A smart screening method and apparatus for arrayed nozzles in inkjet printing.

Abstract

This invention belongs to the field of novel display inkjet printing technology, specifically relating to a method and apparatus for screening arrayed nozzles for inkjet printing. The method includes: based on the feature set of deposited droplets obtained from each nozzle's spray cycle, sequentially passing the data through a classification model and a regression model to obtain a set of state parameters corresponding to normally flying droplets; randomly selecting multiple sets of state parameters and labeling them to train an initial classifier; inputting the remaining sets of state parameters into the initial classifier to obtain corresponding classification probabilities; selecting multiple candidate sets of state parameters based on the classification probabilities, clustering them, and obtaining a label indicating whether each cluster center indicates a normal or non-normal nozzle, further training the initial classifier; and inputting each set of state parameters corresponding to each nozzle into a final classifier to determine whether the nozzle can be used for formal inkjet printing. This invention improves screening efficiency while ensuring screening accuracy.

Description

Technical Field

[0001] This invention belongs to the field of novel display inkjet printing technology, and more specifically, relates to an arrayed nozzle screening method and apparatus for inkjet printing. Background Technology

[0002] Inkjet printing technology is an emerging display panel manufacturing process that primarily uses solvents to melt organic materials and then directly sprays the material onto a substrate to form an organic functional layer. Compared to traditional vapor deposition processes, this method can print according to the required amount of organic material, thus saving 90% of raw materials. It also eliminates the need for a vacuum environment and expensive precision masks, allowing for the fabrication of substrates of any size. It is considered a revolutionary technology poised to replace vapor deposition, and the development of high-precision, highly adaptable industrial-grade inkjet printing display manufacturing equipment is a hot research area for panel manufacturers and research institutions both domestically and internationally.

[0003] Inkjet printing technology, as a manufacturing process for display panels, still faces many challenges. One key issue is the screening of abnormal nozzles during printing. In the process of inkjet printing OLED display devices, due to variations in printhead manufacturing, improper process parameter settings, and the relatively complex printing environment, jetting defects such as nozzle clogging, satellite droplets, and abnormal volume may occur. Abnormal nozzles participating in the printing process can seriously affect the quality of the final display device. Therefore, how to detect abnormal nozzles during the printing process is crucial for achieving high-quality and high-efficiency printing.

[0004] Existing inkjet nozzle screening methods primarily rely on droplet observation to directly obtain droplet volume and velocity for judgment. Droplet observation is mainly based on visual measurement, and traditional visual measurement techniques are mostly based on the principle of stroboscopic imaging. However, with the continuous increase in inkjet resolution and printing area, the substrate area also increases, and the number of nozzles to be detected also increases significantly. Using droplet observation to detect droplet parameters is too time-consuming, inefficient, and consumes a lot of process time, seriously reducing the efficiency of inkjet printing.

[0005] Therefore, how to achieve efficient and intelligent detection and screening of arrayed nozzles has become the key and difficult point for the continued development of inkjet display technology. Summary of the Invention

[0006] In view of the shortcomings and improvement needs of the existing technology, the present invention provides an arrayed nozzle screening method and device for inkjet printing, which aims to solve the problem of low nozzle screening efficiency caused by low efficiency of flying droplet observation, and realize rapid and efficient screening of arrayed nozzles.

[0007] To achieve the above objectives, according to one aspect of the present invention, a smart selection method for arrayed nozzles in inkjet printing is provided, comprising:

[0008] Based on the feature set of deposited droplets obtained from each jet of each nozzle, a trained classification model is used to determine whether the deposited droplets of the nozzle in that jet are normal or not. The feature set of each normal deposited droplet is then input into the trained regression model to obtain the state parameter set of the corresponding flying droplet.

[0009] Multiple sets of state parameters are randomly selected, and the label of whether the nozzle is normal or not corresponding to each set of state parameters is obtained to train an initial classifier. The remaining sets of state parameters are input into the initial classifier to obtain the corresponding classification probabilities. A sample selection strategy is then used to select multiple candidate sets of state parameters. The multiple candidate sets of state parameters are clustered to obtain the label of whether the nozzle is normal or not corresponding to each cluster center. This is used to further train the initial classifier to obtain the final classifier.

[0010] Each set of state parameters corresponding to each nozzle is input into the final classifier, and based on the classification results, it is determined whether the nozzle can be used for formal inkjet printing.

[0011] The beneficial effects of this invention are as follows: First, this invention constructs a sample set of state parameters for each nozzle's flying droplets. Specifically, it utilizes the feature set of deposited droplets obtained from a single injection of each nozzle. Firstly, a classification model is used to filter the feature set corresponding to normal deposited droplets. Then, based on the feature set corresponding to normal deposited droplets, a trained regression model is used to obtain the set of state parameters (including the volume and velocity of the flying droplets) corresponding to the normal deposited droplets, thus constructing a sample set of state parameters. Subsequently, this invention proposes to use the aforementioned constructed sample set of state parameters to construct a classifier capable of accurately predicting whether a nozzle is normal or not based on the state parameter set through a series of specific operations such as pre-training, sample selection, clustering, and retraining. Finally, using the state parameter sets of multiple flying droplets corresponding to each nozzle to be screened after multiple injections, the constructed classifier is used to obtain prediction results for the normality of multiple corresponding nozzles. Based on the multiple prediction results corresponding to the nozzle to be screened, it is determined whether the nozzle can be screened for formal inkjet printing. This invention introduces three machine learning models to collaboratively achieve intelligent nozzle screening, which greatly improves screening efficiency compared to existing methods. In particular, when constructing the classifier, this invention also proposes a series of specific training operations such as pre-training, sample selection, clustering, and retraining, which improves screening efficiency while ensuring screening accuracy.

[0012] Furthermore, the following sample selection strategy in active learning is adopted:

[0013]

[0014] In the formula, This indicates that the initial classifier is used to classify the remaining set of state parameters x without labels. i Predicted as category The probability, and These represent the negative example category and the positive example category, respectively; x * The set of state parameters in the remaining set of state parameters with the smallest difference in probability between the two predicted classes is considered as the candidate set of state parameters.

[0015] A further beneficial effect of this invention is that the sample selection strategy employed prioritizes sample data that are easily classified as either positive or negative examples and lie at the class boundaries in the sample space. This invention selects these samples as candidate set of state parameters to further train the classifier, which greatly improves the predictive performance of the classifier and thus enhances the accuracy of nozzle screening. Positive examples are data whose state parameters are within a threshold, while negative examples are data whose state parameters exceed the threshold.

[0016] Furthermore, the k-means algorithm is used to cluster multiple candidate state parameter sets.

[0017] Further beneficial effects of the present invention are: the present invention adopts the k-means clustering method to realize an unsupervised clustering algorithm, which reduces information redundancy, can effectively reduce the labeling of a large number of similar samples, avoids the slow movement of the classification hyperplane and the waste of iteration costs, and effectively improves the convergence speed and classification accuracy of the algorithm.

[0018] Furthermore, the criteria for determining whether each nozzle is suitable for inkjet printing are as follows:

[0019] When the final classifier outputs all classification results based on the set of state parameters corresponding to the nozzle, and all are normal, the nozzle is selected as suitable for inkjet printing.

[0020] The beneficial effects of this invention are: if all the sampled samples are positive samples, the nozzle is judged to be a normal nozzle; if negative samples appear, the nozzle is judged to be an abnormal nozzle, thereby further improving the screening accuracy.

[0021] Furthermore, the classification model is composed of multiple trained support vector machines. Each feature dimension in the feature set corresponds to a support vector machine in the classification model for predicting the probability that the deposited droplet is abnormal in that feature dimension. At the same time, the classification model also includes a support vector machine for predicting the probability that the deposited droplet is normal.

[0022] The specific method for using a classification model is as follows:

[0023] Each feature set is input into each support vector machine in the classification model, and each support vector machine predicts and outputs a probability; the normality or non-normality of the sediment droplet corresponding to the highest probability is taken as the output result of the classification model.

[0024] A further beneficial effect of this invention is that, for predicting whether a sedimentary droplet is normal or not, this invention employs a multi-class support vector machine. In the feature set, each feature dimension corresponds to a support vector machine in the classification model, used to predict the probability that the sedimentary droplet is abnormal in that feature dimension, and another support vector machine used to predict the probability that the sedimentary droplet is normal. Each support vector machine is specifically used to predict the probability that the sedimentary droplet is abnormal in a certain dimension, rather than using a single machine for prediction. This avoids misclassifying abnormal sedimentary droplets as normal ones, greatly improving the accuracy of determining whether a sedimentary droplet is normal and significantly enhancing the classification accuracy.

[0025] Furthermore, each support vector machine in the classification model is obtained using the following training method:

[0026] Multiple feature set samples are randomly selected from the feature set sample set. Based on the prediction target of the support vector machine, labels are obtained for each feature set sample indicating whether the sedimentary droplets are normal or whether they are abnormal in a certain feature dimension, in order to train an initial support vector machine. The remaining feature set samples are input into the initial support vector machine to obtain the corresponding prediction probabilities. A sample selection strategy is then used to construct multiple candidate feature set samples. The multiple candidate feature set samples are clustered to determine each cluster center. Based on the prediction target, labels are obtained for each cluster center indicating whether the sedimentary droplets are normal or whether they are abnormal in a certain feature dimension, in order to further train the initial support vector machine, resulting in a final support vector machine used to achieve the prediction target.

[0027] A further beneficial effect of the present invention is that, when constructing each support vector machine, the present invention proposes a series of specific training operations such as pre-training, sample selection, clustering, and retraining, which improves the classification accuracy of the support vector machine.

[0028] Furthermore, when training the support vector machine in the classification model, the following active learning sample selection strategy is adopted:

[0029]

[0030] In the formula, X is the sample set composed of the remaining feature set samples, p(y B |x i ) and p(y SB |x i ) represents the initial support vector machine for sample x iDetermine the two highest probability values ​​obtained.

[0031] A further beneficial effect of the present invention is that the sample selection strategy adopted in the present invention prioritizes sample data that are at the class boundary of the sample space. The present invention selects these samples as candidate sediment droplet feature sets to further train the support vector machine, which can greatly improve the prediction performance of the support vector machine.

[0032] Furthermore, the clustering method used in training the support vector machine in the classification model is the k-means algorithm.

[0033] Further beneficial effects of the present invention are: the present invention adopts the k-means clustering method to realize an unsupervised clustering algorithm, which reduces information redundancy, can effectively reduce the labeling of a large number of similar samples, avoids the slow movement of the classification hyperplane and the waste of iteration costs, and effectively improves the convergence speed and classification accuracy of the algorithm.

[0034] Furthermore, the feature set includes the following feature dimensions: whether the number of deposited droplets is 0, whether the number of deposited droplets is greater than 1, whether there is a landing point deviation of the deposited droplets, and whether the diameter of the deposited droplets exceeds the threshold.

[0035] Further beneficial effects of the present invention are: the present invention uses multiple dimensions such as the number of deposited droplets, the deviation of the deposited droplet landing point, and whether the diameter of the deposited droplets exceeds the threshold to label abnormal deposited ink droplets, which can improve the classification accuracy of the sample set, avoid the abnormal deposited ink droplet parameters from affecting the prediction accuracy of the regression model, and ultimately affect the selection of nozzles.

[0036] The present invention also provides an intelligent screening device for an array of nozzles for inkjet printing, comprising: an inkjet printing module, a droplet deposition substrate, a deposited droplet observation module, a motion module, and a control module;

[0037] The control module is used to control the inkjet printing module to move above the droplet deposition substrate for test spraying via the motion module, and to control the deposition droplet observation module to move to the droplet deposition area via the motion module, to acquire images of the deposition droplets ejected from each nozzle and transmit them to the control module; the control module is also used to detect the feature set of the deposition droplets corresponding to each nozzle based on the deposition droplet images, and to execute the intelligent selection method for arrayed nozzles for inkjet printing as described above.

[0038] The beneficial effects of this invention are as follows: The intelligent screening device for arrayed nozzles in inkjet printing proposed in this invention includes an inkjet printing module, a droplet deposition substrate, a deposition droplet observation module, and a motion module, which work together to acquire images of the deposition droplets. The image is processed by a control module to obtain a set of deposition droplet features. Thus, the control module can use the intelligent screening method for arrayed nozzles in inkjet printing as described above to achieve high-precision and high-efficiency nozzle screening.

[0039] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:

[0040] This invention introduces three machine learning models to collaboratively achieve intelligent nozzle screening, significantly improving screening efficiency compared to existing methods. Specifically, in constructing the classifier, this invention proposes a series of specific training operations, including pre-training, sample selection, clustering, and retraining, effectively ensuring screening accuracy while improving efficiency. Attached Figure Description

[0041] Figure 1 A flowchart illustrating an intelligent selection method for arrayed nozzles in inkjet printing, provided in an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram illustrating the acquisition of the feature set of deposited droplets provided in an embodiment of the present invention;

[0043] Figure 3 This is a process flow diagram of the nozzle selection stage in the intelligent nozzle selection method for an arrayed nozzle of an inkjet printing system provided in an embodiment of the present invention.

[0044] Figure 4 This is a flowchart illustrating the construction and application of a two-layer proxy model in the intelligent selection method for arrayed nozzles in an inkjet printing system provided in this embodiment of the invention.

[0045] Figure 5 The structure diagram of the machine learning regression model provided in the embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0047] Example 1

[0048] A smart selection method for arrayed nozzles in inkjet printing, such as Figure 1 As shown, it includes:

[0049] Based on the feature set of deposited droplets obtained from each jet of each nozzle, a trained classification model is used to determine whether the deposited droplets of the nozzle in that jet are normal or not. The feature set of each normal deposited droplet is then input into the trained regression model to obtain the state parameter set of the corresponding flying droplet.

[0050] Multiple sets of state parameters are randomly selected, and the label of whether the nozzle is normal or not corresponding to each set of state parameters is obtained to train an initial classifier. The remaining sets of state parameters are input into the initial classifier to obtain the corresponding classification probabilities. A sample selection strategy is then used to select multiple candidate sets of state parameters. The multiple candidate sets of state parameters are clustered to determine each cluster center. The label of whether the nozzle is normal or not corresponding to each cluster center is obtained to further train the initial classifier and obtain the final classifier.

[0051] Each set of state parameters corresponding to each nozzle is input into the final classifier, and based on the classification results, it is determined whether the nozzle can be used for formal inkjet printing.

[0052] It should be noted that the set of state parameters can include volume and velocity.

[0053] This embodiment first constructs a sample set of state parameters for each nozzle's flying droplets. Specifically, it utilizes the feature set of deposited droplets obtained from a single injection from each nozzle. First, a classification model is used to filter the feature set corresponding to normal deposited droplets. Then, based on the feature set of normal deposited droplets, a trained regression model is used to obtain the set of state parameters (including droplet volume and velocity) for the flying droplets corresponding to these normal droplets, thus constructing the sample set of state parameters. Subsequently, this invention proposes to use the aforementioned sample set of state parameters to construct a classifier capable of accurately predicting the normality of nozzles based on the state parameter set through a series of specific operations such as pre-training, sample selection, clustering, and retraining. Finally, using the state parameter sets of multiple flying droplets corresponding to each nozzle to be screened after multiple injections, the constructed classifier is used to obtain prediction results for the normality of multiple nozzles. Based on the multiple prediction results corresponding to the nozzle to be screened, it is determined whether the nozzle can be screened for formal inkjet printing. This embodiment introduces three machine learning models to collaboratively achieve intelligent nozzle screening, greatly improving screening efficiency compared to existing methods. In this embodiment, when constructing the classifier, a series of specific training operations are proposed, such as pre-training, sample selection, clustering, and retraining, to improve screening efficiency while ensuring screening accuracy.

[0054] The above sample selection strategy can adopt the sample selection strategy in active learning, including the sample selection strategy based on BvSB and the sampling strategy based on uncertainty. As a preferred implementation, this embodiment adopts the following sample selection strategy based on BvSB:

[0055]

[0056] In the formula, represents the probability of predicting the remaining set of unlabeled samples x of one state parameter as the class i using the initial classifier, and and represent the negative example class and the positive example class respectively; x * represents the set of state parameters with the smallest difference in the probabilities predicted as the two classes in the remaining set of state parameters, and is used as the candidate set of state parameters.

[0057] The adopted sample selection strategy gives priority to those sample data that are easily judged as positive and negative examples and are at the class boundary of the sample space. In this embodiment, these samples are selected as the candidate set of state parameters to further train the classifier, which can greatly improve the prediction performance of the classifier and thus is beneficial to improving the accuracy of nozzle screening. Among them, the positive examples are the data with state parameters within the threshold, and the negative examples are the data with state parameters exceeding the threshold.

[0058] As a preferred implementation, the k-means algorithm can be used to cluster multiple candidate sets of state parameters.

[0059] Specifically, t unlabeled samples are selected from the unlabeled sample set X as the candidate sample set, and then k (k < t) cluster centers are initialized on the candidate sample set by the k-means algorithm, which are u j , j = 1, 2,..., k respectively. Each cluster center corresponds to a cluster c j , j = 1, 2,..., k. Calculate the distance from each sample x i in the current cluster to the cluster center of each cluster, and then divide x i into the nearest cluster. After all the samples are divided, re-determine the cluster center, and repeat the above two steps until the distance between the samples in each cluster and the cluster center sample no longer decreases, and return the final clustering result. For these k clustering results, select the samples at the cluster centers to submit for annotation to obtain the corresponding labels. Select unlabeled samples based on the criterion of difference, and the unsupervised clustering algorithm determines the cluster to which the sample belongs. Among them, u j is the cluster center, and c (i) is the sample xj The cluster that is closest to the k clusters.

[0060] The sample selection strategy based on BvSB, combined with k-means clustering, achieves an unsupervised clustering algorithm, reducing information redundancy. Labeling a large number of similar samples leads to slow hyperplane movement, wasting iteration costs and slowing the algorithm's convergence, thus affecting classification accuracy.

[0061] As a preferred implementation scheme, the criteria for determining whether each nozzle is suitable for inkjet printing are:

[0062] When all classification results output by the final classifier based on the set of state parameters corresponding to the nozzle are normal, the nozzle is selected as suitable for inkjet printing.

[0063] If all the sampled samples are positive, the nozzle is judged as a normal nozzle; if any negative samples are found, the nozzle is judged as an abnormal nozzle, thus further improving the screening accuracy.

[0064] As a preferred implementation, the classification model consists of multiple trained support vector machines (SVMs). Each feature dimension in the feature set corresponds to a SVM in the classification model for predicting the probability that the deposited droplet is abnormal in that feature dimension. The classification model also includes an SVM for predicting the probability that the deposited droplet is normal. The specific method of using the classification model is as follows: each feature set is input into each SVM in the classification model, and each SVM predicts and outputs a probability. The normality of the deposited droplet corresponding to the highest probability is taken as the output result of the classification model.

[0065] To predict whether a sediment droplet is normal or not, this invention employs a multi-class support vector machine (SVM). For each feature dimension in the feature set, there is a corresponding SVM in the classification model used to predict the probability that a sediment droplet is abnormal in that feature dimension, and another SVM used to predict the probability that a sediment droplet is normal. Each SVM is specifically designed to predict the probability of a sediment droplet being abnormal in a particular dimension, rather than using a single SVM for all dimensions. This avoids misclassifying abnormal sediment droplets as normal ones, greatly improving the accuracy of determining whether a sediment droplet is normal and significantly enhancing classification precision.

[0066] The classification model used in this embodiment is an active learning classification model, specifically including: an active learning model based on an improved strategy of information entropy, with the initial model consisting of multiple support vector machine models, modeled through five components, A = (G, U m ,T,Q,U n ), where G represents the classifier model, Q represents the sample selection strategy, and U mIndicates a labeled sample set, U n Let T represent the unlabeled sample set, and T represent the human annotators. The initial model selected for the classification model is a support vector machine model, which can be represented as: y(x) = sign(w t Φ(x)+b); where Φ(x) is the kernel function, w is the function control coefficient, b is an unknown constant, and t is the coefficient control factor.

[0067] As a preferred implementation scheme, each support vector machine in the classification model is obtained using the following training method:

[0068] Multiple feature set samples are randomly selected from the feature set sample set. Based on the prediction objective of the support vector machine, labels are obtained for each feature set sample indicating whether the sedimentary droplets are normal or whether they exhibit anomalies in a certain feature dimension, in order to train an initial support vector machine. The remaining feature set samples are input into the initial support vector machine to obtain the corresponding prediction probabilities. A sample selection strategy is then used to construct multiple candidate feature set samples. These candidate feature set samples are clustered to determine the cluster centers. Based on the prediction objective, labels are obtained for each cluster center indicating whether the sedimentary droplets are normal or whether they exhibit anomalies in a certain feature dimension, in order to further train the initial support vector machine, resulting in the final support vector machine used to achieve the prediction objective.

[0069] In constructing each support vector machine, this embodiment proposes a series of specific training operations such as pre-training, sample selection, clustering, and retraining, which can improve the classification accuracy of the support vector machine.

[0070] As a preferred implementation scheme, the following active learning sample selection strategy can be adopted when training the support vector machine in the classification model:

[0071]

[0072] In the formula, X is the sample set composed of the remaining feature set samples, p(y B |x i ) and p(y SB |x i ) represents the initial support vector machine for sample x i Determine the two highest probability values ​​obtained.

[0073] The sample selection strategy adopted prioritizes sample data that are located at the class boundaries of the sample space. This invention selects these samples as candidate feature sets of sediment droplets to further train the support vector machine, which can greatly improve the prediction performance of the support vector machine.

[0074] As a preferred implementation scheme, the clustering method used when training the support vector machine in the classification model is the k-means algorithm.

[0075] We employ a BvSB-based strategy combined with k-means clustering to implement an unsupervised clustering algorithm. This reduces information redundancy, effectively minimizes the need to label a large number of similar samples, avoids the waste of iteration costs due to slow hyperplane movement, and effectively improves the convergence speed and classification accuracy of the algorithm.

[0076] As a preferred implementation scheme, the feature dimensions included in the above feature set are: whether the number of deposited droplets is 0, whether the number of deposited droplets is greater than 1, whether there is a landing point deviation of the deposited droplets, and whether the diameter of the deposited droplets exceeds the threshold.

[0077] By using multi-dimensional annotation of anomalous deposited ink droplets, the classification accuracy of the sample set can be improved, and the parameters of anomalous deposited ink droplets can be avoided from affecting the prediction accuracy of the regression model, which in turn affects the selection of nozzles.

[0078] It should be noted here that, regarding the regression model where the volume and velocity of flying droplets are mapped from the feature set of sedimentary droplets, an approximate formula is used to illustrate the regression construction process:

[0079] Taking the volume mapping process as an example, the same principle applies to velocity. The actual mapping of volume corresponds to the area of ​​the deposited ink droplet, while velocity corresponds to the droplet's landing point deviation. The mapping relationship is given by the following formula:

[0080] y = xA1A2;

[0081]

[0082]

[0083] Where x is a 1×4 vector representing the set of input sediment droplet features, which is the input sample; A1 is a 4×9 matrix representing the set of weights mapped from the input layer to the hidden layer; A2 is a 9×1 vector representing the set of weights mapped from the hidden layer to the output layer; y is the predicted output of the regression model for sample x, ν i,j ω represents the weights mapped from the i-th node in the input layer to the j-th node in the hidden layer; p,1 This represents the weights mapped from the p-th node in the hidden layer to the nodes in the output layer; i represents an integer from 1 to 4; j and p represent integers from 1 to 9.

[0084] To more clearly describe the method of the present invention, the intelligent screening method of this embodiment can be described as including: a model building stage, a droplet parameter detection stage, and a nozzle parameter acquisition and screening stage. The model building stage includes operations S1 to S6, the droplet parameter detection stage includes operation S7, and the nozzle parameter acquisition and screening stage includes operation S8.

[0085] Model building phase:

[0086] Operation S1 is used to perform a trial print, detect the feature set of the deposited droplets corresponding to each nozzle, and use it as input parameters to build an unlabeled classification dataset.

[0087] Operation S2 is used to label the samples in the unlabeled classification dataset based on anomalies in droplet number and position, thus creating a labeled classification dataset.

[0088] Operation S3: Use the labeled classification dataset to train a machine learning classification model to obtain a machine learning classification model for determining whether there is droplet anomaly based on the feature set of deposited droplets.

[0089] Operation S4: Use the flight droplet volume and velocity measurement results to label the normal samples in the labeled classification dataset and establish a regression dataset;

[0090] Operation S5: Use the regression dataset to train a machine learning regression model to obtain a machine learning regression model that calculates droplet volume based on the feature set of deposited droplets.

[0091] Operation S6: Use the established classification and regression models to build a two-layer surrogate model;

[0092] Droplet parameter detection stage:

[0093] Operation S7: Before the actual printing, move the printhead to the test printing area for test printing, detect the set of deposited droplet features corresponding to each nozzle, and input them into the two-layer proxy model to calculate the droplet volume and velocity corresponding to the set of deposited droplet features.

[0094] Operation S8 obtains the droplet velocity and volume parameters. Using an active learning-based sampling method, the overall state of each nozzle is obtained for nozzle selection.

[0095] A backpropagation (BP) neural network regression model can be used as the regression model mentioned above. Specifically, it includes an input layer, a hidden layer, and an output layer. The input layer contains P1 = M nodes, where M is the number of sediment droplet features. No activation function is used in the input layer. The hidden layer contains P2 = 2P1 + 1 nodes, and the activation function is: Where, x p For the p-th input sample, ci Let be the i-th center point, h be the number of nodes in the hidden layer, and n be the number of output samples. The output layer contains P3 = 1 node, and the activation function is a linear function with the expression f(x) = x. The predicted output of the RBF neural network regression model for sample x is: Where, x p For the p-th input sample, c i Let h be the i-th center point, h be the number of nodes in the hidden layer, and n be the number of output samples.

[0096] After obtaining the set of state parameters of the flying droplets using the two-layer surrogate model, it is necessary to process the large set of parameters corresponding to each nozzle to predict the condition of that nozzle. An active learning-based sampling method is adopted, which specifically includes: training a classifier C using a labeled classification sample set T, using a random forest classifier, using a query strategy in active learning to evaluate the importance of unlabeled samples and select the most valuable samples, and further training the classifier.

[0097] In other words, the classifier training method used in this embodiment is as follows: an active learning approach is adopted. The active learning process is a process of continuous iterative training of a classifier. The sample set U1 is labeled samples and U2 is unlabeled samples. A model G is first trained using U1. This model is then used to select samples from U2. Based on the sample selection strategy, the most suitable samples are selected and handed over to experts for labeling. Each time, the samples that can best improve the model performance are selected.

[0098] In summary, the intelligent nozzle screening method of this embodiment includes: performing trial printing to detect the set of deposited droplet features corresponding to each nozzle, and using these features as input parameters to establish an unlabeled classification dataset; labeling the unlabeled classification dataset based on the presence, quantity, position, and diameter anomalies of droplets to establish a labeled classification dataset, and training an active learning model using the labeled classification dataset; using the model to select samples from the unlabeled sample set, identifying the most suitable samples according to the sample selection strategy and having them labeled by experts, selecting the samples that best improve model performance each time for model training, allowing experts to obtain a better model with minimal work; labeling normal samples in the labeled classification dataset using the measured volume and velocity of flying droplets to establish a regression dataset, and then training a BP neural network regression model using the regression dataset; and using the trained active learning classification model and the BP neural network regression model to establish a two-layer surrogate model to calculate the droplet volume and velocity corresponding to any set of deposited droplet features. By employing an active learning-based sampling method to obtain droplet volume and velocity that meet certain conditions, and then using this information for classifier training and nozzle selection, this invention constructs a framework for applying machine learning models to nozzle ejection state prediction and nozzle selection. By applying active learning methods, a high-performance prediction model is obtained with minimal data labeling costs. While ensuring detection accuracy, this invention effectively improves the efficiency of arrayed nozzle selection in large-area inkjet printing production lines, achieving efficient and intelligent manufacturing.

[0099] Example 2

[0100] A smart screening device for an array of nozzles for inkjet printing includes: an inkjet printing module, a droplet deposition substrate, a deposited droplet observation module, a motion module, and a control module.

[0101] The control module is used to control the inkjet printing module to move above the droplet deposition substrate for test spraying via the motion module, and to control the deposition droplet observation module to move to the droplet deposition area via the motion module, to collect images of the deposition droplets ejected from each nozzle and transmit them to the control module; the control module is also used to detect the feature set of the deposition droplets corresponding to each nozzle based on the deposition droplet images, and to execute the intelligent screening method for arrayed nozzles for inkjet printing as described in Embodiment 1 above.

[0102] The intelligent screening device for arrayed nozzles in inkjet printing proposed in this embodiment includes an inkjet printing module, a droplet deposition substrate, a deposition droplet observation module, and a motion module. These components work together to acquire images of the deposited droplets. The image is then processed by a control module to obtain a set of deposited droplet features. Thus, the control module can employ the intelligent screening method for arrayed nozzles in inkjet printing described above to achieve high-precision and high-efficiency nozzle screening.

[0103] To better illustrate the present invention, the following specific detection system is provided to further explain the detection process described in Embodiment 1:

[0104] Figure 2 This is a schematic diagram of obtaining the feature set of deposited droplets provided in an embodiment of the present invention. The diagram clearly shows the deposition of droplets on the substrate, including normally deposited droplets and abnormally deposited droplets, including no deposition and abnormal deposition quantity. Appropriate droplets are selected to measure relevant parameters, including the diameter of the deposited droplets and the deviation of the deposited droplets from the theoretical landing point, to obtain relevant parameters of the feature set of deposited droplets.

[0105] Furthermore, Figure 3 This is a process flow diagram of volume and velocity detection in the intelligent nozzle screening method provided in this embodiment of the invention, specifically including the following steps:

[0106] (1) Conduct test spraying, record the theoretical landing position of the sprayed droplets in each nozzle, and then acquire the image of the deposited droplets through the downward observation camera and obtain the feature set of the deposited droplets corresponding to the current nozzle.

[0107] (2) Input the obtained set of sediment droplet features into the active learning classification model and output the droplet anomaly information. If there is an anomaly in the sediment droplet, mark the current nozzle as an abnormal nozzle and directly proceed to the detection of the next nozzle and jump to step (1); if there is no anomaly in the sediment droplet, proceed to the next step.

[0108] (3) Input the obtained set of sediment droplet features into the BP neural network regression model, output the droplet volume and velocity and record them. Use the sampling detection method based on active learning to determine whether the nozzle is abnormal and perform the next nozzle detection.

[0109] (4) Repeat steps (1) to (3) until all nozzles are detected. Output the volume and velocity of the droplets corresponding to all nozzles, determine the state of the nozzles, and complete the screening of inkjet printing nozzles.

[0110] Furthermore, Figure 4 This is a flowchart illustrating the construction and application of a two-layer proxy model in the intelligent nozzle screening method provided in this embodiment of the invention, specifically including the following steps:

[0111] (1) Move the nozzle module to the ink droplet observation position, flash the nozzle and observe it. Calculate and record the volume and velocity of the droplets corresponding to each nozzle by acquiring images of the flying droplets through the side-view observation camera.

[0112] (2) Move the nozzle module above the droplet deposition substrate for test spraying, record the theoretical landing position of the droplets sprayed from each nozzle, and then move the downward observation camera above the droplet deposition substrate for observation. Collect the image of the deposited droplets and calculate the feature set (d,m,Δx,Δy) of the deposited droplets corresponding to each nozzle. Here, d represents the spreading diameter of the deposited droplets, m represents the number of deposited droplets, Δx represents the x-direction landing point deviation of the deposited droplets, and Δy represents the y-direction landing point deviation of the deposited droplets. Extract a certain number of samples from all the feature sets of deposited droplets and use them as input parameters to establish an unlabeled classification dataset θ.

[0113] According to a preferred embodiment of the present invention, the capacity of the unlabeled classification dataset θ is N. H =3000, the constructed unlabeled classification dataset can be represented as:

[0114]

[0115] Where (d,m,Δx,Δy) represents the set of sediment droplet features for each sample group.

[0116] (3) Label the samples in the unlabeled classification dataset θ based on the abnormal number and position of droplets. When there are no deposited droplets or the number of deposited droplets is greater than or equal to two, or the x / y direction droplet deviation exceeds the preset value, label the unlabeled classification dataset samples as abnormal samples. Otherwise, label the unlabeled classification dataset samples as normal samples, and establish a labeled classification dataset D. c :

[0117]

[0118] Where y represents the sample label, which takes the value of 0 or 1; when y = 0, it means that the sediment droplets corresponding to the sample are abnormal and the sample is an abnormal sample; when y = 1, it means that the sediment droplets corresponding to the sample are not abnormal and the sample is a normal sample.

[0119] (4) Using the labeled classification dataset D c Train a random active learning classification model. The input to the model is a set of features of the sediment droplets. The output of the model is whether the sediment droplets are abnormal.

[0120] Furthermore, the machine learning classification model is an active learning classification model, specifically including: an active learning model based on an improved strategy using information entropy, with the initial model being a support vector machine model, modeled through five components:

[0121] A=(G,U m ,T,Q,U n );

[0122] Among them, G represents the classifier model, Q represents the sample selection strategy, and U m represents the labeled sample set, and U n represents the unlabeled sample set, and T represents the human annotator.

[0123] The initial model selected by the active learning classification model is a support vector machine model, which can be expressed as:

[0124] y(x) = sign(w t Φ(x) + b);

[0125] Among them, Φ(x) is the kernel function, w is the function control coefficient, b is the unknown constant, and t is the coefficient control factor.

[0126] Furthermore, according to the above active learning classification model, the sample strategy selection method is:

[0127]

[0128] Among them, X is the unlabeled sample set, p(y B |x i ) and p(y SB |x i ) are the two highest probability values obtained by the classifier for the sample x i . If the uncertainty of a sample is higher, the C(x i ) value of this sample is smaller.

[0129] According to a preferred embodiment of the present invention, the training process of the active learning classification model is as follows:

[0130] Select t unlabeled samples from the unlabeled sample set X as the candidate sample set. Then, initialize k (k < t) clustering centers on the candidate sample set , which are u j , j = 1, 2,..., k. Each clustering center corresponds to a cluster c j , j = 1, 2,..., k. For each sample x i in the candidate sample set, calculate its distance to each clustering center, and then divide x i into the cluster with the smallest clustering, and calculate the sample and the closest cluster among the k clusters. After all samples are divided, re-determine the cluster center, and repeat the above two steps until the distance between the samples in each cluster and the cluster center sample no longer decreases, and return the final clustering result. For these k clustering results, select the samples at the clustering centers and submit them for annotation.

[0131] (5) Use the flight droplet volume and speed for the labeled classification data set D cNormal samples were labeled to create a droplet volume regression dataset D. r :

[0132]

[0133] Where V represents the sample label, specifically the actual volume of the droplet; V e This represents the sample label, specifically the actual velocity of the droplet; P h This represents a labeled classification dataset D. c The total number of normal samples.

[0134] (6) Using the droplet volume regression dataset D r Train a BP neural network regression model. The input to the model is the set of features of the sediment droplets, and the output of the model is the droplet volume and velocity.

[0135] More specifically, Figure 5 This is a schematic diagram of the BP neural network regression model structure provided in an embodiment of the present invention, including: an input layer, a hidden layer, and an output layer, wherein: the input layer contains P1 = M nodes, where M is the number of feature values ​​of the deposited droplets, and no activation function is used in the input layer; the hidden layer contains P2 = 2P1 + 1 nodes, and the activation function is:

[0136]

[0137] Where, x p For the p-th input sample, c i Let h be the i-th center point, h be the number of nodes in the hidden layer, and n be the number of output samples.

[0138] The output layer contains 1 node (P3 = 1), and the activation function is a linear function with the expression f(x) = x.

[0139] Furthermore, the predicted output of the above BP neural network regression model for sample x is:

[0140]

[0141] Where, x p For the p-th input sample, c i Let h be the i-th center point, h be the number of nodes in the hidden layer, and n be the number of output samples.

[0142] According to a preferred embodiment of the present invention, the training process of the BP neural network regression model is as follows:

[0143] Step 1: Divide the droplet volume regression dataset D into 5:1 ratios. r Divided into training set D re With test set D rtThe learning rate γ = 0.1, the number of iterations n = 300, the number of nodes in the input layer P1 = 5, the number of nodes in the hidden layer P2 = 11, the number of nodes in the output layer P3 = 1, and the loss function is the mean squared error function.

[0144]

[0145] In the formula, k is the number of samples in the training set. For the predicted output of the i-th sample group, y [i] This is the actual output for the i-th sample group.

[0146] Step 2: Initialize all connection weights and biases in the network to 0.

[0147] Step 3: Transfer the training set D re Using the sample features as input parameters, the forward propagation algorithm is applied to calculate the output of the j-th node in the hidden layer under the current weight condition, according to the following formula. With the output α of the output layer node [2] :

[0148]

[0149]

[0150] Where, x i v represents the input feature vector of the i-th node in the input layer. i,j ω represents the weights mapped from the i-th node in the input layer to the j-th node in the hidden layer; p,1 This represents the mapping weight from the p-th node in the hidden layer to the output layer node, where i represents an integer from 1 to 5; j and p represent integers from 1 to 11; O [1] O represents the deviation in the mapping from the input layer to the hidden layer. [2] This represents the deviation in the mapping from the hidden layer to the output layer.

[0151] Step 4: Calculate the propagation error δ at all nodes under the current weight conditions using the following formula:

[0152]

[0153]

[0154]

[0155] in, y represents the propagation error of the output layer nodes, and y represents the actual droplet volume and velocity in the sample. ω represents the propagation error at the p-th node in the hidden layer. i,1 This represents the mapping weight from the i-th node in the hidden layer to the node in the output layer; ν represents the propagation error at the i-th node in the input layer. i,j This represents the mapping weight from the i-th node in the input layer to the j-th node in the hidden layer.

[0156] Step 5: Update all connection weights and biases according to the following formula:

[0157]

[0158]

[0159]

[0160]

[0161] Where γ represents the learning rate, This represents the gradient term that maps the weights from the i-th node in the input layer to the j-th node in the hidden layer. The gradient term represents the deviation in the mapping from the input layer to the hidden layer. This represents the gradient term of the weights mapped from the p-th node in the hidden layer to the output layer nodes. This represents the gradient term representing the deviation in the mapping from the hidden layer to the output layer.

[0162] Step 6: Repeat steps 3 to 5 above, continuously updating the connection weights and biases until the iteration termination condition is met, thus completing the construction of the BP neural network regression model.

[0163] Step 7: Use a BP neural network regression model on the test set D rt Perform prediction verification and record the final prediction error range to complete the construction of the BP neural network regression model.

[0164] After completing the construction of the active learning classification model and the BP neural network regression model, the system is applied to the printing interval stage (trial printing stage) in the actual printing process. After the trial spray, the deposited droplets are detected to obtain the droplet volume and velocity, which are used for nozzle screening.

[0165] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart selection method for arrayed nozzles in inkjet printing, characterized in that, include: Based on the feature set of deposited droplets obtained from each jet of each nozzle, a trained classification model is used to determine whether the deposited droplets of the nozzle in that jet are normal or not. The feature set of each normal deposited droplet is then input into the trained regression model to obtain the state parameter set of the corresponding flying droplet. Multiple sets of state parameters are randomly selected, and the label of whether the nozzle is normal or not is obtained for each set of state parameters in order to train an initial classifier. The remaining state parameter sets are input into the initial classifier to obtain the corresponding classification probabilities. A sample selection strategy is then adopted to select multiple candidate state parameter sets. The multiple candidate state parameter sets are clustered to obtain the label of whether the nozzle is normal or not corresponding to each cluster center, so as to further train the initial classifier and obtain the final classifier. Each set of state parameters corresponding to each nozzle is input into the final classifier, and based on the classification results, it is determined whether the nozzle can be used for formal inkjet printing. The classification model is composed of multiple trained support vector machines. Each feature dimension in the feature set has a corresponding support vector machine in the classification model for predicting the probability that the deposited droplet is abnormal in that feature dimension. The classification model also includes a support vector machine for predicting the probability that the deposited droplet is normal. The specific method for using a classification model is as follows: Each feature set is input into each support vector machine in the classification model, and each support vector machine predicts and outputs a probability; the normality or non-normality of the sediment droplet corresponding to the highest probability is taken as the output result of the classification model.

2. The intelligent selection method for arrayed nozzles in inkjet printing according to claim 1, characterized in that, The following sample selection strategy is adopted in active learning: ； In the formula, This indicates that the initial classifier is used to sample the remaining set of state parameters without labels. Predicted as category The probability, and These represent the negative example category and the positive example category, respectively. The set of state parameters in the remaining set of state parameters with the smallest difference in probability between the two predicted classes is considered as the candidate set of state parameters.

3. The intelligent selection method for arrayed nozzles in inkjet printing according to claim 1, characterized in that, The k-means algorithm is used to cluster multiple candidate sets of state parameters.

4. The intelligent selection method for arrayed nozzles in inkjet printing according to claim 1, characterized in that, The criteria for determining whether each nozzle is suitable for inkjet printing are: When the final classifier outputs all classification results based on the set of state parameters corresponding to the nozzle, and all are normal, the nozzle is selected as suitable for inkjet printing.

5. The intelligent screening method for arrayed nozzles in inkjet printing according to claim 1, characterized in that, Each support vector machine in the classification model is obtained using the following training method: Multiple feature set samples are randomly selected from the feature set sample set. Based on the prediction target of the support vector machine, labels are obtained for each feature set sample indicating whether the sedimentary droplets are normal or whether they are abnormal in a certain feature dimension, in order to train an initial support vector machine. The remaining feature set samples are input into the initial support vector machine to obtain the corresponding prediction probabilities. A sample selection strategy is then used to construct multiple candidate feature set samples. The multiple candidate feature set samples are clustered to determine each cluster center. Based on the prediction target, labels are obtained for each cluster center indicating whether the sedimentary droplets are normal or whether they are abnormal in a certain feature dimension, in order to further train the initial support vector machine, resulting in a final support vector machine used to achieve the prediction target.

6. The intelligent screening method for arrayed nozzles in inkjet printing according to claim 5, characterized in that, When training the support vector machine in the classification model, the following active learning sample selection strategy is adopted: ； In the formula, X is a sample set composed of the remaining feature set samples. and For the initial support vector machine to sample Determine the two highest probability values ​​obtained.

7. The intelligent screening method for arrayed nozzles in inkjet printing according to claim 5, characterized in that, The clustering method used in training the support vector machine in the classification model is the k-means algorithm.

8. The intelligent screening method for arrayed nozzles in inkjet printing according to claim 1, characterized in that, The feature set includes the following feature dimensions: whether the number of deposited droplets is 0, whether the number of deposited droplets is greater than 1, whether there is a landing point deviation of the deposited droplets, and whether the diameter of the deposited droplets exceeds the threshold.

9. A smart screening device for arrayed nozzles in inkjet printing, characterized in that, include: Inkjet printing module, droplet deposition substrate, droplet deposition observation module, motion module, and control module; The control module is used to control the inkjet printing module to move above the droplet deposition substrate for test spraying via the motion module, and to control the deposition droplet observation module to move to the droplet deposition area via the motion module, to acquire images of the deposition droplets ejected from each nozzle and transmit them to the control module; the control module is also used to detect the feature set of the deposition droplets corresponding to each nozzle based on the deposition droplet images, and to execute the intelligent screening method for arrayed nozzles for inkjet printing as described in any one of claims 1 to 8.

Images