 are randomly extracted, mapped to a latent space using an autoencoder, and predicted characteristics are obtained from a prediction model. The method of random extraction for generating a new driving waveform may be the same as the method of random extraction described in step S3.

[0063] [Step S7: Evaluation of predicted characteristics] 1, when the predicted characteristics are acquired in step S6, the second processor determines whether the predicted characteristics satisfy the desired characteristics. For example, if the conditions specified as the target values ​​of the desired characteristics are "droplet speed 7 m / s or more, drop volume 3 pL (picoliters) or more, and no satellite drops," the second processor determines whether the predicted characteristics satisfy all of these target value conditions.

[0064] If the result of the determination in step S7 is No, the second processor returns to step S6 and predicts the characteristics of another drive waveform. In other words, if the predicted characteristics do not satisfy any of the conditions (desired conditions) of the target characteristics specified in advance, the process returns to step S6, a new drive waveform is generated, and the comparison determination (step S7) between the predicted characteristics and the desired characteristics is repeated until successful.

[0065] On the other hand, if the result of the determination in step S7 is Yes, the second processor proceeds to step S8. That is, if the predicted characteristics satisfy all the desired conditions, the extraction is successful, and the process proceeds to step S8.

[0066] [Step S8: Determining the driving waveform] In step S8 of Fig. 1, the second processor determines the drive waveform whose predicted characteristics satisfy the desired conditions as the drive waveform suitable for ejecting the ink to be used. After step S8, the flow chart of Fig. 1 ends.

[0067] When multiple drive waveforms that satisfy the desired conditions are to be created, steps S6 to S8 may be repeated until the number of successes reaches a desired number.

[0068] Also, by scoring the predicted characteristics, those with better scores may be extracted with emphasis. For example, when a prediction model is constructed by Gaussian process regression in step S5, the average value and standard deviation of the predicted characteristics are obtained as the output of the prediction model, and the upper probability of the normal distribution is used to calculate the level of prediction accuracy. For example, the probability that the drop speed is 7 m / s or more can be obtained. Similarly, the probability of satisfying each condition of the drop amount and the presence or absence of satellite drops and the joint probability are obtained, and the level of prediction accuracy of the predicted characteristics can be evaluated. Using the level of prediction accuracy thus obtained as an index, those with higher prediction accuracy can be given a higher priority and used. In other words, those with a higher probability of the predicted characteristics exceeding the target value can be determined as the appropriate drive waveform. Similarly, the predicted characteristics may be scored according to the degree to which they exceed the conditions.

[0069] <<Ingenuity in searching for driving waveforms>> When randomly extracting a new drive waveform in step S6, a method of extracting from a general probability distribution such as a uniform distribution or a normal distribution may be used, as in step S3. However, by utilizing the evaluation indexes described above, various combinatorial optimization methods (gradient descent method), genetic algorithms, Markov chain Monte Carlo methods, bandit algorithms, etc. can be used.

[0070] In addition, when performing random extraction, it is essential to adjust the search resolution from the perspective of search efficiency (time-to-benefit), and variogram analysis can be used as an indicator of how large the search increment should be set. Variogram analysis is a method for analyzing the correlation between the distance in the x coordinates of any sample points and the difference in y values. In general, if the distance between x coordinates is small, the difference in y values ​​is small, so for vector combinations of equal distances, the variance of y values ​​is small. Increasing the distance between x coordinates increases the variance of y values, and eventually the distance and y value variance become uncorrelated (see Figures 7 to 9). Since little information can be obtained by extracting distances shorter than this uncorrelated distance, the search increment can be set to a value larger than this to further improve search efficiency.

[0071] Figures 7 to 9 are graphs showing examples of variogram analysis, with Figure 7 showing the analysis results for the drop speed, Figure 8 showing the drop volume, and Figure 9 showing the number of drops. In these figures, the distances shown by dashed lines represent the distances (ranges) at which the variance of the distance and the y value are uncorrelated. In this way, it is preferable to perform a variogram analysis in advance and set the search step size based on the results.

[0072] Example of Hardware Configuration of Information Processing Device The processes from step S1 to step S8 can be executed by a computer system including one or more computers.

[0073] FIG. 10 is a block diagram showing an example of a hardware configuration of an information processing device 100 that executes at least a part of the processing of the driving waveform generating method according to the embodiment.

[0074] The information processing device 100 includes a processor 102, a computer-readable medium 104 which is a non-transient tangible entity, a communication interface 106, an input / output interface 108, and a bus 110. The processor 102 is connected to the computer-readable medium 104, the communication interface 106, and the input / output interface 108 via the bus 110. The form of the information processing device 100 is not particularly limited, and may be a server, a personal computer, a workstation, a tablet terminal, or the like.

[0075] The processor 102 can be at least one of a first processor and a second processor. The processor 102 includes a CPU (Central Processing Unit). The processor 102 may include a GPU (Graphics Processing Unit). The computer-readable medium 104 includes a memory 112 which is a main storage device and a storage 114 which is an auxiliary storage device. The computer-readable medium 104 may be, for example, a semiconductor memory, a hard disk drive (HDD) device, a solid state drive (SSD) device, or a combination of a plurality of these. The computer-readable medium 104 is an example of a "storage device" in this disclosure.

[0076] A plurality of programs and data for performing various processes are stored in the computer-readable medium 104. The term "program" includes the concept of a program module. The processor 102 executes the instructions of the programs stored in the computer-readable medium 104 to function as various processing units.

[0077] The information processing device 100 can be connected to an electric communication line (not shown) via the communication interface 106. The electric communication line may be a wide area communication line, a private network communication line, or a combination of these.

[0078] The information processing device 100 may include an input device 152 and a display device 154. The input device 152 is configured, for example, by a keyboard, a mouse, a multi-touch panel, or other pointing device, or a voice input device, or an appropriate combination of these. The display device 154 is configured, for example, by a liquid crystal display, an organic electro-luminescence (OEL) display, or a projector, or an appropriate combination of these. The input device 152 and the display device 154 are connected to the processor 102 via the input / output interface 108.

[0079] 《How to collect data for learning》 FIG. 11 is an explanatory diagram that shows a schematic configuration example of an inkjet device 200 used in a discharge experiment for obtaining data used for learning.

[0080] The inkjet device 200 includes an inkjet head 202 , a drive circuit 250 , an information processing device 300 , and a camera 320 .

[0081] 11 shows a cross-sectional view of the three-dimensional structure of one ejector 210 in the inkjet head 202, but the inkjet head 202 includes a plurality of ejectors 210. The inkjet head 202 is an example of a "liquid ejection head" in this disclosure. The inkjet device 200 may be an experimental device or an inkjet printing device used for printing.

[0082] The ejector 210 of the inkjet head 202 includes a nozzle 212, a pressure chamber 214, and a piezoelectric element 216. The nozzle 212 communicates with the pressure chamber 214 via a nozzle flow path 218. The pressure chamber 214 communicates with a supply-side common flow path 224 via an individual supply path 220.

[0083] The vibration plate 226 that constitutes the top surface of the pressure chamber 214 includes a conductive layer (not shown) that functions as a common electrode corresponding to the lower electrode of the piezoelectric element 216. The pressure chamber 214, the walls of the other flow path parts, and the vibration plate 226 can be made of silicon.

[0084] The material of the diaphragm 226 is not limited to silicon, and it may be made of a non-conductive material such as resin, etc. The diaphragm 226 itself may be made of a metal material such as stainless steel, and may serve as a diaphragm that also serves as a common electrode.

[0085] A piezoelectric unimorph actuator is formed by a structure in which a piezoelectric element 216 is laminated on a vibration plate 226. The piezoelectric element 216 is connected to a drive circuit 250 and is driven by a drive voltage supplied from the drive circuit 250. A drive voltage is applied to an individual electrode 228, which is the upper electrode of the piezoelectric element 216, to deform the piezoelectric body 230, which deflects the vibration plate 226 and changes the volume of the pressure chamber 214. A pressure change accompanying the change in volume of the pressure chamber 214 acts on the ink, causing the ink to be ejected from the nozzle 212.

[0086] When the piezoelectric element 216 returns to its original state after ejecting the ink, new ink is filled into the pressure chamber 214 from the supply-side common flow path 224 through the individual supply path 220. The inkjet head 202 may be provided with an ink recovery path (not shown) for recovering ink not used for ejection.

[0087] The shape of the pressure chamber 214 in a plan view is not particularly limited, and may be a rectangle, another polygon, a circle, an ellipse, etc. A cover plate 232 is provided above the individual electrode 228. The cover plate 232 is a member that ensures a movable space 234 for the piezoelectric element 216 and seals the periphery of the piezoelectric element 216.

[0088] A supply-side ink chamber (not shown) and a recovery-side ink chamber (not shown) are formed above the cover plate 232. The supply-side ink chamber is connected to the supply-side common flow path 224 via a communication path (not shown). The recovery-side ink chamber is connected to the recovery-side common flow path (not shown) via a communication path (not shown).

[0089] The information processing device 300, which controls the ejection operation of the inkjet head 202, includes a control unit 302, a waveform generating unit 304, an image processing unit 306, and a data storage unit 308. The information processing device 300 may include a driving circuit 250. The hardware configuration of the information processing device 300 may be similar to that of Fig. 7. The processing functions of each unit of the information processing device 300 can be realized by the processor 102 executing program instructions.

[0090] The information processing device 300 is connected to a camera 320. The camera 320 is placed at a position where it can capture an image of the flying state of the ink ejected from the nozzle 212. The control unit 302 controls the entire system including the inkjet head 202 and the camera 320. The waveform generating unit 304 can generate various types of driving waveforms DWj according to instructions from the control unit 302. For example, the waveform generating unit 304 can generate multiple driving waveforms DWj with different combinations of the values ​​of the 12 parameters described in FIG. 2. The subscript j represents an index for identifying multiple driving waveforms. For example, when generating 100 types of driving waveforms DWj, j is an integer from 1 to 100.

[0091] The driving circuit 250 supplies the driving voltage of the driving waveform DWj generated by the waveform generating unit 304 to the piezoelectric element 216. In this manner, the piezoelectric element 216 is driven, causing ink to be ejected from the nozzle 212. The camera 320 captures images of the flight state of the ink ejected from the nozzle 212 at regular time intervals. The control unit 302 controls the timing of image capture by the camera 320 in synchronization with the drive of the piezoelectric element 216. A group of time-series images captured by the camera 320 is sent to the image processing unit 306.

[0092] The image processor 306 performs necessary processing such as extraction of a region of interest and cropping on the acquired images to generate a time-series flight shape image group FSj(t) showing the flight shape of the ink. The subscript t indicates the time in the time series.

[0093] The control unit 302 associates (links) the driving waveforms DWj with the flight shape image groups FSj(t) and stores the driving waveforms DWj and the flight shape image groups FSj(t) in the data storage unit 308. In this way, a data set including a plurality of driving waveforms DWj and a plurality of corresponding flight shape image groups FSj(t) is created. A part or the whole of this data set is used as a learning data set. Note that such a data set is created for each combination of the ink used and the inkjet head 202.

[0094] How to create an autoencoder Fig. 12 is a block diagram showing a schematic functional configuration of an information processing device 170 that executes the creation process of an autoencoder. The hardware configuration of the information processing device 170 may be the same as the configuration described in Fig. 10. The processing functions of each unit of the information processing device 170 are realized by the processor 102 executing program instructions.

[0095] The information processing device 170 includes a waveform generating unit 172, an autoencoder 50, a loss calculating unit 174, a parameter update amount calculating unit 176, and a parameter update processing unit 178. The waveform generating unit 172 may have a similar configuration to the waveform generating unit 304 in FIG.

[0096] The autoencoder 50 includes an encoder unit 52 and a decoder unit 54. A driving waveform generated by a waveform generating unit 172 is input to the autoencoder 50, and features of the latent space are extracted by the encoder unit 52. The features of the latent space extracted by the encoder unit 52 are reconstructed by the decoder unit 54, and a reconstructed waveform is output. A loss calculation unit 174 calculates a loss indicating an error between the reconstructed waveform output by the autoencoder 50 and the original input driving waveform. A parameter update amount calculation unit 176 calculates an update amount of the parameters of the autoencoder 50 based on the calculated loss. A parameter update processing unit 178 updates the parameters of the autoencoder 50 according to the calculated update amount.

[0097] The parameters of the autoencoder 50 are optimized by updating the parameters of the autoencoder 50 multiple times using multiple drive waveforms so that a reconstructed waveform identical to the input drive waveform is obtained. The information processing device 170 functions as a machine learning system that executes machine learning processing to create the autoencoder 50. The processing function of the information processing device 170 may be incorporated into the information processing device 300 in FIG. 11.

[0098] How to create a predictive model: Example 1 Fig. 13 is a block diagram showing a schematic functional configuration of an information processing device 400 that executes a process of creating a prediction model. The hardware configuration of the information processing device 400 may be the same as the configuration described in Fig. 10. The information processing device 400 includes a data storage unit 402, a data acquisition unit 404, an autoencoder 50, an image processing unit 408, a prediction model 410, and an optimizer 412.

[0099] A data set TDS1 including a plurality of data sets each associated with a drive waveform TDWj and a corresponding flight shape TFSj is stored in the data storage unit 402. The drive waveform TDWj and the corresponding flight shape TFSj may be a drive waveform DWj and a time-series flight shape image group FSj(t) collected by the method described in FIG.

[0100] The data acquisition unit 404 acquires a data set including the drive waveform TDWj and the flight shape TFSj from the data storage unit 402. The drive waveform TDWj acquired via the data acquisition unit 404 is input to the autoencoder 50. The flight shape TFSj acquired via the data acquisition unit 404 is input to the image processing unit 408.

[0101] The autoencoder 50 is the trained autoencoder described in Fig. 5. Instead of the autoencoder 50, only the encoder unit 52 in the trained autoencoder 50 may be used.

[0102] The autoencoder 50 compresses the input driving waveform TDWj into a latent space and outputs waveform features TWFj represented by vectors in the latent space. The waveform features TWFj are input to the prediction model 410.

[0103] The prediction model 410 is a machine learning model that receives the waveform features TWFj as input and outputs the predicted characteristics PFFj. For example, a Gaussian process regression model can be applied as the prediction model 410. In this case, the prediction model 410 outputs the average value and standard deviation of the predicted characteristics.

[0104] The prediction model 410 is actually a program, and together with the autoencoder 50, causes the computer to realize a function of predicting the behavior of the inkjet head 202. The predicted characteristic PFFj output by the prediction model 410 corresponds to the predicted result of flight prediction when the drive waveform TDWj is applied. The predicted characteristic PFFj is sent to an optimizer 412.

[0105] The image processing unit 408 extracts characteristics TFFj such as droplet speed, droplet volume, and the presence or absence of satellite droplets from the input flight shape TFSj. The characteristics TFFj extracted by image processing from the actual flight shape TFSj correspond to the correct characteristics obtained when the drive waveform TDWj is applied.

[0106] The optimizer 412 performs the following processes: comparing the predicted characteristic PFFj with the correct (actual) characteristic TFFj, calculating a loss indicating the error between them, calculating an update amount for the parameters of the prediction model 410 based on the loss, and updating the parameters of the prediction model 410 according to the calculated update amount. The parameters of the prediction model 410 are called model parameters.

[0107] The optimizer 412 updates the model parameters so that the predicted characteristics PFFj approach the correct characteristics TFFj.

[0108] By updating the model parameters multiple times using multiple data sets included in the data set TDS1, the model parameters of the prediction model 410 are optimized, and a prediction model 410 capable of highly accurate prediction is created. A corresponding prediction model 410 is created for each combination of the ink used and the inkjet head 202. The information processing device 400 functions as a machine learning system that executes machine learning processing to create the prediction model 410.

[0109] How to create a predictive model: Example 2 In the example of Figure 13, an example was described in which the dataset TDS1 was used as a learning dataset, but as shown in Figure 14, a data set of waveform features TWFj and characteristics TFFj may be created in advance based on the dataset TDS1, and a learning dataset TDS2 including the waveform features TWFj and characteristics TFFj may be constructed.

[0110] Fig. 14 is a block diagram showing a schematic functional configuration of an information processing device 420 that executes a process of creating a learning dataset TDS2. The hardware configuration of the information processing device 420 may be the same as the configuration described in Fig. 10. In Fig. 14, the same reference numerals are used to designate the same or similar configurations as those of the information processing device 400 shown in Fig. 13, and descriptions thereof will be omitted.

[0111] The information processing device 420 includes a data storage unit 422 that associates and stores the waveform features TWFj output from the autoencoder 50 with the characteristics TFFj extracted by the image processing of the image processing unit 408. The data storage unit 422 stores a data set TDS2 including a plurality of data sets of the waveform features TWFj and the characteristics TFFj corresponding to the waveform features TWFj.

[0112] The data storage unit 422 may be configured by a storage device different from or the same as the data storage unit 402. The processing function of the information processing device 420 may be incorporated into the information processing device 300 of FIG.

[0113] 15, the data set TDS1 and the data set TDS2 may be integrated and stored in the data storage unit 422 as a new data set TDS3.

[0114] Fig. 16 is a block diagram showing a schematic functional configuration of an information processing device 430 that executes a process of creating a prediction model 410 using the data set TDS2. The hardware configuration of the information processing device 430 may be similar to the configuration described in Fig. 10. In Fig. 16, the same reference numerals are used to designate the same or similar configurations as those of the information processing device 400 shown in Fig. 13, and descriptions thereof will be omitted.

[0115] The information processing device 430 includes a data storage unit 422 storing a data set TDS2, a data acquisition unit 432, a prediction model 410, and an optimizer 412. The data acquisition unit 432 acquires a data set including waveform features TWFj and characteristics TFFj from the data storage unit 422. The waveform features TWFj are input to the prediction model 410. The characteristics TFFj are sent to the optimizer 412. Other operations are similar to those in FIG. 13.

[0116] <<Driving waveform search method using trained prediction model 410>> Fig. 17 is a block diagram showing a schematic functional configuration of an information processing device 500 that executes a process of searching for a promising driving waveform using the trained autoencoder 50 and the prediction model 410 constructed according to the present embodiment. The hardware configuration of the information processing device 500 may be the same as the configuration described in Fig. 10. The processing functions of each unit of the information processing device 500 can be realized by the processor 102 executing program instructions.

[0117] The information processing device 500 includes a control unit 502, a waveform generating unit 504, an autoencoder 50, a prediction model 410, a characteristic evaluation unit 506, a drive waveform determination unit 508, and a storage unit 510. The control unit 502 controls the processing of each unit. The control unit 502 instructs the waveform generating unit 504 to generate an unknown drive waveform. The unknown drive waveform is a new drive waveform other than the drive waveform used when learning the prediction model 410 (i.e., other than the drive waveform used in the ejection experiment described in FIG. 11), and is a drive waveform with unknown ejection characteristics.

[0118] The waveform generating unit 504 generates a plurality of drive waveforms CDWk of various waveforms according to instructions from the control unit 502. The subscript k is an index that identifies the drive waveform. For example, when 400 types of drive waveforms are generated, k can be an integer from 1 to 400. The waveform generating unit 504 generates a new drive waveform CDWk by, for example, randomly changing the parameter values ​​of the drive waveforms.

[0119] The autoencoder 50 is the trained autoencoder described in Fig. 5. Instead of the autoencoder 50, only the encoder unit 52 in the trained autoencoder 50 may be used. The autoencoder 50 receives an input of a driving waveform CDWk, compresses the driving waveform CDWk into a latent space, and outputs a waveform feature FVk.

[0120] The prediction model 410 is a trained model created using the information processing device 400 described in Fig. 9. The prediction model 410 receives the waveform feature FVk and outputs the predicted characteristic PFCk. The information processing device 500 uses a combination of the autoencoder 50 and the prediction model 410 to perform forward prediction of the ejection characteristic from the drive waveform CDWk.

[0121] The characteristic evaluation unit 506 evaluates whether or not the predicted characteristic PFCk output from the prediction model 410 satisfies the conditions of a target characteristic. The characteristic evaluation unit 506 compares the predicted characteristic PFCk with a predetermined target value to determine whether or not the desired characteristic has been achieved.

[0122] The control unit 502 associates the drive waveform CDWk, the waveform feature FVk, and the predicted characteristic PFCk, and stores these data in the storage unit 510. In this manner, a set of data including a plurality of new drive waveforms CDWk (k=1, 2, . . .) and a plurality of corresponding waveform features FVk and predicted characteristics PFCk is stored in the storage unit 510.

[0123] The drive waveform determination unit 508 determines a promising drive waveform based on the evaluation result of the predicted characteristic PFCk by the characteristic evaluation unit 506. The drive waveform determination unit 508 may determine, as the optimal drive waveform, a drive waveform that makes the predicted characteristic PFCk satisfy a predetermined target value condition and achieves the best characteristics.

[0124] For example, the drive waveform determination unit 508 determines the optimal drive waveform based on an evaluation value for a characteristic characterized by at least one of the droplet speed, droplet volume, and the presence or absence of satellite droplets.

[0125] If the predicted characteristic PFCk does not satisfy the condition of the target value, the drive waveform may be excluded from the candidates and the data may not be stored in the storage unit 510.

[0126] In this way, a prediction model 410 corresponding to the combination of the ink and the inkjet head 202 to be used is used to create a drive waveform suitable for ejecting that ink.

[0127] About the programs that run computers A program that causes a computer to realize some or all of the processing functions of each of information processing devices 170, information processing device 300, information processing device 400, information processing device 420, information processing device 430, and information processing device 500 can be recorded on a computer-readable medium that is a non-transitory information storage medium such as an optical disk, a magnetic disk, a semiconductor memory, or other tangible object, and the program can be provided through this information storage medium.

[0128] In addition, instead of providing the program by storing it on such a tangible, non-transitory computer-readable medium, it is also possible to provide the program signal as a download service using a telecommunications line such as the Internet.

[0129] Furthermore, some or all of the processing functions of each of the above-mentioned devices may be realized by cloud computing, and may also be provided as SaaS (Software as a Service).

[0130] <Hardware configuration of each processing unit> The hardware structure of the processing units that execute various processes, such as the waveform generation unit 172, autoencoder 50, loss calculation unit 174, parameter update amount calculation unit 176, and parameter update processing unit 178 in the information processing device 170, the control unit 302, waveform generation unit 304, and image processing unit 306 in the information processing device 300, the data acquisition unit 404, autoencoder 406, image processing unit 408, prediction model 410, and optimizer 412 in the information processing device 400, and the control unit 502, waveform generation unit 504, characteristic evaluation unit 506, and drive waveform determination unit 508 in the information processing device 500, is, for example, various processors as shown below.

[0131] Various types of processors include CPUs, which are general-purpose processors that execute programs and function as various processing units, GPUs, programmable logic devices (PLDs), such as FPGAs (Field Programmable Gate Arrays), which are processors whose circuit configuration can be changed after manufacture, and dedicated electrical circuits, such as ASICs (Application Specific Integrated Circuits), which are processors with a circuit configuration designed specifically to execute specific processes.

[0132] A processing unit may be composed of one of these various processors, or may be composed of two or more processors of the same type or different types. For example, a processing unit may be composed of multiple FPGAs, or a combination of a CPU and an FPGA, or a combination of a CPU and a GPU. Also, multiple processing units may be composed of one processor. As an example of multiple processing units being composed of one processor, first, as represented by a computer such as a client or a server, there is a form in which one processor is composed of a combination of one or more CPUs and software, and this processor functions as multiple processing units. Second, as represented by a system on chip (SoC), there is a form in which a processor that realizes the functions of the entire system including multiple processing units is used in one IC (Integrated Circuit) chip. In this way, the various processing units are composed of one or more of the above various processors as a hardware structure.

[0133] Furthermore, the hardware structure of these various processors is, more specifically, an electric circuit that combines circuit elements such as semiconductor elements.

[0134] Advantages of the embodiment According to the above-described embodiment, the following effects can be obtained.

[0135] [1] For a combination of ink and inkjet head used, machine learning is used to automatically learn the relationship between the drive waveform and the characteristics, using characteristics extracted from the actual flight shape, to create a prediction model 410 that can accurately predict the characteristics for unknown drive waveforms.

[0136] [2] Even engineers without expertise in fluid mechanics and numerical calculations can create a high-performance predictive model 410 by repeatedly collecting data through experiments and using images of the actual flight shape.

[0137] [3] By automating the process of searching for drive waveforms using the trained predictive model 410, even engineers who do not have expertise in creating drive waveforms can create drive waveforms that can achieve a certain level of characteristics.

[0138] [4] The user can select a drive waveform that suits their purpose. For example, if the quality of a solid print is important, the amount of droplets can be prioritized and satellite droplets can be tolerated. By setting the conditions (target values) that must be met for various characteristics according to the purpose, the user can create a drive waveform that matches those conditions.

[0139] [5] It can be used to evaluate inks. Until now, it has been difficult to rank ink types based on the quality of optimization of the drive waveform, but the method of this embodiment makes it possible to compare the ejection suitability of inks based on indicators such as the number of candidate waveforms that can obtain favorable flight characteristics.

[0140] [6] It is possible to find drive waveforms that can achieve high-quality characteristics that would be difficult to achieve by conventional engineers. In other words, it is possible to search for completely unknown drive waveforms that would be difficult to find by hand.

[0141] [7] Promising drive waveforms can be created more efficiently and in a shorter time than if engineers were to create the drive waveforms themselves.

[0142] [8] It is possible to automatically and efficiently find drive waveforms with good characteristics from the countless drive waveforms available.

[0143] [9] When a drive waveform is searched manually, only drive waveforms with typical properties that are based on conventional knowledge can be found. However, when the automatic search of this embodiment is used, it is possible to suggest unexpected drive waveforms as candidates that a human would not try.

[0144] Variation 1 In the above embodiment, an example was described in which a time-series image group taken at a fixed time interval is used as data relating to the actual flight shape, but instead of this time-series image group, it is also possible to use, for example, a single image taken of the ink flight state a predetermined time after the application of the drive waveform. In this case, it is desirable to set the predetermined time so that characteristics such as the ink droplet speed, droplet volume, length of the liquid column, and the presence or absence of satellite droplets can be identified from the position of the ink shown in a single image taken at that timing.

[0145] As described in the embodiment, by using a group of two or more time-series images taken at different times, it is possible to grasp characteristics such as droplet speed more accurately, and a prediction model 410 with high prediction accuracy can be created.

[0146] Variation 2 In the above embodiment, an example of a machine learning model that receives input of coordinates in the latent space of a drive waveform and outputs predicted characteristics has been described, but the prediction model is not limited to this example. For example, a prediction model that receives input of a drive waveform and outputs a predicted flight shape may be constructed by machine learning using the data set TDS1 of FIG. 13. In this case, the model parameters of the machine learning model are optimized while evaluating the error between the predicted flight shape output by the prediction model and the actual (correct) flight shape.

[0147] Then, for the predicted flight shape output by this trained prediction model, evaluation values ​​such as droplet volume, droplet speed, and the presence or absence of satellite droplets may be calculated, and an appropriate drive waveform may be determined based on the evaluation values.

[0148] <Device application examples> In the above embodiment, an example of an inkjet device used for inkjet printing has been described, but the scope of application of the present invention is not limited to this example. Regardless of the type and purpose of the liquid used, the technology disclosed herein can be applied to devices that eject liquid using a piezoelectric liquid ejection head. For example, the technology can be widely applied to liquid ejection devices that draw various shapes and patterns using liquid functional materials (collectively referred to as "liquids"), such as wiring drawing devices that draw wiring patterns for electronic circuits, manufacturing devices for various devices, resist printing devices that use resin liquid as a functional liquid for ejection, color filter manufacturing devices, and microstructure forming devices that form microstructures using materials for material deposition.

[0149] "others" The present disclosure is not limited to the above-described embodiment, and various modifications are possible without departing from the spirit and scope of the technical idea of ​​the present disclosure. [Explanation of symbols]

[0150] 20 Driving Waveform 22 Preliminary vibration pulse 24 Discharge Pulse 26 Reverberation suppression pulse 50 Autoencoder 52 Encoder section 54 Decoder section 100 Information processing device 102 processors 104 Computer-readable medium 106 Communication Interface 108 Input / Output Interface 110 Bus 112 Memory 114 Storage 152 Input Device 154 Display device 170 Information processing equipment 172 Waveform generator 174 Loss Calculation Department 176 Parameter update amount calculation unit 178 Parameter update processing section 200 Inkjet device 202 Inkjet head 210 Ejector 212 Nozzle 214 Pressure Chamber 216 Piezoelectric element 218 Nozzle flow path 220 Individual supply route 224 Supply side common flow path 226 Diaphragm 228 individual electrodes 230 Piezoelectric 232 Cover Plate 234 Movable space 250 Drive circuit 300 Information processing device 302 Control section 304 Waveform generator 306 Image Processing Unit 308 Data Storage Department 320 Camera 400 Information processing device 402 Data storage unit 404 Data Acquisition Department 406 Model Parameter Update Unit 408 Image Processing Unit 410 Predictive Models 412 Optimizer 420 Information processing equipment 422 Data Storage Unit 430 Information processing equipment 432 Data Acquisition Department 500 Information processing equipment 502 Control section 504 Waveform generator 506 Characterization Department 508 Drive Waveform Determination Unit 510 Storage section CDWk drive waveform DWj drive waveform FVk waveform characteristics F4A left diagram F4B right FSj(t) Flight shape image set PFCk predicted characteristics PFFj predicted properties WFi High Dimension Vector RWFi reconstruction vector TDWj drive waveform TFFj characteristics TFSj flight shape TDS1, TDS2, TDS3 Data Sets S1~S8 Steps for creating driving waveforms

Claims

1. A method for creating a driving waveform used to drive a piezoelectric element of a liquid ejection head including a piezoelectric element, comprising the steps of: using a machine learning model that has been trained by machine learning using data on the liquid to be discharged by the liquid discharge head and the actual flight shape of the liquid when each of a plurality of drive waveforms is applied to the piezoelectric element using the liquid discharge head, one or more processors perform a flight prediction of the liquid when an unknown drive waveform is input; determining a drive waveform suitable for ejecting the liquid based on the flight prediction; A driving waveform generating method including the steps of:

2. The parameters of the drive waveform include at least one of a pulse width, a slope, a pulse height, and a pulse interval. A method for generating a drive waveform as claimed in claim 1.

3. The learning phase of the machine learning model includes compressing each of the plurality of drive waveforms into a latent space having a dimension smaller than the dimension of the drive waveform.

2. The driving waveform generating method according to claim 1.

4. The drive waveform is input to an autoencoder and converted into coordinates in the latent space.

4. The driving waveform generating method according to claim 3.

5. In the learning phase, the machine learning model is trained to predict the evaluation value when the drive waveform is applied, using a correspondence relationship between coordinates in the latent space of each of the plurality of drive waveforms and an evaluation value based on the actual flight shape.

4. The driving waveform generating method according to claim 3.

6. The data on the actual flight shape includes the evaluation value indicating a characteristic extracted from an image of the actual flight shape.

6. The driving waveform generating method according to claim 5.

7. the evaluation value includes at least one of a droplet speed, a droplet volume, and a value indicating the presence or absence of satellite droplets of the liquid ejected from the liquid ejection head; 6. The driving waveform generating method according to claim 5.

8. The flight prediction includes a prediction of the evaluation value, the one or more processors, generating one or more unknown drive waveforms that are different from the plurality of drive waveforms; Calculating coordinates in the latent space from the unknown drive waveform; Calculating the evaluation value predicted from coordinates in the latent space of the unknown drive waveform using the machine learning model; comparing the evaluation value calculated using the machine learning model with a target value to determine a drive waveform that satisfies the target value; 6. The driving waveform generating method according to claim 5.

9. The machine learning model is a model that outputs an average value and a standard deviation of the evaluation value predicted from coordinates in the latent space.

6. The driving waveform generating method according to claim 5.

10. the one or more processors, generating one or more unknown drive waveforms that are different from the plurality of drive waveforms; Calculating coordinates in the latent space from the unknown drive waveform; Calculating an average value and a standard deviation of the evaluation value predicted from the coordinates in the latent space using the machine learning model; Calculating the probability that the evaluation value exceeds a target value from an average value and a standard deviation of the evaluation value calculated using the machine learning model; determining a drive waveform having a high probability of exceeding the target value as an appropriate drive waveform; 10. The driving waveform generating method according to claim 9.

11. the one or more processors: Calculating coordinates in the latent space from the unknown drive waveform using an autoencoder; 9. The driving waveform generating method according to claim 8.

12. the one or more processors: randomly extracting parameter values ​​of the drive waveform based on a uniform distribution to generate a plurality of unknown drive waveforms different from the plurality of drive waveforms, and performing the flight prediction using the machine learning model for each of the drive waveforms; 2. The driving waveform generating method according to claim 1.

13. the one or more processors: In the case where a plurality of unknown drive waveforms different from the plurality of drive waveforms are generated by randomly extracting values ​​of parameters of the drive waveform based on a uniform distribution, a variogram analysis is performed in advance to clarify a relationship between the distance in the latent space and the variance of the evaluation value, and a step size for searching the drive waveform is set based on the variogram analysis; 6. The driving waveform generating method according to claim 5.

14. the step size of the search is set to be equal to or larger than a distance at which the distance in the latent space and the variance of the evaluation value are uncorrelated based on the variogram analysis; The driving waveform generating method according to claim 13.

15. An information processing device for executing the driving waveform generating method according to any one of claims 1 to 14, the one or more processors; and one or more storage devices in which the machine learning model is stored; An information processing device comprising:

16. A program that causes a computer to execute the drive waveform generating method according to claim 1 .

Images