Skin-feel UV ink generation method based on deep learning and skin-feel UV ink

Abstract

The present invention relates to technical fields such as deep learning and ink production, and provides a skin-feel UV ink generation method based on deep learning and skin-feel UV ink. The method obtains raw material composition data and performance parameter data and inputs them into the input layer corresponding to the deep learning model respectively. The deep learning model is iteratively trained by a back propagation algorithm, so that the trained model can output an optimal raw material feeding ratio. The optimal raw material feeding ratio is applied to production, including feeding and mixing, dispersing machine stirring, and adding auxiliary agents. After stirring is completed, a UV curing test is performed. The deep learning model is updated again according to the test feedback, and finally a skin-feel UV ink with expected curing speed, adhesion and feel performance is generated, thereby improving the efficiency and accuracy of analyzing the multi-parameter coupling relationship of the skin-feel UV ink and avoiding inconsistent ink batches and performance fluctuations.

Description

Technical Field

[0001] The present invention relates to technical fields such as deep learning and ink production, and in particular to a skin-feeling UV ink generation method based on deep learning and skin-feeling UV ink. Background Art

[0002] With the continuous advancement of printing technology, UV inks have gained widespread application in various high-quality packaging and specialty printing applications. However, traditional UV ink formulations often require a balanced mix of multiple resins, additives, and photoinitiators to meet multi-dimensional performance requirements, including cure speed, adhesion, abrasion resistance, and tactile feel. In particular, applications such as cigarette packaging and cosmetics packaging, where both tactile and visual effects are highly demanded, often require additional "skin feel" enhancements. These include adding specific ratios of matte resins, matte-resistant resins, and tactile resins. However, existing processes often rely on manual experience or limited trial-and-error processes to determine the proportions of these raw materials, which can lead to insufficient control of key parameters in complex formulations and make it difficult to achieve a consistent balance of fast cure, high adhesion, and a pleasant feel. Any deviation in a particular formulation ratio can result in inadequate cure, reduced adhesion, substandard abrasion resistance, or even unpleasant tactile feel. Furthermore, with increasing customer demands, repeated adjustments or small-batch customization are increasingly necessary to accommodate the appearance and functional requirements of different products. Traditional methods are often costly, time-consuming, and lack flexibility and precision. Given this situation, how to use intelligent methods to rapidly iteratively optimize the feed ratios of multiple raw materials has become a pressing technical challenge. On the one hand, it is necessary to extract feasible feed ratio patterns from a large amount of historical formula and performance test data to avoid blind manual experimentation. On the other hand, it is necessary to achieve a comprehensive balance between objectives such as curing speed, adhesion, abrasion resistance, and feel scores to accommodate different performance requirements. Under traditional formula adjustment methods, it is difficult to accurately analyze the multi-parameter coupling relationship manually in a short period of time, which can easily lead to misjudgment or over-adjustment, resulting in inconsistent ink batches and performance fluctuations. Summary of the Invention

[0003] In response to the above-mentioned shortcomings of the existing technology, the present invention provides a skin-feeling UV ink generation method based on deep learning to improve the efficiency and accuracy of analyzing the multi-parameter coupling relationship of skin-feeling UV ink, thereby avoiding ink batch inconsistency and performance fluctuations.

[0004] In a first aspect, the present invention provides a method for generating skin-feel UV ink based on deep learning, comprising:

[0005] Obtain raw material composition data and performance parameter data of skin-feel UV ink, wherein the raw material composition data includes the weight percentages of polyester matte resin, polyurethane easy-matting resin, skin-feel resin, photoinitiator, slip agent, and wax agent; and the performance parameter data includes curing speed, adhesion level, abrasion resistance times, and hand feel score;

[0006] Establishing a deep learning model of a multi-layer neural network, inputting the raw material composition data and the performance parameter data into the corresponding input layers of the deep learning model respectively; iteratively training the deep learning model through a back propagation algorithm so that the trained model can output the optimal raw material feed ratio;

[0007] The optimal raw material ratio is applied to production, including mixing, stirring in a disperser, and adding additives. After stirring, a UV curing test is carried out. The deep learning model is updated again based on the test feedback, and finally a skin-feel UV ink with the expected curing speed, adhesion, and feel is generated.

[0008] In a second aspect, the present invention provides a skin-feeling UV ink, which is produced using the above-mentioned skin-feeling UV ink generation method based on deep learning.

[0009] Compared with the prior art, the present invention has the following beneficial effects:

[0010] The present invention provides a skin-feel UV ink generation method and skin-feel UV ink based on deep learning. Raw material composition data and performance parameter data of the skin-feel UV ink are obtained, wherein the raw material composition data include weight percentages of polyester matte resin, polyurethane easy-matting resin, skin-feel resin, photoinitiator, slip additive and wax additive, and the performance parameter data include curing speed, adhesion level, wear resistance times and hand feel score. A deep learning model of a multi-layer neural network is established, and the raw material composition data and the performance parameter data are respectively input into corresponding input layers of the deep learning model. The deep learning model is iteratively trained by a back propagation algorithm so that the trained model can output an optimal raw material feeding ratio. The optimal raw material feeding ratio is applied to production, including feeding and mixing, dispersing machine stirring and adding additives. After stirring, a UV curing test is performed, and the deep learning model is updated again according to the test feedback. Finally, a skin-feel UV ink with expected curing speed, adhesion and hand feel performance is generated, thereby improving the efficiency and accuracy of analyzing the multi-parameter coupling relationship of the skin-feel UV ink and avoiding ink batch inconsistency and performance fluctuation. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their description are used to explain the present invention and do not constitute an undue limitation of the present invention. Some specific embodiments of the present invention will be described in detail in an illustrative and non-restrictive manner with reference to the drawings. The same reference numerals in the drawings indicate the same or similar components or parts. It should be understood by those skilled in the art that these drawings are not necessarily drawn to scale. In the drawings:

[0012] Figure 1 This is a flow chart of a method for generating skin-feel UV ink based on deep learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0013] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0014] Example 1

[0015] See also Figure 1 This embodiment provides a method for generating skin-feel UV ink based on deep learning, comprising the following steps:

[0016] S101. Obtain raw material composition data and performance parameter data of skin-feel UV ink, wherein the raw material composition data includes weight percentages of polyester matte resin, polyurethane easy matting resin, skin-feel resin, photoinitiator, slip agent, and wax additive, and the performance parameter data includes curing speed, adhesion level, number of abrasion resistance, and hand feel score; for example, the weight percentages of polyester matte resin, polyurethane easy matting resin, skin-feel resin, photoinitiator, slip agent, and wax additive meet the range of 40%-60% of polyester matte resin, 10%-25% of polyurethane easy matting resin, 20%-30% of skin-feel resin, 4%-5% of photoinitiator, 2%-5% of slip agent, and 2%-3% of wax additive;

[0017] S102, establishing a deep learning model of a multi-layer neural network, inputting the raw material composition data and the performance parameter data into the input layers corresponding to the deep learning model respectively; iteratively training the deep learning model using a back propagation algorithm so that the trained model can output an optimal raw material feed ratio;

[0018] S103. Apply the optimal raw material feeding ratio to production, including feeding and mixing, dispersing machine stirring, and adding additives. After stirring, perform UV curing test, update the deep learning model again based on the test feedback, and finally generate a skin-feel UV ink with the expected curing speed, adhesion, and feel.

[0019] It should be noted that the skin-feel UV ink generation method based on deep learning proposed in this embodiment can realize the automatic prediction and output of the optimal raw material feeding ratio through a multi-layer neural network model in a scenario where the ratio of multiple raw materials (polyester matte resin, polyurethane easy-to-matt resin, skin-feel resin, photoinitiator, slip additive, wax additive) is complex and the performance targets (curing speed, adhesion, wear resistance, feel score) affect each other. In traditional formula design, the ratio is often determined only by experience or limited experiments, and it is difficult to take into account multiple performance dimensions in a timely manner. In this embodiment, the use of deep learning models for iterative training can effectively extract the potential rules in a large amount of data. In this embodiment, by obtaining raw material composition data and performance parameter data, and inputting these data into a deep learning model (for example, a neural network), the back propagation algorithm is continuously iterated until the model can output the optimal formula, and then after verification by actual production and UV curing tests, it is updated again to achieve closed-loop optimization of skin-feel UV ink. In this embodiment, the entire process from data collection and modeling, to training and output, and then to actual production and feedback solves the recipe uncertainty caused by multi-parameter coupling, thereby improving the efficiency and accuracy of recipe design.

[0020] In some preferred embodiments, before obtaining raw material composition data and performance parameter data, a sample collection subsystem is further included. The sample collection subsystem conducts experimental measurements on multiple batches of skin-feel UV inks to obtain accurate experimental measurement data. The experimental measurement data includes the corresponding curing speed, adhesion level, number of wear resistance, and feel score of polyester matte resin, polyurethane easy-matting resin, skin-feel resin, photoinitiator, slip agent, and wax agent at different ratios. During the experimental measurement, unified test standards are used for measurement and temperature and humidity are stably controlled. The measured data is transmitted to a database for summary and collation. The obtained collated experimental measurement data is input into the training set and validation set of the deep learning model to improve the generalization ability and prediction accuracy of the model. After the deep learning model training is completed, the model parameters are fixed and the first round of optimal formula recommendations are output to facilitate the subsequent actual production verification and retraining stage. It should be noted that in this embodiment, the step of establishing a sample collection subsystem is added before obtaining raw material composition data and performance parameter data to conduct experimental measurements on multiple batches of skin-feel UV inks to form training sets and validation sets. The reason for this setting is that the performance of the deep learning model is closely related to the quality of the data used, and traditional manual experiments often lack systematic collection and management, resulting in incomplete or inaccurate model training data. In this embodiment, the experimental measurement data is collected and sorted in the database through unified test standards and stable control of temperature and humidity, and then input into the model to ensure the reliability and consistency of the data. At the same time, the model parameters are fixed and the first round of optimal formula suggestions are output. After the initial training is completed, the model has a certain generalization ability and prediction accuracy, and a preliminary formula is formed first to lay the foundation for subsequent retraining.

[0021] In some preferred embodiments, when establishing a deep learning model of a multi-layer neural network, an input layer is set to receive raw material composition data and performance parameter data, and at least three hidden layers are set to construct a feature extraction and nonlinear mapping structure; residual connections are introduced in the hidden layers to avoid the phenomenon of gradient disappearance or explosion when the network layer is deep; a multi-objective regression unit is set up in the output layer to simultaneously predict the curing speed, adhesion level, wear resistance and feel score of the skin-feel UV ink; a mean square error loss function is used in the training process, and L2 regularization is combined to constrain the weights to prevent overfitting; through training, the deep learning model can accurately predict the corresponding performance indicators according to the feed ratio, and realize intelligent optimization of the ratio of polyester matte resin, polyurethane easy matting resin, skin-feel resin, photoinitiator, slip additive and wax additive. It should be noted that in this embodiment, it is clear that when establishing a multi-layer neural network, at least three hidden layers need to be set, and residual connections are introduced in the hidden layers, and a multi-objective regression unit, mean square error loss function, L2 regularization and other means are combined. The reason for this setting is that there is often a high nonlinear correlation between the various performance indicators of skin-feel UV inks, and simple shallow network models are difficult to capture complex coupling relationships. At the same time, residual connections can effectively alleviate the gradient vanishing or explosion problems of deep networks and enhance the sustainability of training. The multi-target regression unit is designed in the output layer because it is necessary to simultaneously predict four indicators: curing speed, adhesion level, wear resistance and feel score, and then make a comprehensive trade-off. The use of mean square error loss function and L2 regularization can reduce the risk of overfitting and enable the model to maintain good generalization ability when facing various raw material ratios. Therefore, this embodiment clearly stipulates the network structure, training strategy and loss form, so as to solve the technical problem of how to extract key features in high-dimensional, multi-target ink formulas and ensure stable training, thereby enhancing the accuracy and reliability of model predictions.

[0022] In some preferred embodiments, during the mixing and stirring process of the feed machine, an online monitoring module is added to detect the stirring temperature, viscosity and speed in real time, and the detection results are input into the process monitoring sub-network in the deep learning model; the process monitoring sub-network compares the difference between the predicted viscosity and the actual viscosity in real time. If the deviation exceeds the threshold, the disperser speed is automatically adjusted or the additive addition speed is changed; after the slip additive and the wax additive are added, a layered stirring method is used for secondary mixing, and the curing efficiency is monitored by UV curing pre-test; finally, the measured data is sent back to the deep learning model for fine-tuning, so that the generated skin-feeling UV ink maintains consistency and controllability during the production process, avoiding formula inaccuracies caused by temperature changes or uneven stirring. It should be noted that an online monitoring module is added in this embodiment to input the real-time detected stirring temperature, viscosity and speed information into the process monitoring sub-network in the deep learning model, and automatically adjust the disperser speed or the additive addition speed when the deviation exceeds the threshold. The reason for this setting is that even under the same ratio conditions, changes in environmental factors such as temperature and humidity at the production site can cause the stirring process to be unstable, thereby causing fluctuations in ink performance. If we rely solely on offline detection, it is difficult to detect and correct deviations in a timely manner. In this embodiment, a process monitoring subnetwork is used to implement online closed-loop regulation. Once the detection results show obvious deviations, the production process parameters can be automatically intervened to maintain the formula adjustment within the target range. Through such online monitoring and feedback, the quality inconsistency caused by uneven stirring or temperature fluctuations can be significantly reduced, and the measured data can be fed back to the deep learning model for fine-tuning, forming a cycle of continuous learning and correction, thereby effectively solving the problem of inaccurate deep learning formulas caused by fluctuations in environmental factors on the production site, and significantly improving the industrial feasibility and robustness of the entire method.

[0023] In some preferred embodiments, during the deep learning model training process, four sub-goals are set: curing speed, adhesion level, wear resistance, and feel score. A weighted multi-objective optimization strategy is used, assigning each sub-goal a corresponding weight coefficient. Through iterative training, multiple rounds of testing are conducted on different weight coefficient configurations to screen out a comprehensive formula that meets both fast curing requirements and high adhesion and good feel. The weighted multi-objective optimization strategy generates a feasible solution set at the model output layer, and then selects the optimal solution based on pre-set target sorting rules. Based on this optimal solution, a set of recommended ratios of polyester matte resin, polyurethane easy-to-matt resin, skin-feel resin, photoinitiator, slip agent, and wax additive are obtained during the feeding process to balance the requirements of fast printing and durable feel. It should be noted that for skin-feel UV inks, indicators such as fast curing and high adhesion often conflict with each other, and balancing speed and performance is not easy. A multi-objective optimization strategy that uses each indicator as a sub-goal and gives it different weights helps to find an optimal solution set based on a comprehensive balance, and then screens out the final formula according to the pre-set priority, thereby preventing a certain indicator from being ignored and avoiding the blind pursuit of a certain goal that leads to a serious decline in other performance. This embodiment uses a weighted multi-objective optimization strategy to screen out a comprehensive formula that meets both rapid curing and high adhesion and good feel. It can solve the problem of mutual influence between multiple performance indicators, allowing the model to simultaneously evaluate and weigh the priorities of different indicators, thereby outputting the global optimal or approximately optimal feeding plan, enabling the deep learning model to adapt to diverse needs and meet the actual industrial production.

[0024] In some preferred embodiments, before outputting the optimal raw material feed ratio, a simulation verification phase is performed. The generated formula parameters are input into a simulation module with physical constraints. The simulation module calculates the theoretical curing efficiency and surface tension based on the known kinetic models of the chemical reaction and photocuring process of polyester matte resin, polyurethane matt resin, skin-feel resin, photoinitiator, slip agent, and wax additive. If the simulation results deviate from the predicted results of the deep learning model, the deviation information is fed back to the model for secondary training until the simulation results and the predicted results match within a preset error range. It should be noted that deep learning models can sometimes only capture statistical correlations at the data level. When encountering extreme cases or special conditions not covered by the training data, the model prediction may deviate. Introducing a simulation module for chemical and photocuring kinetics can review the rationality of the formula from a mechanistic perspective. If the simulation results deviate significantly from the model predictions, feedback is provided to the model for correction. This combination of data-driven and mechanistic models can effectively reduce blind spots and improve the credibility of formula predictions, while also saving the losses that may be caused by direct testing at the production site and solving the problem of uncontrolled deviations that may arise from relying solely on statistical models.

[0025] In some preferred embodiments, a feel scoring subnetwork is set in the deep learning model to separately extract key features related to the content of skin-feel resin and the type of slip additive; the feel scoring subnetwork receives feel scoring training samples and uses an attention mechanism to weight the main features that affect the feel; the output of the feel scoring subnetwork is combined with the remaining prediction results of the main model of the deep learning model to obtain a final score, so that the feed ratio can more accurately match the excellent tactile performance; during verification in production, a standardized blind touch evaluation process is adopted, a fixed number of evaluators are invited to score, and the scores are input into the feel scoring subnetwork for error backtracking correction again to achieve refined control of the skin feel effect. It should be noted that compared with quantitative measurement indicators such as curing speed or wear resistance, feel scoring is more subjective and diverse, and the connection between it and raw materials such as skin-feel resin and slip additive is more complex. A separate feel scoring subnetwork can focus more on tactile performance and avoid being overwhelmed by other indicators in the comprehensive model. The final score is obtained after being combined with the output of the main model. In production verification, it can be combined with blind touch evaluation, and the feedback score can be input into the sub-network for correction again, so that the model can achieve deep learning of the subjective evaluation of the feel, so that the feed ratio can be more accurately matched with the excellent touch, and the problem that the feel indicators are often ignored or difficult to quantify under multi-target prediction is solved, so that the final skin-feel UV ink is fully optimized in terms of tactile experience.

[0026] In some preferred embodiments, a chemical surface energy calculation model is incorporated into the adjustment of the ratio of slip additives to wax additives. This model estimates the impact of different additive combinations on the surface friction coefficient of printed cigarette packs. In training samples, high friction coefficients are associated with scratch rates, and a penalty term is established during training for friction coefficient prediction errors. When an additive combination poses a potential scratch risk, the chemical surface energy calculation model increases the penalty value of the additive combination at the output layer, incentivizing the final formulation to reduce the friction coefficient. It should be noted that in production practice, excessively high surface friction coefficients can lead to white scratches or scuffs on finished cigarette packs, impacting packaging quality. Slip additives and wax additives determine surface lubricity and anti-friction properties. By incorporating the chemical surface energy calculation model into a deep learning system, the performance of additive combinations in real-world usage scenarios can be more accurately assessed, addressing the issue of inappropriate scratch rates caused by previous empirical judgments. Furthermore, penalizing friction coefficient prediction errors can proactively favor more optimal combinations during training, thereby minimizing the risk of scratches while ensuring wear resistance.

[0027] In some preferred embodiments, an automatic learning rate adjustment mechanism is introduced during backpropagation training of raw material feed ratios. This mechanism dynamically adjusts the learning rate based on the convergence rate of the loss function during training. If the loss value decreases slowly over several iterations, the learning rate is automatically increased to accelerate convergence. If the loss value fluctuates or increases, the learning rate is reduced to avoid excessive updates. It should be noted that in deep learning, the learning rate determines the stride of parameter updates. If it is too large, model training may fluctuate or even fail to converge. If it is too small, the training iteration speed is too slow and may fall into a local minimum. Especially in chemical formulation scenarios, with a wide parameter space and numerous performance indicators, a suitable learning rate strategy is required to balance speed and stability. Automatic learning rate adjustment can adaptively increase or decrease the learning rate based on the training curve, avoiding the tedious and imprecise manual parameter adjustment. In this way, increasing the learning rate when the loss function decreases slowly can accelerate convergence; reducing the learning rate when the loss value fluctuates violently can stabilize the update stride. This can significantly reduce the number of iterations and improve the model's generalization ability, solving the overshoot or undershoot problems that may occur in traditional fixed learning rate training, making the deep learning process of feed ratio more efficient and reliable.

[0028] In some preferred embodiments, the deep learning-based skin-feel UV ink production method further includes configuring a time series-based model update module to periodically retrain the long-term collected feeding and printing data and compare it with the current model parameters. If the retrained model performs better on the validation set, the old parameters are replaced; if the two models converge, the current parameters are maintained. It should be noted that in actual production, raw material batches, environmental conditions, and even market demand may change over time, causing the existing model parameters to gradually become out of touch with reality. To maintain high prediction accuracy over the long term, it is necessary to periodically update the model. The time series-based model update module can perform retraining after collecting a certain amount of new production data and determine whether the new model is more suitable for the current environment. If so, the new model is adopted; otherwise, the current model is retained. This ensures that the deep learning model remains in sync with production practice and prevents parameter aging. Furthermore, the validation set can be used to detect outliers or erroneous measurements in the new data, preventing model degradation.

[0029] In some preferred embodiments, the adhesion level is taken as a key focus during the back-propagation process, and when the adhesion level range exceeds a predefined threshold, the deep learning model gives priority to increasing the precision adjustment of the ratio of polyurethane-containing easy-to-matt resin to polyester matte resin; if the adhesion training error is still high, additional constraints are introduced to limit the use of high-slip additives and high-wax additives to prevent excessive improvement of the feel and matte effect at the expense of adhesion performance. It should be noted that adhesion is one of the important indicators of whether UV ink can be firmly bonded to a specific substrate, but it often conflicts with smoothness and skin feel. If you pursue super smoothness and matte effects, it sometimes causes a decrease in adhesion. This embodiment triggers a priority strategy by setting an adhesion level range threshold during the training process. Once the model detects that the adhesion deviation is too large, it will constrain the main influencing factors (such as matte resin or slip additive), increase attention to adhesion, and ensure stable printing quality.

[0030] In some preferred embodiments, when measuring curing speed, a high-precision UV energy meter is used to transmit the actual irradiation energy, exposure time, and surface hardness test results to the training platform in real time. By comparing the deviation between the model-predicted and measured curing speeds, the gradient weight of the photoinitiator ratio in the output layer is adjusted incrementally. If the deviation persists, the deep learning model re-examines the photoinitiator ratio range in the raw material composition data to adjust the photoinitiator ratio to a more reasonable range in the next iteration. It should be noted that curing speed is a key property of UV ink, affecting both printing efficiency and determining the final performance of the ink layer. Relying solely on model predictions can easily lead to deviations from actual UV irradiation conditions. By incorporating a high-precision energy meter and surface hardness testing, a more comprehensive reflection of the actual curing conditions can be achieved, especially under UV lamps of varying intensities and wavelength distributions. If a mismatch between the photoinitiator addition amount and the actual curing speed is detected, corrections are made promptly during training or inference. This allows the model to more flexibly adapt to diverse UV equipment and lighting conditions, addressing the issue of inconsistent curing speeds and ensuring stable ink curing performance under different workshop environments or lighting sources.

[0031] In some preferred embodiments, a color matching subsystem specifically for cigarette packages is configured to calculate the possible optimal ink layer thickness based on the chromaticity values ​​and apparent gloss requirements given in the cigarette package design file, combined with the effects of polyester matte resin and polyurethane easy-matting resin on diffuse reflection of light after UV curing; the color matching subsystem fuses the calculated value with the optimal feed ratio predicted by the deep learning model to obtain a formula that can meet both matte appearance and accurate chromaticity; during actual silk-screen printing, the CI E Lab value of the sample is measured and compared with the target value, and the difference is fed back to the color matching subsystem to continue to correct the fusion result and achieve accurate restoration of the cigarette package color. It should be noted that cigarette package printing usually requires not only stable performance, but also matching the specific design in appearance, such as the chromaticity and glossiness of the brand logo. Matt resin and easy-matting resin will affect the light reflection characteristics, thereby affecting the color visual effect. This embodiment uses a separate color matching subsystem to compare color spaces (such as CIE Lab values) to help the model meet visual requirements. If a large difference from the target value is detected during actual screen printing, feedback is provided to the subsystem, and the fusion result is then corrected, thereby maintaining the matte effect while accurately presenting the specified color, solving the problem of the coupling effect between pigments and resins on appearance and color tone, and ensuring that the printed products can obtain consistent and expected color performance.

[0032] In some preferred embodiments, all collected raw material composition data and performance parameter data are subjected to feature selection and dimensionality reduction processing to remove redundant features that have very little impact on the four indicators of curing speed, adhesion level, wear resistance, and feel score; in this process, a feature selection algorithm based on information gain is used, combined with a dimensionality reduction algorithm based on principal component analysis, and the simplified key features are input into the deep learning model. It should be noted that although deep learning can process high-dimensional data, too many redundant or noisy features will lead to a decrease in training efficiency and even interfere with model convergence. At the same time, a large number of irrelevant or minimally influential features may increase the risk of overfitting. In this embodiment, by providing a feature selection + dimensionality reduction process, the model maintains high prediction accuracy and stability under a more streamlined data dimension. In this way, it can not only improve computational efficiency, but also strengthen the model's learning ability for truly key variables (such as resin ratio, photoinitiator dosage), avoid the negative effects of data expansion, and help to achieve a better balance between iteration speed and stability in industrial sites.

[0033] In some preferred embodiments, a linkage process of a programmable logic controller is established between a disperser and a UV curing device, and the programmable logic controller automatically performs feeding, stirring and adding auxiliary agents according to the feed ratio and the disperser speed suggestion instruction output by the deep learning model; after the addition is completed, the UV curing device is triggered by the programmable logic controller to perform a preliminary curing test. If the test result shows that the curing speed or adhesion level is not up to standard, the iterative update program of the model is called again, the feed ratio is adjusted and the feeding and stirring process is re-executed. It should be noted that even if the model can give a more accurate feed ratio, if the production link lacks automated linkage, the actual execution is prone to deviation due to human error or lagging operation. In the present embodiment, by a programmable logic controller (PLC), the model output and the device instruction can be seamlessly connected, so that feeding, mixing and curing test form a closed-loop mechanism. When the actual test result is not up to standard, the PLC immediately triggers the model iterative update, without manual waiting or repeated operation, and the whole process greatly improves production efficiency and consistency, realizes the full process automation from algorithm to industrial equipment, and provides an operational path for large-scale industrial applications.

[0034] Example 2

[0035] See also Figure 1 This embodiment provides a skin-feel UV ink, which is produced using the skin-feel UV ink generation method based on deep learning in any of the above embodiments. Raw material composition data and performance parameter data of the skin-feel UV ink are obtained, wherein the raw material composition data includes the weight percentages of a polyester matte resin, a polyurethane matting resin, a skin-feel resin, a photoinitiator, a slip agent, and a wax agent, and the performance parameter data include a curing speed, an adhesion level, a number of abrasion resistance times, and a feel score. A deep learning model of a multi-layer neural network is established, and the raw material composition data and the performance parameter data are respectively input into corresponding input layers of the deep learning model. The deep learning model is iteratively trained using a backpropagation algorithm so that the trained model can output an optimal raw material feed ratio. The optimal raw material feed ratio is applied to production, including mixing, stirring in a disperser, and adding additives. A UV curing test is performed after stirring is completed. The deep learning model is updated again based on the test feedback, and ultimately a skin-feel UV ink with the expected curing speed, adhesion, and feel performance is generated. This improves the efficiency and accuracy of analyzing the multi-parameter coupling relationship of the skin-feel UV ink and avoids ink batch inconsistency and performance fluctuations.

[0036] It should be pointed out that the above embodiments are only preferred specific implementation methods of the present invention, and the protection scope of the present invention is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention. The protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for generating skin-feel UV ink based on deep learning, characterized in that: include: Obtain raw material composition data and performance parameter data of skin-feel UV ink, wherein the raw material composition data includes the weight percentages of polyester matte resin, polyurethane easy-matting resin, skin-feel resin, photoinitiator, slip agent, and wax agent; and the performance parameter data includes curing speed, adhesion level, abrasion resistance times, and hand feel score; Establishing a deep learning model of a multi-layer neural network, inputting the raw material composition data and the performance parameter data into the corresponding input layers of the deep learning model respectively; iteratively training the deep learning model through a back propagation algorithm so that the trained model can output the optimal raw material feed ratio; The optimal raw material ratio is applied to production, including mixing, stirring in a disperser, and adding additives. After stirring, a UV curing test is performed. The deep learning model is updated again based on the test feedback, and ultimately a skin-feel UV ink with the expected curing speed, adhesion, and feel is generated; During the mixing and dispersing process, an online monitoring module is added to monitor the stirring temperature, viscosity, and speed in real time, and the detection results are input into the process monitoring sub-network of the deep learning model. The process monitoring sub-network compares the difference between the predicted viscosity and the actual viscosity in real time. If the deviation exceeds a threshold, the dispersing machine speed is automatically adjusted or the additive addition rate is changed. After the slip additive and wax additive are added, a secondary mixing method is used. The curing efficiency is monitored through UV curing pre-test. Finally, the measured data is fed back to the deep learning model for fine-tuning, so that the generated skin-feel UV ink remains consistent and controllable during the production process, avoiding formulation inaccuracies caused by temperature changes or uneven stirring. A feel scoring subnetwork is set up in the deep learning model to separately extract key features related to the skin-feel resin content and the type of slip additive; the feel scoring subnetwork receives feel scoring training samples and uses an attention mechanism to weight the main features affecting the feel; the output of the feel scoring subnetwork is combined with the remaining prediction results of the deep learning model main model to obtain a final score, so that the material ratio can be more accurately matched with excellent tactile performance; during verification in production, a standardized blind touch evaluation process is adopted, and a fixed number of evaluators are invited to score. The scores are then input into the feel scoring subnetwork for error backtracking correction to achieve refined control of the skin feel effect.

2. The method for generating skin-feel UV ink based on deep learning according to claim 1, characterized in that: Before obtaining the raw material composition data and performance parameter data, it also includes establishing a sample collection subsystem, which conducts experimental measurements on multiple batches of skin-feel UV inks to obtain accurate experimental measurement data. The experimental measurement data include the curing speed, adhesion level, wear resistance and feel score corresponding to different proportions of polyester matte resin, polyurethane easy-to-matt resin, skin-feel resin, photoinitiator, slip agent and wax agent; in the experimental measurement, a unified test standard is used for measurement and the temperature and humidity are stably controlled, and the measured data is transmitted to the database for summary and collation, and the collated experimental measurement data is obtained and input into the training set and validation set of the deep learning model to improve the generalization ability and prediction accuracy of the model; after the training of the deep learning model is completed, the model parameters are fixed and the first round of optimal formula recommendations are output to enter the subsequent actual production verification and retraining stage.

3. The method for generating skin-feel UV ink based on deep learning according to claim 1, characterized in that: When establishing a deep learning model of a multi-layer neural network, an input layer is set to receive raw material composition data and performance parameter data, and at least three hidden layers are set to construct a feature extraction and nonlinear mapping structure; residual connections are introduced in the hidden layers to avoid the phenomenon of gradient disappearance or explosion when the network layer is deep; a multi-objective regression unit is established in the output layer to simultaneously predict the curing speed, adhesion level, wear resistance and feel score of skin-feel UV ink; the mean square error loss function is used in the training process, and the weights are constrained in combination with L2 regularization to prevent overfitting; through training, the deep learning model can accurately predict the corresponding performance indicators according to the feed ratio, and realize intelligent optimization of the ratio of polyester matte resin, polyurethane easy-to-matt resin, skin-feel resin, photoinitiator, slip additive and wax additive.

4. The method for generating skin-feel UV ink based on deep learning according to claim 1, characterized in that: During the deep learning model training process, four sub-goals were set: curing speed, adhesion level, wear resistance, and feel score. A weighted multi-objective optimization strategy was used to assign a corresponding weight coefficient to each sub-goal. Through iterative training, multiple rounds of testing were conducted on different weight coefficient configurations to screen out a comprehensive formula that meets both fast curing requirements and ensures high adhesion and good feel. The weighted multi-objective optimization strategy generated a set of feasible solutions at the model output layer, and then selected the optimal solution based on pre-set target sorting rules. Under this optimal solution, a set of recommended ratios for polyester matte resin, polyurethane easy-to-matt resin, skin-feel resin, photoinitiator, slip additive, and wax additive were obtained in the feeding stage to take into account the needs of both fast printing and durable feel.

5. The method for generating skin-feel UV ink based on deep learning according to claim 1, characterized in that: Before outputting the optimal raw material feed ratio, a simulation verification phase is first carried out, and the generated formula parameters are input into a simulation module with physical constraints. The simulation module calculates the theoretical curing efficiency and surface tension based on the known kinetic models of polyester matte resin, polyurethane easy-matting resin, skin-feel resin, photoinitiator, slip agent and wax additive in the chemical reaction and photocuring process; if the simulation results deviate from the results predicted by the deep learning model, the deviation information is fed back to the model for secondary training until the simulation results match the predicted results within the preset error range.

6. The method for generating skin-feel UV ink based on deep learning according to claim 1, characterized in that: When adjusting the ratio of the slip agent to the wax agent, a chemical surface energy calculation model is introduced to estimate the effect of different combinations of additives on the surface friction coefficient of the cigarette package printed material; In the training samples, high friction coefficients are stored corresponding to scratch rates, and a penalty term is set for the friction coefficient prediction error during training; when the additive combination leads to potential scratch risks, the chemical surface energy calculation model increases the loss value of the additive combination scheme in the output layer, prompting the final formula to tend to reduce the friction coefficient.

7. The method for generating skin-feel UV ink based on deep learning according to claim 1, characterized in that: When performing backpropagation training on the raw material feed ratio, an automatic learning rate adjustment mechanism is introduced to dynamically modify the learning rate according to the convergence speed of the loss function during training. If the loss value decreases slowly within several iterations, the learning rate is automatically increased to accelerate convergence. If the loss value fluctuates or increases, the learning rate is reduced to avoid excessive updates.

8. A skin-feel UV ink, characterized in that: The skin-feel UV ink is produced using the skin-feel UV ink generation method based on deep learning as described in any one of claims 1 to 7.

Images