Inkjet printing droplet volume mean regulation method for large number of nozzles

Abstract

The present application belongs to the field of inkjet printing manufacturing, and particularly relates to a method for regulating the mean value of droplet volume of a large number of nozzles in inkjet printing, which comprises the following steps: setting initial waveform parameter vectors of a plurality of nozzles and driving all the nozzles to print, collecting the droplet volume of each nozzle; calculating the droplet volume deviation of each nozzle relative to the target mean volume, judging whether the average value of the droplet volume deviation of all the nozzles is less than the precision threshold value, if yes, the regulation is completed, if not, inputting the droplet volume deviation of each nozzle of each nozzle and the current waveform parameter vector of the nozzle into the pre-constructed waveform parameter feedback control model of the nozzle to obtain the control increment of each waveform parameter, adding the control increment of each waveform parameter to the value of each corresponding parameter in the current waveform parameter vector to obtain a new waveform parameter vector, and re-executing the driving printing. The present application can efficiently realize that the mean value of the droplet volume of a large number of nozzles in multi-nozzle printing meets the precision requirement of the target mean volume.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of inkjet printing manufacturing, and more particularly relates to a method for regulating the mean value of droplet volume of inkjet printing of a large number of nozzles. BACKGROUND

[0002] The inkjet printing technology has the advantages of simple process, low cost and suitability for large-area manufacturing, and has become an additive manufacturing technology for mass production of large-area electronic devices. When printing large-area electronic devices, multiple printheads are often used in parallel. Each printhead has hundreds or thousands of nozzles, so a large number of nozzles participate in the ejection of ink material into the fixed position of the substrate during the printing process.

[0003] Each nozzle of each printhead can be independently switched on or off, but all nozzles of the same printhead can only share one ejection driving voltage waveform. However, due to manufacturing errors, ink supply distribution, surface wettability and other problems, the droplet volumes ejected by multiple nozzles under the same driving waveform will be different. Therefore, in order to ensure the accuracy of inkjet printing manufacturing of a large number of nozzles, the driving waveform parameters of all printheads need to be regulated so that the droplet volumes of a large number of nozzles participating in the ejection are distributed around the target volume. However, the prior art can only regulate the ejection droplet of a single nozzle, and does not consider the difference in droplet volume of multiple nozzles under the same driving waveform. It only relies on regulating the droplet volume of a certain nozzle to meet the requirements, and cannot meet the requirement that all nozzle droplets reach the target volume. Therefore, there is an urgent need in the art to further improve and improve the prior art to meet the increasingly high process requirements. SUMMARY

[0004] In view of the defects and improvement needs of the prior art, the present application provides a method for regulating the mean value of droplet volume of inkjet printing of a large number of nozzles, which aims to solve the technical problem that the mean value of droplet volume of a large number of nozzles participating in printing does not meet the accuracy requirements of the target volume.

[0005] To achieve the above-mentioned purpose, according to one aspect of the present application, a method for regulating the mean value of droplet volume of inkjet printing of a large number of nozzles is provided, comprising:

[0006] Randomly setting the initial waveform parameter vector of n printheads participating in printing, and driving all printheads to print ink, and collecting the droplet volume of each nozzle;

[0007] The droplet volume deviation of each nozzle relative to the target mean volume is calculated to determine whether the average of the droplet volume deviations of all nozzles is less than a preset precision threshold. If yes, the regulation is completed and the process is ended. If no, the droplet volume deviation of each nozzle of each print head and the current waveform parameter vector of the print head are input into the pre-constructed waveform parameter feedback control model of the print head to obtain the control increment of each waveform parameter. The control increment of each waveform parameter is added to the value of each corresponding parameter in the current waveform parameter vector to obtain a new waveform parameter vector, and the driving printing is re-executed.

[0008] Further, the waveform parameter feedback control model is constructed in the following manner:

[0009] A set of driving voltage waveform parameter vectors is constructed. Each waveform parameter vector in the set is used to drive each print head to collect the droplet volume of each nozzle under each waveform parameter vector. A state set of each print head including two dimensions of the waveform parameter vector and the corresponding droplet volume of each nozzle is constructed to train a jet data model of the print head for predicting the droplet volume of each nozzle.

[0010] A strategy network of each print head and m evaluation networks and m target networks corresponding to the m nozzles of the print head are initialized. The input of the strategy network is the current waveform parameter vector of the print head and the current droplet volume deviation of the m nozzles relative to the target mean volume, and the output is the control increment of each waveform parameter. The output is added to each corresponding parameter in the current waveform parameter vector as a new waveform parameter vector to be fed back to the jet data model of the print head to predict the droplet volume of each nozzle in the next state. Under the assistance of the m evaluation networks and the m target networks, the strategy network of the print head and the corresponding jet data model are interactively learned to train the strategy network as the waveform parameter feedback control model of the print head.

[0011] Further, the implementation manner of the interactive learning is as follows:

[0012] S1, start an epoch, randomly set a target mean volume V in the adjustable volume range t , and randomly set the same initial waveform parameter vector for each print head;

[0013] S2, based on the jet data model of each print head and the current waveform parameter vector, obtain the droplet volume of each nozzle in the print head, calculate the droplet volume deviation of each nozzle relative to the target mean volume under the driving of the current waveform parameter vector, and construct the current state of the print head including the current waveform parameter vector and the current droplet volume deviation of each nozzle.

[0014] S3, input the current state of each nozzle into the strategy network of the nozzle to obtain a control increment of each waveform parameter used to drive the nozzle, add each waveform parameter in the current waveform parameter vector to the corresponding control increment to obtain a new waveform parameter vector of the nozzle, repeat S2 once to obtain a next state of the nozzle; calculate a reward value of each nozzle in the next state, and an identifier d used to represent whether an average deviation of droplet volumes of all nozzles in the next state from a target average volume reaches a precision threshold value; splice the current state, the control increment, the next state, and the reward value of all nozzles in the next state and the identifier d of each nozzle to form a sample;

[0015] S4, take the next state of each nozzle as a new current state of the nozzle, and determine whether a current sample quantity reaches a threshold value, if yes, perform an iterative training of the strategy network, m evaluation networks and m target networks of the nozzle based on the current sample, and perform S5; if no, directly perform S5;

[0016] S5, determine whether the identifier d meets the requirement in a current epoch, if yes, if a judgment epoch quantity reaches a threshold value, end, if the judgment epoch quantity does not reach the threshold value, re-perform S1; if no, re-perform S3.

[0017] Further, the identifier d is expressed as:

[0018]

[0019] In the formula, Δμ represents an average deviation of droplet volumes of all nozzles of each nozzle from a target average volume, and ξ represents the precision threshold value.

[0020] Further, in each iterative training, a ratio of a droplet volume deviation of each nozzle to a sum of droplet volume deviations of all nozzles in the nozzle is taken as a weight of an output of the evaluation network of the nozzle, weighted summation is performed on outputs of the m evaluation networks to construct a loss of the strategy network, parameters of the strategy network are updated; a target network of each nozzle is used to construct a target output of the evaluation network of the nozzle, a difference between the target output and an actual output of the evaluation network is taken to construct a loss of the evaluation network, parameters of the evaluation network of each nozzle are synchronously updated; parameters of the target network of each nozzle are soft-updated by using current parameters of the evaluation network of the nozzle after a preset number of iterations.

[0021] Further, the loss of the strategy network of each nozzle is specifically:

[0022]

[0023] In the formula, φ i is a parameter of the strategy network π of the nozzle i, N bV t V′ ijk Q(s ik ,a ik |θ ij ) represents the evaluation network output of the jth nozzle of the i th nozzle head, θ ij represents the evaluation network parameter of the jth nozzle of the i th nozzle head, s ik and a ik are the inputs of the evaluation network of the jth nozzle of the i th nozzle head, s ik represents the current state of the i th nozzle head in the k th sample, a ik represents the control increment of each waveform parameter of the i th nozzle head in the k th sample.

[0024] Further, the loss of the evaluation network of each nozzle is specifically:

[0025]

[0026] In the formula, N b represents the sample size required for one iteration of training; γ is a discount coefficient, which is selected from 0 to 1; V′ ik represents the next state of the i th nozzle head in the k th sample, a′ ik represents the control increment of each waveform parameter calculated by the policy network based on the next state; θ ij represents the evaluation network parameter of the jth nozzle of the i th nozzle head; s ik represents the current state of the i th nozzle head in the k th sample, a ik represents the control increment of the i th nozzle head in the k th sample;

[0027] r ij represents the reward value of the jth nozzle of the i th nozzle head in the next state, which is calculated in the following manner: the deviation of the average volume of all nozzles of the nozzle head from the target average volume is multiplied by a reward coefficient to obtain the reward of the nozzle head, and the ratio of the deviation of the droplet volume of each nozzle from the target average volume to the total deviation of the droplet volume of all nozzles is taken as the distribution coefficient of the reward of the nozzle head to the reward of the nozzle, and the reward of the nozzle is obtained by distributing the reward of the nozzle head.

[0028] Further, the reward value r ij of the jth nozzle of the i th nozzle head is calculated in the following manner:

[0029]

[0030] In the formula, η is a reward coefficient; μ iV is the average value of the droplet volume of all the nozzles of the nozzle i in the next state t V is the target average volume corresponding to the current epoch ij V is the droplet volume of the jth nozzle of the nozzle i in the next state.

[0031] Further, the reward coefficient η is in the range of [1, 100].

[0032] The application also provides a computer readable storage medium comprising a stored computer program, wherein the computer program, when executed by a processor, controls a device in which the storage medium is located to perform the above-mentioned droplet volume average value regulation method for a large number of nozzles of an inkjet printer.

[0033] Overall, the above technical solutions conceived by the application can achieve the following beneficial effects:

[0034] (1) The application detects the droplet volume of multiple nozzles of multiple nozzles, and compares the feedback with the target average volume, so as to utilize the volume difference of each nozzle, combine the constructed waveform parameter feedback control model of the corresponding nozzle, regulate the driving waveform parameter of the multiple nozzles, and thus complete the closed-loop regulation of the waveform parameter of the multiple nozzles, which can effectively cope with the droplet volume deviation caused by environmental factors. Through online regulation of the driving waveform parameter, the droplet volume distribution of the large number of nozzles of the multiple nozzles can be near the target volume, and the average volume requirement of the inkjet printing manufacturing process can be ensured.

[0035] (2) The application proposes a construction method of the waveform parameter feedback control model of each nozzle, specifically, a strategy network is established for each nozzle, and m evaluation networks and m target networks corresponding to m nozzles, under the assistance of the m evaluation networks and the m target networks, the strategy network and the corresponding nozzle ejection data model interact to learn, so as to train the strategy network as the waveform parameter feedback control model of the nozzle, the feedback control model is established based on data, the control strategy is extracted from industrial data, and the ejection trend of all nozzles participates in the construction process of the nozzle regulation strategy, so that the control model can make the optimal strategy to improve the overall distribution of the nozzle volume.

[0036] (3) The application involves the calculation of the reward value of each nozzle when calculating the loss of the evaluation network of each nozzle. In the calculation of the reward value of each nozzle, the application proposes the distribution of the reward value of the entire nozzle to the m nozzles, which comprehensively considers the spraying relationship of multiple nozzles on the same nozzle. The loss of the evaluation network of each nozzle is used for parameter adjustment of the evaluation network, and the output of the parameter-adjusted evaluation network finally acts on the parameter update of the policy network. The parameter-updated policy network is used to adjust the driving voltage waveform of the corresponding nozzle. This way reduces the difference in droplet volume of the numerous nozzles sharing the waveform parameters on the same nozzle, and provides a basis for the mean value regulation of the droplet volume of the large number of nozzles. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 A schematic block diagram of a droplet volume mean value regulation method for a large number of nozzles of an inkjet printer is provided for the embodiments of the application.

[0038] Figure 2 A process diagram of the driving voltage parameter and large number of nozzle droplet volume data collection process of a multi-nozzle is provided for the embodiments of the application.

[0039] Figure 3 A structure diagram of each nozzle policy network is provided for the embodiments of the application.

[0040] Figure 4 A structure diagram of each nozzle evaluation network and target network is provided for the embodiments of the application.

[0041] Figure 5 A flowchart of the interactive learning process of the policy network and the nozzle data model is provided for the embodiments of the application.

[0042] Figure 6 A training framework diagram of each nozzle policy network is provided for the embodiments of the application.

[0043] Figure 7 A droplet volume mean value regulation method for a large number of nozzles of an inkjet printer is provided for the embodiments of the application. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application. In addition, the technical features involved in each embodiment of the application described below can be combined with each other as long as they do not conflict with each other.

[0045] Embodiment one

[0046] A droplet volume mean value regulation method for a large number of nozzles of an inkjet printer is provided for the embodiments of the application.Figure 1 Shown, including:

[0047] Randomly set the initial waveform parameter vectors of the n nozzles involved in the printing, drive all the nozzles to print ink, and collect the droplet volume of each nozzle;

[0048] Calculate the droplet volume deviation of each nozzle relative to the target mean volume to determine whether the average value of the droplet volume deviation of all nozzles is less than the preset accuracy threshold. If so, complete the control and end the process. If not, input the droplet volume deviation of each nozzle of each nozzle and the current waveform parameter vector of the nozzle into the pre-constructed waveform parameter feedback control model of the nozzle to obtain the control increment of each waveform parameter, add the control increment of each waveform parameter to the corresponding parameter value in the current waveform parameter vector to obtain a new waveform parameter vector, and re-execute the drive printing.

[0049] It should be noted that the initial waveform parameter vectors of the n nozzles can be the same or different. Figure 2 As shown in the figure, the relationship between the droplet spreading area and volume is calibrated, and the actual volume V0 of the droplet is obtained by weighing. Then, its landing point image is collected and the spreading area S0 is calculated. Therefore, when measuring the volume of multi-nozzle droplets, the dot matrix image is first collected, and then the image is segmented to obtain the spreading area s of each nozzle droplet. ij , from which the droplet volume of the nozzle is calculated as follows: This method of trial printing dot arrays with multiple nozzles quickly collects the landing point images of droplets from a large number of nozzles, and then calculates the droplet volume of each nozzle through image segmentation, thereby greatly improving the detection efficiency of the droplet volume of a large number of nozzles.

[0050] This embodiment detects droplet volumes from multiple nozzles across a multi-nozzle system and compares the resulting droplet volumes with a target mean volume. This differential volume across each nozzle is then used to adjust the drive waveform parameters for the multi-nozzle system, achieving closed-loop control of the waveform parameters. This effectively addresses droplet volume deviations caused by environmental factors. By online control of the drive waveform parameters, the droplet volumes of a large number of nozzles across a multi-nozzle system can be distributed near the target volume, ensuring the average volume requirements of the inkjet printing process.

[0051] That is, the present invention detects the droplet volume of multiple nozzles of multiple printheads and obtains the volume deviation of each nozzle by comparing the feedback with the target mean volume, and then calculates the waveform parameter control increment by the feedback control model in combination with the current waveform parameters of the printhead for closed-loop regulation. This can make the droplet volume of a large number of nozzles of multiple printheads distributed near the target volume, ensure the average volume requirement of the inkjet printing manufacturing process, and effectively deal with the droplet volume deviation caused by environmental factors.

[0052] As a preferred embodiment, the waveform parameter feedback control model is constructed in the following way:

[0053] A set of driving voltage waveform parameter vectors P = {X 1 ,X 2 ,...,X k ,...} is constructed, where is a waveform parameter vector; each nozzle droplet volume under each waveform parameter vector is collected by driving each nozzle with each waveform parameter vector in the set, and a state set of each nozzle including two dimensions of waveform parameter vectors and their corresponding nozzle droplet volumes is constructed to train a jet data model of the nozzle for predicting its nozzle droplet volume;

[0054] A strategy network of each nozzle and m evaluation networks and m target networks corresponding to the m nozzles of the nozzle are initialized; the input of the strategy network is the current waveform parameter vector of the nozzle and the current droplet volume deviation of the m nozzles relative to the target mean volume, and the output is the control increment of each waveform parameter, which is added to the corresponding parameter in the current waveform parameter vector as a new waveform parameter vector to be fed back to the jet data model of the nozzle to predict the next state nozzle droplet volume, that is, to simulate the real environment by using the jet data model of the nozzle; under the assistance of the m evaluation networks and the m target networks, the strategy network of the nozzle and the corresponding jet data model are interactively learned to train the strategy network as the waveform parameter feedback control model of the nozzle.

[0055] Regarding the construction of the set of driving voltage waveform parameter vectors, for example, the peak voltage up and the duration td of the driving voltage waveform parameter of each nozzle are selected, and their value sequences are U = {U min :U step :U max}, where U min and U max are the upper and lower limits of the voltage parameter determined by the nozzle, and U step is the value step length which can be set, such as 0.5; T = {T min :T step :T max}, where T min and T max are the upper and lower limits of the time parameter determined by the nozzle, and T step is the value step length which can be set, such as 0.2. Therefore, the set of waveform parameters to be tested is P = {(u p ,t d )|u p ∈U,t d ∈T}.

[0056] Regarding driving each printhead with every waveform parameter vector in the set, for example, driving n printheads with waveform parameters in set P, for example, X i = X k where X i is the driving waveform parameter of printhead i, printing a dot pattern on a test print substrate, where each drop represents the ejection status of a nozzle, a detection camera can be used to capture the dot pattern image to collect the drop volume V ij of all nozzles, then automatically wiping the test print substrate for the next round of dot pattern printing.

[0057] Regarding the state set of each printhead, the state set D established using the waveform parameter vector and the drop volume of each nozzle is as follows:

[0058] D = {(X1,..., X n , V 11 ,..., V nm )...}

[0059] The ejection data model M1,..., M n of each printhead is established using a neural network, where the input of M i is the waveform parameter vector X i of printhead i, and the output is the drop volume V i1 ,..., V im of each nozzle of printhead i. Based on the state set D, the network parameters of the ejection data model of each printhead can be trained using the method of stochastic gradient descent.

[0060] Regarding the strategy network, for example, the structure diagram of the strategy network of each printhead as shown in Figure 3 In this embodiment, a strategy network π i of printhead i, whose input is the current waveform parameter X i of the printhead and the drop volume deviation ΔV i1 ,..., ΔV im of the m nozzles of the printhead, and whose output is the control increment ΔX i of the waveform parameter. Exemplarily, the number of neurons in each layer (including the input layer and the output layer) in the strategy network is (2+m)-128-256-128-2.

[0061] Regarding the evaluation network and the target network, for example, the structure diagram of the evaluation network and the target network of each nozzle as shown in Figure 4 The nozzle j of printhead i corresponds to an evaluation network Q ij to be trained and a target network not participating in training. They have the same structure, and the input is the previous waveform parameter X i and the drop volume deviation ΔV i1..., ΔV im and control increment ΔX i , output is evaluation value q ij (the output of the evaluation network is represented as q ij , and the output of the target network is represented as , and the number of neurons in each layer (including the input layer and the output layer) in the evaluation network and the target network is (4+m)-128-256-128-1, respectively.

[0062] The above-mentioned interactive learning process is a cyclic process of building samples and training, which can be an embodiment, as shown in Figure 5 The implementation of the above-mentioned interactive learning is as follows:

[0063] S1, start an epoch, randomly set the target mean volume V t in the adjustable volume range, and randomly set the same initial waveform parameter vector for each nozzle;

[0064] S2, based on the spray data model of each nozzle and the current waveform parameter vector, obtain the droplet volume of each nozzle in the nozzle, calculate the droplet volume deviation of each nozzle under the driving of the current waveform parameter vector relative to the target mean volume, and build the current state of the nozzle including the current waveform parameter vector and the current droplet volume deviation of each nozzle.

[0065] S3, input the current state of each nozzle into the strategy network of the nozzle to obtain the control increment of each waveform parameter for driving the nozzle, add each waveform parameter in the current waveform parameter vector to the corresponding control increment to obtain the new waveform parameter vector of the nozzle, repeat S2 once to obtain the next state of the nozzle; calculate the reward value of each nozzle in the next state, and the identifier d representing whether the average deviation of the droplet volume of all nozzles in the next state relative to the target mean volume reaches the accuracy threshold; splice the current state, the control increment, the next state, and the reward value and identifier d of all nozzles in the next state to form a sample;

[0066] S4, take the next state of each nozzle as the new current state of the nozzle, and judge whether the current sample quantity reaches the threshold value, if yes, based on the current sample, perform one iteration training on the strategy network, m evaluation networks and m target networks of the nozzle, and execute S5; if not, directly execute S5;

[0067] S5, judge whether the identifier d meets the requirements in the current epoch, if yes, if the number of epochs judged reaches the threshold value, end, if the number of epochs judged does not reach the threshold value, re-execute S1; if not, re-execute S3.

[0068] The identifier d represents whether the next state after the current regulation meets the accuracy requirement of the volume mean target, and can be represented as d = | Δμ | < ξ as an embodiment.

[0069]

[0070] In the formula, Δμ represents the average deviation of the droplet volume of all nozzles of each nozzle from the target mean volume, and ξ represents the accuracy threshold.

[0071] Figure 6 As a preferred embodiment, in each iteration training, the droplet volume deviation of each nozzle and the sum of the droplet volume deviations of all nozzles in the nozzle are used as the weight of the evaluation network output of the nozzle to weight and sum the outputs of the m evaluation networks, the loss of the strategy network is constructed to update the parameters of the strategy network, so that the waveform parameter increment output by the strategy network to drive the nozzle to print can make the overall evaluation of each evaluation network to each nozzle high; in addition, the target output of the evaluation network of each nozzle is constructed by the target network of the nozzle, and the loss of the evaluation network is constructed by the difference between the target output and the actual output of the evaluation network to update the parameters of the evaluation network of each nozzle synchronously; the parameters of the target network of each nozzle are soft updated by the current parameters of the evaluation network of the nozzle after a preset number of iterations.

[0072] As a preferred embodiment, the loss of the strategy network of each nozzle is specifically:

[0073]

[0074] In the formula, φ i is the parameter of the strategy network π of the nozzle i, N b is the sample size required for one iteration training, V t is the target mean volume corresponding to the current epoch, V ij ' k is the droplet volume of the jth nozzle of the nozzle i in the next state; Q(s ik , a ik |θ ij ) represents the output of the evaluation network of the jth nozzle of the nozzle i, θ ij represents the parameter of the evaluation network of the jth nozzle of the nozzle i, s ik and a ik are the inputs of the evaluation network of the jth nozzle of the nozzle i, s ik represents the current state of the nozzle i in the kth sample, a ik represents the control increment of each waveform parameter of the nozzle i in the kth sample.

[0075] Therefore, the policy network parameter of the nozzle i is updated as φ i = φ i - λ π ▽J π (φ i ), wherein λ π is a learning rate, preferably in the range of (0, 0.1), and can be preferably 0.001.

[0076] As a preferred embodiment, the target of the evaluation network is approximated based on the reward value and the output of the target network in the next state, and the loss of the evaluation network of each nozzle is based on the deviation of the output of the target network and the evaluation network of each nozzle, specifically, the loss of the evaluation network of each nozzle is:

[0077]

[0078] In the formula, N b represents the sample size required for one iteration of training; γ is a discount factor, and is in the range of 0 to 1, and can be preferably 0.95; is the target network parameter of the jth nozzle of the nozzle i, s' ik is the next state in the kth sample of the nozzle i, a' ik is the control increment of each waveform parameter calculated by the policy network based on the next state; θ ij is the evaluation network parameter of the jth nozzle of the nozzle i; s ik is the current state in the kth sample of the nozzle i, a ik is the control increment in the kth sample of the nozzle i;

[0079] r ij represents the reward value of the jth nozzle of the nozzle i in the next state, which refers to the expected return that the evaluation network of each nozzle can obtain after the control in the current state, and is used to update the parameters of the evaluation network Q ij , which is updated in the following manner: the deviation of the average volume of all nozzles of the nozzle from the target average volume is multiplied by a reward coefficient to obtain the reward of the nozzle, and the ratio of the droplet volume deviation of each nozzle to the total droplet volume deviation of all nozzles is taken as the distribution coefficient of the reward of the nozzle to the reward of the nozzle, and the reward of the nozzle is obtained by distributing the reward of the nozzle.

[0080] The evaluation network parameter of the nozzle is updated as θ ij = θ ij - λ Q ▽J Q (θ ij ), wherein λ Q is a learning rate, preferably in the range of (0, 0.1); and the update of the target network parameter can be wherein τ is a soft update coefficient, which is between 0 and 1, and can be preferably 0.01.

[0081] As a preferred embodiment, the reward value r of the jth nozzle of the printhead i is calculated as follows: ij The calculation is as follows:

[0082]

[0083] wherein η is a reward coefficient determined by the regulation process; μ i V is the average droplet volume of all nozzles of the printhead i in the next state, V t V is the target average volume corresponding to the current epoch, V ij V is the droplet volume of the jth nozzle of the printhead i in the next state.

[0084] As a preferred embodiment, the reward coefficient η is in the range of [1, 100].

[0085] In general, as Figure 7 shown in the figure, the embodiment quickly detects the droplet volume data of the ten thousand nozzles of the multiple printheads under different driving voltage waveform parameters through test printing dot matrix, and then establishes a jetting data model of each printhead based on the data. Then a strategy network for controlling the average droplet volume of the multiple printheads and multiple nozzles is constructed, and iterative learning is performed through interaction with the printhead data model, so that it converges. Finally, in online control, through the actual feedback of the droplet volume deviation, the output waveform parameters are regulated to make the average droplet volume of the multiple printheads and multiple nozzles converge to the accuracy requirement range of the target volume. In this way, the number of nozzles within the target volume range can be maximized, thereby improving the use efficiency and printing accuracy of the multiple nozzles.

[0086] Embodiment Two

[0087] A computer readable storage medium comprising a stored computer program, wherein the computer program, when executed by a processor, controls the device where the storage medium is located to perform the method for regulating the average droplet volume of the multiple nozzles of the inkjet printer as described above.

[0088] The related technical solutions are the same as those of Embodiment One, which will not be repeated here.

[0089] Those skilled in the art will readily understand that the above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for controlling the mean droplet volume of a large number of nozzles in inkjet printing, characterized in that: include: Randomly set the printing n The initial waveform parameter vector of each nozzle is generated, and all nozzles are driven to print ink, and the droplet volume of each nozzle is collected; Calculate the droplet volume deviation of each nozzle relative to the target mean volume to determine whether the average value of the droplet volume deviation of all nozzles is less than the preset accuracy threshold. If so, complete the control and end the process. If not, input the droplet volume deviation of each nozzle of each nozzle and the current waveform parameter vector of the nozzle into the pre-built waveform parameter feedback control model of the nozzle to obtain the control increment of each waveform parameter. Add the control increment of each waveform parameter to the corresponding parameter value in the current waveform parameter vector to obtain a new waveform parameter vector, and re-execute the drive printing; The waveform parameter feedback control model is constructed in the following way: Constructing a set of driving voltage waveform parameter vectors, using each waveform parameter vector in the set to drive each printhead to collect the droplet volume of each nozzle in each printhead under each waveform parameter vector, and constructing a state set for each printhead including two dimensions: the waveform parameter vector and its corresponding droplet volume of each nozzle, to train a spray data model for the printhead for predicting the droplet volume of each nozzle; Initialize a strategy network for each nozzle and m evaluation networks and m target networks corresponding one-to-one to the m nozzles of the nozzle; the input of the strategy network is the current waveform parameter vector of the nozzle and the current droplet volume deviation of the m nozzles relative to the target mean volume, and the output is the control increment of each waveform parameter, which is used to be added to the corresponding parameters in the current waveform parameter vector as a new waveform parameter vector to be fed back to the injection data model of the nozzle to predict the droplet volume of each nozzle in the next state; with the assistance of m evaluation networks and m target networks, the strategy network of the nozzle and the corresponding injection data model are used for interactive learning to train the strategy network as the waveform parameter feedback control model of the nozzle.

2. The method for controlling the mean droplet volume according to claim 1, wherein: The interactive learning is implemented as follows: S1. Start an epoch and randomly set the target mean volume within the adjustable volume range. , and randomly set the same initial waveform parameter vector for each nozzle; S2. Based on the ejection data model and current waveform parameter vector of each nozzle, obtain the droplet volume of each nozzle in the nozzle, calculate the droplet volume deviation of each nozzle relative to the target mean volume under the drive of the current waveform parameter vector, and construct the current state of the nozzle including the current waveform parameter vector and the current droplet volume deviation of each nozzle; S3. Input the current state of each nozzle into the strategy network of the nozzle to obtain the control increment of each waveform parameter used to drive the nozzle. Add each waveform parameter in the current waveform parameter vector to the corresponding control increment to obtain a new waveform parameter vector for the nozzle. Repeat S2 to obtain the next state of the nozzle. Calculate the reward value for each nozzle in the next state, as well as an indicator to indicate whether the average deviation of the droplet volume of all nozzles in the next state relative to the target mean volume reaches the accuracy threshold d ; The current state of each nozzle, the control increment and the next state and all nozzle reward values ​​and identifiers in the next state d Splice to form a sample; S4. Take the next state of each nozzle as the new current state of the nozzle and determine whether the current sample size reaches the threshold. If so, perform an iterative training of the strategy network, m evaluation networks, and m target networks of the nozzle based on the current samples and execute S5. If not, execute S5 directly. S5. Judgment mark Whether the requirements are met in the current epoch. If so, if the number of epochs is judged to have reached the threshold, then the process ends. If the number of epochs is judged to have not reached the threshold, then S1 is re-executed; if not, S3 is re-executed.

3. The method for controlling the mean droplet volume according to claim 2, wherein: The identification d Expressed as: ； Where, It represents the average deviation of the droplet volume of all nozzles of each printhead relative to the target mean volume, Indicates the accuracy threshold.

4. The method for controlling the mean droplet volume according to claim 1 or 2, wherein: In each iterative training, the ratio of the droplet volume deviation of each nozzle to the sum of the droplet volume deviations of all nozzles in the nozzle is used as the weight of the evaluation network output of the nozzle, and the outputs of m evaluation networks are weighted and summed to construct the loss of the strategy network to update the parameters of the strategy network; the target network of each nozzle is used to construct the target output of the evaluation network of the nozzle, and the loss of the evaluation network is constructed by subtracting the target output from the actual output of the evaluation network to synchronously update the parameters of the evaluation network of each nozzle; the parameters of the target network of each nozzle are soft-updated using the current parameters of the evaluation network of the nozzle after a preset number of iterations.

5. The method for controlling the mean droplet volume according to claim 4, wherein: The loss of the strategy network of each sprinkler is: ； Where, For nozzle i Strategy Network Parameters, is the sample size required for one iterative training, is the target mean volume corresponding to the current epoch, For nozzle i No. j The droplet volume of each nozzle in the next state; Indicates the nozzle i No. j The evaluation network output of each nozzle is Indicates the nozzle i No. j Evaluation network parameters of each nozzle, and For nozzle i No. j The input of the evaluation network of each nozzle is Indicates the nozzle i In the k The current state in samples, Indicates the nozzle i In the k Each waveform parameter controls the increment in samples.

6. The method for controlling the mean droplet volume according to claim 4, wherein: The loss of the evaluation network for each nozzle is specifically: ； Where, Indicates the sample size required for one iterative training; is the discount coefficient, which ranges from 0 to 1; For nozzle i No. j Target network parameters for each nozzle, For nozzle i No. k The next state in samples, is the control increment of each waveform parameter calculated by the policy network based on the next state; For nozzle i No. j Evaluation network parameters of each nozzle; For nozzle i No. k The current state in samples, For nozzle i No. k the control increment in samples; Indicates the next state of the nozzle i No. j The reward value of each nozzle is determined as follows: the deviation of the average volume of all nozzles of the nozzle relative to the target mean volume is multiplied by the reward coefficient as the nozzle reward, and the ratio of the droplet volume deviation of each nozzle relative to the target mean volume to the sum of the droplet volume deviations of all nozzles is used as the distribution coefficient of the nozzle to the nozzle reward, and the nozzle reward is distributed to obtain the reward value of the nozzle.

7. The method for controlling the mean droplet volume according to claim 6, wherein: nozzle i No. j Nozzle Reward Value The calculation method is: ； Where, is the reward coefficient; For nozzle i The average droplet volume of all nozzles in the next state is, is the target mean volume corresponding to the current epoch, For nozzle i No. j The droplet volume of each nozzle in the next state.

8. The method for controlling the mean droplet volume according to claim 7, wherein: The reward coefficient The value range is [1,100].

9. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute a method for controlling the droplet volume mean of a large number of inkjet printing nozzles as described in any one of claims 1 to 8.

Images