A method for constructing an electrohydrodynamic jet drop volume prediction model and a drop volume prediction method

Abstract

The application belongs to the field of electrohydrodynamic jet printing, and particularly relates to a method for constructing an electrohydrodynamic jet printing droplet volume prediction model and a droplet volume prediction method, which comprises the following steps: constructing a training sample set, each training sample comprising process parameter features and corresponding image features of the same dimension; iteratively training a neural network using the training sample set to obtain an electrohydrodynamic jet printing droplet volume prediction model; in each iteration training, firstly, according to the feature quantity ratio between the current process parameter features and corresponding image features obtained by training optimization, the process parameter features and corresponding image features in each training sample are compressed or expanded in the condition of keeping the dimension unchanged, respectively, the compressed or expanded two kinds of features are spliced and fused, the spliced and fused features are taken as network input, and the feature quantity ratio is optimized and updated after each iteration training. The application can balance the accuracy and real-time performance in electrohydrodynamic jet printing droplet volume prediction.

Description

Technical Field

[0001] This invention belongs to the field of electrohydraulic inkjet printing, and more specifically, relates to a method for constructing an electrohydraulic inkjet printing droplet volume prediction model and a method for predicting droplet volume. Background Technology

[0002] Electrohydraulic inkjet printing is a branch of inkjet printing technology. The driving force generated by its electric field is far greater than the extrusion force generated by expansion, enabling solution printing over a wider viscosity range (1–10000 cPs). Furthermore, because droplets form at the tip of the Taylor cone, the resulting droplet diameter is much smaller than the nozzle diameter, allowing for the printing of ultra-high resolution structures at the micrometer and even nanometer scales. It has attracted widespread research attention in the fabrication of ultra-high resolution devices such as microlens arrays, biomimetic compound eyes, and display device repair.

[0003] The mechanism of electro-ink printing is complex and influenced by numerous factors. The volume of the ejected droplets is primarily affected by process parameters (voltage, inlet flow rate), printing material properties (surface tension, viscosity, density, etc.), and structural parameters (nozzle inner diameter, nozzle-to-substrate height), among others. These influences are mostly non-linear, making it difficult to theoretically deduce the relationship between droplet volume and process parameters. Conventional inkjet printing methods can capture the volume of airborne droplets using high-speed cameras. However, electro-ink printing involves high-speed (5 m / s), small (fl-level), and narrow (micrometer-level) droplets, making camera capture challenging. Expensive equipment such as phase-Doppler particle size distribution (PDD) velocimetry is also impractical. Therefore, accurately obtaining droplet volume remains the biggest challenge in electro-ink printing.

[0004] Research reveals that current technologies for predicting droplet volume largely rely on manual observation of the spread of filled pixel pits to infer the volume of the printed droplet, or employ an end-to-end data-driven approach from process parameters to droplet volume. Conventional prediction models suffer from low accuracy when dealing with complex fluid dynamics conditions, making online prediction impossible. Balancing the need for online prediction with high accuracy has become the biggest challenge in this field. Summary of the Invention

[0005] To address the shortcomings and improvement needs of existing technologies, this invention provides a method for constructing a predictive model for the volume of electro-hydraulic inkjet printing droplets and a method for predicting droplet volume. The purpose is to propose a predictive method for the volume of electro-hydraulic inkjet printing droplets that balances high accuracy and real-time performance.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for constructing a predictive model for the volume of electrohydraulic inkjet printing droplets is provided, comprising:

[0007] Construct a training sample set, wherein each training sample includes: process parameter features and their corresponding image features and droplet volume, and the process parameter features and image features have the same dimension;

[0008] Using the training sample set, a neural network is iteratively trained to obtain a liquid droplet volume prediction model for electrohydraulic inkjet printing. The neural network's input is multi-source features, and its output is the droplet volume. During each training iteration, based on the feature ratio between the current process parameter features and their corresponding image features obtained through training optimization, the number of features in each training sample is compressed or expanded while maintaining the same dimensionality. The compressed or expanded features are then concatenated and fused. The resulting feature is used as the multi-source feature, and the feature ratio is optimized and updated after each training iteration.

[0009] Furthermore, the process parameter features are constructed as follows: the process parameter dataset composed of various inkjet printing process parameter values ​​is normalized and standardized to obtain a one-dimensional data vector, which is used as the process parameter features.

[0010] Normalization process:

[0011]

[0012] In the formula, x represents the original data, x min x is the minimum value of the sample data in the process parameter dataset. max The maximum value of the sample data in the process parameter dataset;

[0013] Standardization process:

[0014]

[0015] In the formula, x nor To normalize the data, μ is the mean of the data, and σ is the standard deviation of the data.

[0016] The image features are constructed as follows: a Taylor cone image is acquired, features are extracted from the Taylor cone image to obtain Taylor cone image features, and the Taylor cone image features are transformed into a one-dimensional vector as the image features.

[0017] Furthermore, the various printing process parameters include: liquid supply pressure, power supply waveform, and motion parameters, which are obtained using the following methods:

[0018] (a) A high-precision air pressure pump system is used to record the liquid supply pressure at the tip of the needle in real time during the printing process, and the liquid supply pressure value P at the moment of ejection is obtained from it;

[0019] (b) A precision voltage sensor system is used to record the power supply waveform at the tip of the needle in real time during the printing process, thereby obtaining the power supply waveform at the moment of ejection, wherein the power supply waveform includes the bias voltage V. b Amplitude voltage V a Frequency V f and duty cycle V d ;

[0020] (c) A laser interferometer system is used to record the motion parameters of the needle tip in real time during the printing process, and the motion parameters at the moment of ejection are obtained from it. The motion parameters include the needle tip height h and the motion speed v.

[0021] Furthermore, the Taylor cone image is the last frame of the captured droplet before it falls, and it is obtained in the following way:

[0022] A high-speed camera module is used to focus on the Taylor cone tip and observe the tip ejection state in real time. The high-speed camera takes pictures at the same time as the high-voltage power supply induces the droplets to fall, and a set of pictures of the droplets falling is taken. The Naive Bayes image classification algorithm is used to automatically select the last Taylor cone image before the droplets fall.

[0023] Furthermore, a deep learning VGG16 network is used to extract features from the Taylor cone image.

[0024] Furthermore, the feature extraction is implemented as follows:

[0025] (a) Perform grayscale processing on the last frame of the Taylor cone image before the captured droplet falls;

[0026] (b) Input the grayscale image into a deep learning VGG16 network with an image attention mechanism, and extract the image after the last pooling layer as the Taylor cone image feature to complete the feature extraction. The image attention mechanism enables the deep learning VGG16 network to focus on the tip of the Taylor cone. The deep learning VGG16 network contains 13 convolutional layers, 5 pooling layers and 3 fully connected layers.

[0027] Furthermore, the splicing and fusion method is direct splicing, wherein the splicing weight is achieved by expanding or compressing the process parameter features and image features.

[0028] Furthermore, the neural network is a multilayer perceptron neural network; principal component analysis is used to perform the dimensionality increase / decrease operation.

[0029]

[0030] In the formula, m is the total amount of data in the feature to be upgraded or downgraded, and x is... iLet x be the i-th data point before dimensionality reduction. approx For data mapped to the target dimension, t is the preset amount of information to retain.

[0031] This invention also provides a method for predicting the volume of droplets in electrohydrodynamic printing, comprising:

[0032] Based on the training sample construction method in the above-described electrohydraulic inkjet printing droplet volume prediction model construction method, the current inkjet printing process parameter features and their corresponding image features are obtained.

[0033] Based on the optimal feature ratio obtained by the above-described method for constructing a current-current-printing droplet volume prediction model, the obtained process parameter features and their corresponding image features are compressed or expanded while maintaining the same dimension. The compressed or expanded features are then spliced ​​and fused. The spliced ​​and fused multi-source features are input into the current-current-printing droplet volume prediction model to predict the current-current-printing droplet volume.

[0034] The present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed by a processor, it controls the device where the storage medium is located to execute a method for constructing a current-current inkjet printing droplet volume prediction model as described above and / or a current-current inkjet printing droplet volume prediction method as described above.

[0035] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:

[0036] (1) This invention proposes a novel method for predicting droplet volume in electrohydraulic inkjet printing. First, training samples are constructed, each including process parameter features and their corresponding image features. The process parameter features and image features have the same dimension. When training the network model, the process parameter features and image features are first spliced ​​and fused. The resulting multi-source features are used as input to the neural network to comprehensively utilize information from both process parameters and droplet morphology during prediction, ensuring accuracy. Furthermore, this embodiment uses the feature dimension (i.e., the feature ratio) of the two features during splicing and fusion as the optimization update quantity. During iterative training, it is trained and optimized together with the parameters of the neural network to fully learn the contribution ratio of the two feature data to the prediction accuracy, aiming to ensure prediction accuracy. In addition, process parameters can be directly retrieved, while image acquisition and feature extraction can be performed online in real time. Therefore, compared with traditional electrohydraulic inkjet printing droplet volume prediction methods, this embodiment can effectively improve the efficiency and accuracy of droplet volume prediction, avoid waste of manpower and resources, meet the needs of online prediction, and ensure high prediction accuracy. It should be further noted that the selected features of this invention include process parameters, which can retain the influence of process parameters on droplet volume on the basis of image prediction, and can provide a window for subsequent droplet volume control, which meets the actual printing needs and has high application value in the field of electrohydraulic inkjet printing.

[0037] (2) The present invention proposes a method for constructing process parameter features by directly normalizing and standardizing the process parameter dataset composed of various inkjet printing process parameter values. The resulting one-dimensional vector can be used as process parameter features for subsequent training, which is convenient, fast, and can also ensure accuracy. At the same time, it is easy to control the droplet volume in the future. Attached Figure Description

[0038] Figure 1 A schematic diagram of a method for constructing a predictive model for the volume of electrofluid inkjet printing droplets, provided in an embodiment of the present invention;

[0039] Figure 2 A schematic diagram of the prediction architecture of the electrovolute droplet volume prediction method based on multi-source data fusion provided in an embodiment of the present invention;

[0040] Figure 3 A schematic diagram of the feature-level fusion method for the electrovolute droplet volume prediction method based on multi-source data fusion provided in this embodiment of the invention;

[0041] Figure 4 This is a data collection experiment diagram for the electrofluid droplet volume prediction method based on multi-source data fusion provided in an embodiment of the present invention.

[0042] Figure 5 This is an overall framework diagram of a method for predicting the volume of electrofluid inkjet printing droplets provided in an embodiment of the present invention. Detailed Implementation

[0043] 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.

[0044] Example 1

[0045] A method for constructing a predictive model for droplet volume in electrohydraulic inkjet printing, such as... Figure 1 As shown, it includes:

[0046] Construct a training sample set, wherein each training sample includes: process parameter features and their corresponding image features and droplet volume, and the process parameter features and image features have the same dimension;

[0047] Using the training sample set, a neural network is iteratively trained to obtain a liquid droplet volume prediction model for electrohydraulic inkjet printing. The neural network's input is multi-source features, and its output is the droplet volume. During each training iteration, based on the feature ratio between the current process parameter features and their corresponding image features obtained through training optimization, the number of features in each training sample is compressed or expanded while maintaining the same dimensionality. The compressed or expanded features are then concatenated and fused. The resulting feature is used as the multi-source feature, and the feature ratio is optimized and updated after each training iteration.

[0048] In current industrial production, the measurement of droplet volume in electrofluids typically involves using a white-light interferometer. After printing, the droplet is moved to the image features of the white-light interferometer to measure its morphology and calculate its volume. However, this method is time-consuming, slowing down the overall printing process. Conventional inkjet printing can use a high-speed camera to capture airborne droplets and obtain their size, but electrofluids have high droplet velocity (5 m / s), small size (fl level), and narrow space (micrometer level), making camera capture difficult. Furthermore, data-driven methods can predict droplet volume, but conventional prediction models cannot perform online predictions when dealing with the complex operating conditions of electrofluids. In summary, current technologies cannot simultaneously meet the needs of online prediction while ensuring high prediction accuracy.

[0049] This embodiment proposes a novel method. First, training samples are constructed, each including process parameter features and their corresponding image features. The process parameter features and image features have the same dimensionality. When training the network model, the process parameter features and image features are first concatenated and fused. The resulting multi-source features are used as input to the neural network to comprehensively utilize information from both process parameters and droplet morphology during prediction, ensuring accuracy. Furthermore, this embodiment uses the feature dimension (i.e., the feature ratio) of the two features during concatenation and fusion as a value for optimization and update. During iterative training, this value is trained and optimized along with the neural network parameters to fully learn the contribution ratio of the two feature data to prediction accuracy. Additionally, process parameters can be retrieved directly, and image acquisition and feature extraction can be performed online in real time. Therefore, compared to traditional electrohydraulic inkjet printing droplet volume prediction methods, this embodiment can effectively improve the efficiency and accuracy of droplet volume prediction, avoid waste of manpower and resources, meet the needs of online prediction, and ensure high prediction accuracy, making it highly valuable in the field of electrohydraulic inkjet printing.

[0050] In practice, for a fixed solution and electro-hydraulic nozzle, the printing process parameters are adjusted to achieve a stable spraying state within an appropriate range of process parameters. During the printing process, the process parameters, Taylor cone images during spraying, and the corresponding droplet volumes are collected.

[0051] Among them, such as Figure 2 As shown, preferably, the process parameters may include the liquid supply pressure P. n The power supply waveform and motion parameters are as follows: the power supply waveform is mainly a square wave, including the bias voltage V. b Amplitude voltage V a Frequency V f and duty cycle V d The motion parameters are needle tip height h and motion speed v. The process parameter dataset D can be represented as {D=(P nk V bk V ak V fk V dk ,h k ,v k )} k＝1,2…N Where the subscript N represents the number of process parameter groups, i.e., the total number of training samples. Preferably, a high-precision air pump system is used to record the liquid supply pressure at the tip of the needle in real time during the printing process, obtaining the liquid supply pressure value P at the instant of ejection; a precision voltage sensor system is used to record the power supply waveform at the tip of the needle in real time during the printing process, obtaining the power supply waveform at the instant of ejection, wherein the power supply waveform includes the bias voltage V. b Amplitude voltage V a Frequency V f and duty cycle Vd A laser interferometer system is used to record the motion parameters of the needle tip in real time during the printing process, thereby obtaining the motion parameters at the moment of ejection, including the needle tip height h and the motion speed v.

[0052] Additionally, Taylor cone images are captured during the spraying process. These images can be captured by a high-speed camera and are preferred. The last frame image before the droplet falls is selected for model training. Furthermore, the droplet volume can be measured using a white light interferometer and represented by the dataset E, where {E=(V k )} k＝1,2…N .

[0053] For the construction of process parameter features and image features, the preferred implementation method is as follows: the process parameter features are constructed by normalizing and standardizing the process parameter dataset composed of various inkjet printing process parameter values ​​to obtain a one-dimensional data vector, which is used as the process parameter features; the image features are constructed by acquiring a Taylor cone image, extracting features from the Taylor cone image to obtain Taylor cone image features, and converting the Taylor cone image features into a one-dimensional vector, which is used as the image features.

[0054] More specifically, process parameter characteristics can be directly applied to data analysis. Through normalization and standardization, data features can be extracted to obtain the process parameter characteristics {T=(P nk V bk V ak V fk V dk ,h k ,v k )} k＝1,2…N .

[0055] Normalization process:

[0056]

[0057] In the formula, x represents the original data, x min x is the minimum value of the sample data in the process parameter dataset. max The maximum value of the sample data in the process parameter dataset;

[0058] Standardization process:

[0059]

[0060] In the formula, x nor To normalize the data, μ is the mean of the data, and σ is the standard deviation of the data.

[0061] For image features, a deep learning VGG16 network can be preferably used to extract features from Taylor cone images. For example... Figure 3As shown, as a further preferred implementation, grayscale processing is performed on the last frame of the image before the droplet falls to reduce noise and interference from color information and speed up image processing. During feature extraction, the focus is on the tip of the Taylor cone, and an image attention mechanism is added to enable the neural network to adaptively focus on the break of the Taylor cone, thereby improving the accuracy of image processing. The deep learning VGG16 network contains 13 convolutional layers, 5 pooling layers, and 3 fully connected layers. Data is extracted after the last pooling layer to ensure accurate image parsing, ultimately resulting in an image feature of size 7×7×512, which is then transformed into a vector of size 1×25088. This vector array is the aforementioned image feature.

[0062] For multi-source feature acquisition, such as Figure 3 As shown, more specifically, the acquisition method is as follows: using a feature-level fusion method (preferably principal component analysis), the process parameter features and image features in each training sample are fused. Based on the currently optimized feature quantity ratio, the number of features in each process parameter feature and its corresponding image feature is adjusted so that the process parameter features and image features with adjusted feature quantities satisfy the currently optimized feature quantity ratio after stitching. The feature quantity ratio is dynamically adjusted according to the feedback results of subsequent predictions to select the optimal ratio. The stitched multi-source feature is actually a one-dimensional linearly uncorrelated quantity.

[0063] It should be noted that the process parameter features and their corresponding image features in each training sample are compressed or expanded while keeping the dimension unchanged. The dimension can refer to the number of one-dimensional vectors.

[0064] For the obtained high-dimensional image features, principal component analysis is used to compress or expand the number of features, where: mean squared projection error: For data mapped to the target dimension; total changes in the data: Based on formula In the formula, m is the total amount of data in the feature to be upgraded or downgraded, and x is... i Let x be the i-th data point before dimensionality reduction. approx For the data mapped to the target dimension, t is the preset amount of information to retain. The value of t is 0.99, which means that the PCA algorithm retains 99% of the main information to ensure good accuracy after feature dimensionality reduction.

[0065] The aforementioned neural network to be trained is preferably a multilayer perceptron neural network (MLP). Specifically, its training method can be as follows: a set of training samples includes a multi-source feature set M that has fused process parameter features and image features, and a droplet volume set E. A network structure is constructed that includes m input layer neurons, n output layer neurons, and q hidden layer neurons, with the sigmoid function as the activation function.

[0066] Assuming the input layer is represented by vector X, the output of the hidden layer is f(W1x+b1), where W is the weight and b is the bias.

[0067] Finally, there is the output layer, where the process from the hidden layer to the output layer can be viewed as a multi-class logistic regression, with the output being softmax(W2X1+b2), where X1 represents the output of the hidden layer f(W1x+b1). The overall MLP model is shown below:

[0068] f(x)=S(b (2) +W (2) (s(b (1) +W (1) x)))

[0069] The mapping relationship between ink droplet volume and multi-source features is expressed as follows:

[0070]

[0071] Among them, y j Let x represent the j-th output value. i Let v represent the i-th input. ih w represents the network connection weights between input layer neuron i and hidden layer neuron h. hj θ represents the network connection weights between hidden layer neuron h and output layer neuron j. j γ represents the threshold of neuron j in the output layer. h This represents the threshold of the hidden layer neuron h.

[0072] The formula for the convolution operation in the l-th layer is as follows:

[0073]

[0074]

[0075] Max pooling formula for layer l:

[0076]

[0077] In the formula, n l The number of convolutional kernels in the l-th layer. The convolution kernels corresponding to the p-channel of the l-th layer and the q-channel of the (l-1)-th layer are... For the bias of node p in layer l, W l Let z be the weight of the l-th fully connected network. l For the forward input of the l-th layer that has not undergone the activation function, a l This is the forward output of the l-th layer after the activation function.

[0078] During MLP training, the weights and biases of the model are updated by minimizing the mean squared error loss function through backpropagation to achieve accurate prediction of droplet volume, while outputting the coefficient of determination R. 2 To judge the quality of the model, R is required. 2 A score of ≥0.9 indicates that the model training results are good.

[0079]

[0080]

[0081] In the formula Y i For the true value, For the predicted values, the basic structure of the overall model is as follows: Figure 3 As shown.

[0082] In an online printing system, for the process of a droplet falling, the process parameters are composed of the liquid supply pressure, bias voltage, amplitude voltage, frequency, duty cycle, tip height, and movement speed captured by the sensor at that moment. The image of the frame before the droplet falls is selected as the corresponding Taylor cone image of the droplet at that moment. The process parameters and the Taylor cone image are fed into a multi-source data fusion model to obtain a fusion vector, which is automatically fed into a trained MLP for training. The overall data acquisition method is as follows: Figure 4 As shown.

[0083] Example 2

[0084] A method for predicting the droplet volume in electrohydraulic inkjet printing includes:

[0085] According to the training sample construction method in the construction method of the electro-hydraulic inkjet printing droplet volume prediction model as described in Example 1, the process parameter features used for inkjet printing and their corresponding image features are obtained.

[0086] Based on the optimal feature quantity ratio obtained by the construction method of the electro-hydraulic inkjet printing droplet volume prediction model as described in Example 1, and the electro-hydraulic inkjet printing droplet volume prediction model, the obtained process parameter features and their corresponding image features are compressed or expanded respectively while keeping the dimension unchanged. The two types of features after compression or expansion are spliced ​​and fused. The spliced ​​and fused multi-source features are input into the electro-hydraulic inkjet printing droplet volume prediction model to predict the electro-hydraulic inkjet printing droplet volume.

[0087] In actual implementation, the system automatically collects Taylor cone images and process parameters during the spraying process. For a droplet falling process, the Taylor cone image of the frame before the droplet falls is selected as the corresponding Taylor cone image of the droplet at this time. The process parameters are composed of the liquid supply pressure, bias voltage, amplitude voltage, frequency, duty cycle, needle tip height, and movement speed captured by the sensor at this time. After feature extraction, feature compression or expansion and splicing and fusion, the Taylor cone image and process parameters are fed into the pre-trained model to predict the droplet volume.

[0088] The relevant technical solutions are the same as in Embodiment 1, and will not be repeated here.

[0089] In general, such as Figure 5 As shown, the method of the present invention includes three aspects: collecting data during printing, training a model, and predicting the droplet volume using the trained model, the process parameters during spraying, and the Taylor cone image for the droplet volume to be measured. Specifically, it includes: (1) collecting process parameters, Taylor cone images during spraying, and corresponding droplet volumes for a certain printing solution; (2) using a deep learning network to extract features from the Taylor cone image, forming an image feature extraction method; (3) using a feature-level fusion method to fuse process parameter data and Taylor cone image features, forming multi-source data features that fuse process parameters and images; (4) using a neural network to train and learn the multi-source data features and the corresponding droplet volume, establishing a neural network model, and inputting data for training to meet the prediction accuracy; (5) during online printing, automatically collecting process parameters and Taylor cone images during spraying, inputting them into the model to obtain the predicted droplet volume. The present invention can quickly predict the droplet volume during electrohydraulic printing, meeting the needs of online printing.

[0090] Example 3

[0091] A computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed by a processor, it controls the device where the storage medium is located to execute a method for constructing a hydrostatic printing droplet volume prediction model as described in Embodiment 1 above and / or a method for predicting hydrostatic printing droplet volume as described in Embodiment 2 above.

[0092] The relevant technical solutions are the same as those in Embodiment 1 and Embodiment 2, and will not be repeated here.

[0093] 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 method for constructing a predictive model for droplet volume in electrohydraulic inkjet printing, characterized in that, include: Construct a training sample set, wherein each training sample includes: process parameter features and their corresponding image features and droplet volume, and the process parameter features and image features have the same dimension; Using the training sample set, a neural network is iteratively trained to obtain a liquid droplet volume prediction model for electrohydraulic inkjet printing. The neural network's input is multi-source features, and its output is the droplet volume. During each training iteration, based on the feature ratio between the current process parameter features and their corresponding image features obtained through training optimization, the number of features in each training sample is compressed or expanded while maintaining the same dimensionality. The compressed or expanded features are then concatenated and fused. The resulting feature is used as the multi-source feature, and the feature ratio is optimized and updated after each training iteration. Among them, various printing process parameters include: liquid supply pressure, power supply waveform, and motion parameters; The process parameter features are constructed by normalizing and standardizing the process parameter dataset composed of various inkjet printing process parameter values ​​to obtain a one-dimensional data vector, which is used as the process parameter features. The image features are constructed as follows: a Taylor cone image is acquired, features are extracted from the Taylor cone image to obtain Taylor cone image features, and the Taylor cone image features are transformed into a one-dimensional vector as the image features.

2. The construction method according to claim 1, characterized in that, The normalization process: In the formula, x The original data, x min The minimum value of the sample data in the process parameter dataset. x max The maximum value of the sample data in the process parameter dataset; The standardization process: In the formula, To normalize the data, μ The mean of the data. denoted as the standard deviation of the data.

3. The construction method according to claim 2, characterized in that, The various printing process parameters were obtained using the following methods: (a) A high-precision air pressure pump system is used to record the liquid supply pressure at the tip of the needle in real time during the printing process, and the liquid supply pressure value P at the moment of ejection is obtained from it; (b) A precision voltage sensor system is used to record the power supply waveform at the tip of the needle in real time during the printing process, thereby obtaining the power supply waveform at the moment of ejection, wherein the power supply waveform includes the bias voltage. V b Amplitude voltage V a ,frequency V f and duty cycle V d ; (c) A laser interferometer system is used to record the motion parameters of the needle tip in real time during the printing process, thereby obtaining the motion parameters at the moment of ejection, wherein the motion parameters include the needle tip height. h and speed of movement v .

4. The construction method according to claim 2, characterized in that, The Taylor cone image is the last frame of the captured droplet before it falls, and it is obtained in the following way: A high-speed camera module is used to focus on the Taylor cone tip and observe the tip ejection state in real time. The high-speed camera takes pictures at the same time as the high-voltage power supply induces the droplets to fall, and a set of pictures of the droplets falling is taken. The Naive Bayes image classification algorithm is used to automatically select the last Taylor cone image before the droplets fall.

5. The construction method according to claim 2, characterized in that, The Taylor cone image was feature extracted using a deep learning VGG16 network.

6. The construction method according to claim 5, characterized in that, The feature extraction is implemented as follows: (a) Perform grayscale processing on the last frame of the Taylor cone image before the captured droplet falls; (b) Input the grayscale image into a deep learning VGG16 network with an image attention mechanism, and extract the image after the last pooling layer as the Taylor cone image feature to complete the feature extraction. The image attention mechanism enables the deep learning VGG16 network to focus on the tip of the Taylor cone. The deep learning VGG16 network contains 13 convolutional layers, 5 pooling layers and 3 fully connected layers.

7. The construction method according to claim 1, characterized in that, The splicing and fusion method is direct splicing, where the splicing weight is achieved by expanding or compressing the process parameter features and image features.

8. The construction method according to claim 1, characterized in that, The neural network is a multilayer perceptron neural network; principal component analysis is used to perform dimensionality adjustment and dimensionality reduction operations. In the formula, m The total amount of data in the features to be upgraded or downgraded. The first before dimensionality reduction i One data point, x approx For data mapped to the target dimension, t The preset amount of information to retain.

9. A method for predicting the volume of droplets in electrohydraulic inkjet printing, characterized in that, include: According to the training sample construction method in the construction method of the electro-hydraulic inkjet printing droplet volume prediction model as described in any one of claims 1 to 8, the process parameter features used for inkjet printing and their corresponding image features are obtained. Based on the optimal feature quantity ratio obtained by the construction method and the electro-hydraulic inkjet droplet volume prediction model, the obtained process parameter features and their corresponding image features are compressed or expanded while keeping the dimension unchanged. The two types of features after compression or expansion are spliced ​​and fused. The spliced ​​and fused multi-source features are input into the electro-hydraulic inkjet droplet volume prediction model to predict the electro-hydraulic inkjet droplet volume.

10. 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, it controls the device where the storage medium is located to perform a method for constructing a hydroelectric inkjet printing droplet volume prediction model as described in any one of claims 1 to 8 and / or a method for predicting hydroelectric inkjet printing droplet volume as described in claim 9.

Images