An automatic proportioning and stable control system for gilding glue

Abstract

The application discloses a gilding glue automatic proportioning and stable control system and relates to a gilding production auxiliary system in the textile field. The gilding glue automatic proportioning and stable control system comprises a glue state sensing module, an evaporation trend prediction module, an intelligent proportioning decision module, a self-adaptive adjustment control module and an online closed-loop correction module. The gilding glue automatic proportioning and stable control system constructs a full-closed loop system of sensing, prediction, decision, control and correction, realizes long-term automatic stabilization of the glue concentration in a target interval, effectively improves gilding fastness and brightness, and reduces defects such as gold dropping and glue overflow.

Description

Technical Field

[0001] This invention relates to an auxiliary system for textile hot stamping production, and more specifically, to an automatic and stable control system for hot stamping adhesive mixing. Background Technology

[0002] In the textile hot stamping production process, adhesive is the core consumable for achieving a firm bond between the hot stamping layer and the textile fabric. Its concentration stability directly determines the fullness, brightness, and fastness of the hot stamping product. Currently, textile hot stamping factories generally use manual preparation of adhesive and natural storage. In actual production, due to factors such as the production heat generated by the hot stamping equipment and fluctuations in the ambient air temperature in the workshop, the diluent in the adhesive will continuously evaporate and decrease, leading to a gradual increase and unstable fluctuation in the adhesive concentration.

[0003] Excessive glue concentration can lead to uneven bonding between the hot stamping layer and the fabric, resulting in glue overflow and affecting the clarity and brightness of the hot stamping pattern. It also increases curing time, reducing production efficiency. Conversely, insufficient glue concentration can cause poor adhesion, leading to issues like foil flaking and peeling, severely impacting product durability and increasing scrap rates. Furthermore, manual monitoring of glue concentration is prone to errors and has a slow response time. Manually adding diluents to adjust the glue concentration cannot achieve real-time adjustment, further exacerbating the instability of the glue concentration and failing to meet the demands of large-scale, high-quality hot stamping production.

[0004] Currently, some automated glue mixing equipment only has a single mixing function and lacks real-time concentration monitoring and dynamic adjustment capabilities. It cannot solve the problem of concentration fluctuations caused by the evaporation of diluents and is difficult to adapt to the special needs of textile hot stamping production. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide an automatic and stable control system for hot stamping adhesive.

[0006] The above-mentioned technical objective of this invention is achieved through the following technical solution: an automatic mixing and stabilization control system for hot stamping adhesive, comprising: The adhesive state sensing module is used to acquire real-time concentration data, adhesive surface exposure area, and multi-dimensional evaporation influencing factors of hot stamping adhesive online. The real-time concentration data is obtained by converting at least one parameter among viscosity, density, or refractive index. The multi-dimensional evaporation influencing factors include at least the ambient temperature and humidity around the adhesive, the local heat radiation intensity generated by the hot stamping equipment, and the air flow speed above the adhesive surface. The evaporation trend prediction module is connected to the glue state sensing module. It internally deploys a time-series convolutional attention network prediction model. The time-series convolutional attention network prediction model takes the time-series data output by the glue state sensing module as input, extracts the nonlinear dynamic features of glue concentration change, and integrates external evaporation influencing factors to output the prediction curve of the evaporation consumption of diluents within a future time window and the corresponding glue concentration change trajectory. The intelligent proportioning decision module is connected to the evaporation trend prediction module. It has a built-in deep reinforcement learning agent. The deep reinforcement learning agent takes the glue concentration change trajectory, the current glue balance and the target process concentration range as the state, the staged replenishment flow curve of the diluent as the action, and the comprehensive benefit of the actual concentration being kept within the target range and the cost of the diluent consumption as the reward to generate a real-time diluent compensation proportioning strategy. An adaptive mixing control module, connected to the intelligent mixing decision module, is used to generate a mixing control law based on a compensation mixing strategy, and adjust the dynamic mixing ratio of diluent and adhesive base material to offset the concentration drift caused by evaporation. The online closed-loop correction module is connected to the glue state perception module, the evaporation trend prediction module, and the intelligent proportioning decision module, respectively. It acquires the glue concentration detection value after mixing in real time, calculates the residual between the concentration detection value and the target concentration, and uses the residual to perform incremental transfer learning correction on the weights of the temporal convolutional attention network prediction model. At the same time, it adjusts the exploration strategy of the deep reinforcement learning agent to maintain the long-term stability of the glue concentration in a global closed-loop manner.

[0007] The present invention is further configured such that: when the glue state sensing module is used to acquire real-time concentration data, it first collects viscosity, density and refractive index signals, and then performs online fusion regression on the signals through a soft measurement model based on weighted least squares support vector machine to output real-time concentration data.

[0008] The present invention is further configured such that: the multidimensional evaporation influencing factor also includes the hot stamping equipment operating cycle signal, the glue circulation flow rate and the glue container opening geometry factor, and the real-time concentration data is obtained by multi-sensor fusion calculation of viscosity, density and refractive index.

[0009] The present invention is further configured such that: the temporal convolutional attention network prediction model consists of a dilated causal convolutional layer, a multi-head self-attention mechanism and a gated residual connection; the multidimensional evaporation influence factor, after being mapped by an independent embedding layer, is concatenated with the temporal features output by the dilated causal convolutional layer in the channel dimension and input together into the multi-head self-attention mechanism.

[0010] The present invention is further configured such that the deep reinforcement learning agent adopts the maximum entropy deep reinforcement learning algorithm.

[0011] The present invention is further configured such that: the phased replenishment flow curve of the dilution aid output by the deep reinforcement learning agent is parameterized as a time-series pulse sequence, wherein the amplitude and start time of each pulse in the time-series pulse sequence are determined by the agent's actions.

[0012] The present invention is further configured such that the adaptive allocation control module includes a feedforward calculation unit and a feedback fine-tuning unit.

[0013] The present invention is further configured such that: the feedforward calculation unit generates a baseline replenishment flow rate based on the evaporation consumption prediction curve of the diluent; the feedback fine-tuning unit collects the concentration detection value of the mixed glue and calculates its deviation from the target concentration; and generates a flow correction amount using a fuzzy adaptive PID algorithm.

[0014] The present invention is further configured such that: when the online closed-loop correction module performs incremental transfer learning correction on the weights of the temporal convolutional attention network prediction model, it uses the importance of each network parameter to the Fisher information matrix of the historical prediction task based on the elastic weight consolidation mechanism to differentially constrain the parameter update step size.

[0015] The present invention is further configured such that: the intelligent proportioning decision module is connected to an adhesive lifecycle management module, the adhesive lifecycle management module receives the current adhesive balance and the predicted curve of the evaporation consumption of diluent, calculates the remaining time of adhesive availability, and outputs a signal that adhesive needs to be replaced or replenished when the remaining time is lower than the safety threshold and the current compensation proportioning strategy cannot maintain the concentration within the target process concentration range.

[0016] In summary, the present invention has the following beneficial effects: 1. Construct a closed-loop system for sensing, prediction, decision-making, control, and correction to achieve long-term automatic stability of glue concentration within the target range, effectively improving hot stamping fastness and brightness, and reducing defects such as gold peeling and glue overflow. 2. Online integration of viscosity, density, refractive index, and acquisition of multi-dimensional evaporation factors such as temperature, thermal radiation, and airflow overcomes the problems of large errors and slow response of manual monitoring, providing high-precision real-time data for precise control; 3. Utilize temporal convolutional attention networks to predict future evaporation consumption and concentration change trends of dilution additives, providing reliable forward-looking information for early intervention. Employ maximum entropy deep reinforcement learning to dynamically optimize additive replenishment strategies and achieve compensation through parameterized temporal pulses, thereby saving additive costs while ensuring quality. 4. The combination of feedforward compensation and fuzzy adaptive feedback can quickly offset the concentration drift caused by evaporation, resulting in strong dispensing accuracy and anti-interference ability; 5. Utilize the residual concentration after adjustment to perform incremental transfer learning on the prediction model and dynamically adjust the decision-making strategy to enable the system to adapt to changes during long-term operation and prevent concentration runaway. 6. Adhesive lifecycle management: Calculate the available time based on evaporation prediction and provide timely warnings for replacement to avoid production interruptions caused by adhesive aging or depletion. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the application scenario of the automatic mixing and stabilization control system for hot stamping adhesive of the present invention. Figure 2 This is a block diagram of the overall structure of the automatic mixing and stable control system for hot stamping adhesive of the present invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 , Figure 2 This invention provides an automatic proportioning and stabilization control system for hot stamping adhesive, deployed in the adhesive circulation supply unit of a textile hot stamping production line. The system comprises an adhesive status sensing module, an evaporation trend prediction module, an intelligent proportioning decision module, an adaptive dispensing control module, and an online closed-loop correction module, forming a complete closed-loop control loop. The modules exchange data via industrial Ethernet and synchronize time to ensure consistency of timing data.

[0020] The adhesive status sensing module is responsible for acquiring the raw data needed for online decision-making and prediction. This module first deploys a multi-sensor fusion array at key nodes in the adhesive storage tank and circulation pipeline. This array includes an insertion-type vibration viscometer, a Coriolis mass flow density meter, and an online refractometer, all connected in parallel bypass to the main adhesive circulation pipeline to ensure continuous adhesive flow through the measuring tank. Simultaneously, multiple environmental sensors are deployed above the adhesive surface in the storage tank, including a temperature and humidity sensor to collect ambient temperature and humidity, an infrared thermometer to capture the localized thermal radiation intensity generated by the hot stamping equipment, and a differential pressure anemometer to measure the airflow velocity above the adhesive surface. The adhesive status sensing module also receives operating cycle signals from the hot stamping equipment PLC and flow rate feedback signals from the adhesive circulation pump, and calculates the exposed area of ​​the adhesive surface in real time using a laser rangefinder sensor installed on the top of the storage tank. The static parameter of the adhesive container opening geometry factor is manually entered by the operator through the human-machine interface during system initialization.

[0021] The viscosity, density, and refractive index signals collected from sensors are not directly equivalent to glue concentration and require software fusion and conversion. In this embodiment, the glue state sensing module incorporates a soft measurement model based on a weighted least squares support vector machine. Before the system is put into use, standard glue samples with different concentration gradients are prepared in the laboratory, their true concentrations are calibrated using standard instruments, and the viscosity, density, and refractive index signals output by the sensor array are recorded simultaneously to construct a training dataset. This dataset is used to train the weighted least squares support vector machine model offline to determine the model parameters. During online operation, the model takes the three real-time collected physical signals as input and outputs high-precision glue concentration data in real time through nonlinear mapping and weighted regression. All sensor data and the fused concentration data, along with multi-dimensional evaporation influencing factors such as ambient temperature, ambient humidity, local thermal radiation intensity, air flow velocity, equipment operating cycle, glue circulation flow rate, and liquid surface exposure area, are timestamped to form a time-series data stream, which is sent to the evaporation trend prediction module.

[0022] The core of the evaporation trend prediction module is an internally deployed temporal convolutional attention network prediction model. This model consists of three stacked parts: a dilated causal convolutional layer, a multi-head self-attention mechanism, and gated residual connections. Upon receiving continuous temporal data from the glue state perception module, the model first normalizes the raw data. Then, multidimensional evaporation influencing factors, including ambient temperature, ambient humidity, local thermal radiation intensity, airflow velocity, operating cycle, circulation velocity, and opening geometry, are input into an independent embedding layer. This embedding layer, composed of a fully connected network, maps heterogeneous influencing factors to a high-dimensional feature space, generating an evaporation condition feature vector. Simultaneously, the historical concentration sequence enters the dilated causal convolutional layer. Through progressively expanding receptive fields, the dilated causal convolutional layer efficiently extracts long-range nonlinear dynamic features of glue concentration changes without introducing future information leakage. Subsequently, the concentration temporal features output by the dilated causal convolutional layer and the evaporation condition feature vector output by the embedding layer are concatenated along the channel dimension to form a comprehensive feature map rich in internal state change patterns and external evaporation driving factors. The synthesized feature map is then fed into a multi-head self-attention mechanism layer. This mechanism computes the correlation weights between any two time steps in the feature sequence in parallel across multiple representation subspaces, accurately capturing which historical evaporation conditions and concentration states have a decisive impact on future evaporation trends. Finally, the output passes through a gated residual connection consisting of gated linear units and residual connections, accelerating model convergence and preventing gradient vanishing. The model ultimately outputs a predicted curve of the diluent's evaporation consumption and the corresponding glue concentration change trajectory within a future time window, such as the next 30 minutes.

[0023] After receiving the predicted curve and trajectory, the intelligent mixing decision module treats them as part of the environmental state. Together with the current glue balance obtained from the glue state perception module and the externally set target process concentration range, they form a complete state vector, which is then input to the internally deployed deep reinforcement learning agent. In this embodiment, the deep reinforcement learning agent is implemented based on the maximum entropy deep reinforcement learning algorithm. The agent's actions are parameterized as a time-series pulse sequence, which precisely describes the phased replenishment flow curve of the diluent over a future period. The amplitude of each pulse represents the magnitude of the replenishment flow, and the start time of the pulse represents the start time of the replenishment action. The core objective of the agent is to maximize a composite reward function, which consists of a weighted sum of two parts: a positive reward proportional to the duration for which the measured glue concentration remains within the target process concentration range, and a negative cost proportional to the total amount of diluent consumed during that period. By continuously exploring and pursuing the maximum entropy objective during the learning process, the agent can find the optimal balance between "precisely maintaining concentration" and "saving on diluent usage," ultimately outputting a carefully calculated diluent compensation mixing strategy.

[0024] The adaptive dispensing control module consists of a feedforward calculation unit and a feedback fine-tuning unit working in parallel. The feedforward calculation unit directly reads the predicted consumption curve of the diluent evaporation from the evaporation trend prediction module and converts it into a reference replenishment flow signal as the feedforward control quantity. This feedforward control quantity can quickly and proactively compensate for most of the concentration changes caused by evaporation. Simultaneously, the feedback fine-tuning unit collects data in real time from a concentration detector installed downstream of the dispensing pipeline and calculates the slight deviation between the detected value and the target concentration. This deviation signal is input to a fuzzy adaptive PID controller, which tunes the proportional, integral, and derivative coefficients online according to a fuzzy rule base and generates a flow correction quantity. The reference replenishment flow rate and the flow correction quantity are summed in a signal superimposed unit to form the final control command, driving a high-precision metering pump to precisely inject the diluent concentrate into the glue mixer, achieving dynamic mixing with the circulating glue base material, thereby offsetting the concentration drift caused by evaporation.

[0025] To ensure the system's stability and adaptability during long-term operation, an online closed-loop calibration module is activated. This module periodically acquires the concentration of the mixed adhesive and calculates the residual between it and the target setpoint. This residual signal is used for calibration at two levels. At the first level, it is used for online incremental transfer learning correction of the temporal convolutional attention network prediction model. The module employs an elastic weight consolidation mechanism, calculating the Fisher information matrix of each model parameter for historical prediction tasks to assess the importance of each parameter. When backpropagating to update network weights using the new residual, the update step size of important parameters is strongly constrained, while less important parameters can be flexibly adjusted. This effectively prevents catastrophic forgetting while absorbing new operating conditions, ensuring the model's prediction accuracy remains stable despite environmental shifts such as equipment aging and seasonal changes. At the second level, the residual signal triggers adjustments to the exploration strategy of the deep reinforcement learning agent. When the concentration residual is consistently large and exceeds the expected range, it indicates that the current strategy may have fallen into a local optimum or that the operating conditions have changed significantly. At this time, the system temporarily increases the noise variance of the agent's action output to encourage it to explore a wider range in order to find a better compensation strategy until the residual stabilizes again within an acceptable range.

[0026] The system also integrates an adhesive lifecycle management module. This module continuously receives the current adhesive balance and the diluent evaporation consumption prediction curve output by the evaporation trend prediction module. Through integral calculation, the module extrapolates the maximum remaining time before the adhesive concentration exceeds the required process range. When the calculated remaining time is lower than a preset safety threshold, and all compensation ratio strategies output by the deep reinforcement learning agent fail to maintain the predicted concentration trajectory within the target process concentration range, it means that the adhesive itself has aged or the diluent has been exhausted. The module will trigger a tiered warning signal, prompting the operator through the human-machine interface and an audible and visual alarm to replace or replenish the adhesive in a timely manner, thereby avoiding production interruptions and batch quality accidents caused by adhesive failure.

[0027] The soft measurement model, based on weighted least squares support vector machines, introduces sparsity techniques to reduce online computation. For evaporation influencing factors, the model also uses a feature selection algorithm to identify the three key factors with the greatest impact on the evaporation rate of the current glue formulation; the dimensionality-reduced features are used for prediction to balance prediction efficiency. The dilated causal convolution in the temporal convolutional attention network increases its dilation rate layer by layer in powers of two, thus achieving a large receptive field with fewer layers. During the training phase, the deep reinforcement learning agent also incorporates an imitation learning warm-up process, first cloning the policy network's behavior using matching data from human experts, and then incorporating self-learning, significantly reducing the risk and time of random exploration in the initial online phase. Online closed-loop correction adjusts the agent's exploration strategy by dynamically adjusting the action variance parameter output by its policy network. When the residual is below a threshold, the variance gradually shrinks for refined utilization; when the residual exceeds the threshold, the variance expands to promote exploration.

[0028] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An automatic mixing and stabilization control system for hot stamping adhesive, characterized in that: include: The adhesive state sensing module is used to acquire real-time concentration data, adhesive surface exposure area, and multi-dimensional evaporation influencing factors of hot stamping adhesive online. The real-time concentration data is obtained by converting at least one parameter among viscosity, density, or refractive index. The multi-dimensional evaporation influencing factors include at least the ambient temperature and humidity around the adhesive, the local heat radiation intensity generated by the hot stamping equipment, and the air flow speed above the adhesive surface. The evaporation trend prediction module is connected to the glue state sensing module. It internally deploys a time-series convolutional attention network prediction model. The time-series convolutional attention network prediction model takes the time-series data output by the glue state sensing module as input, extracts the nonlinear dynamic features of glue concentration change, and integrates external evaporation influencing factors to output the prediction curve of the evaporation consumption of diluents within a future time window and the corresponding glue concentration change trajectory. The intelligent proportioning decision module is connected to the evaporation trend prediction module. It has a built-in deep reinforcement learning agent. The deep reinforcement learning agent takes the glue concentration change trajectory, the current glue balance and the target process concentration range as the state, the staged replenishment flow curve of the diluent as the action, and the comprehensive benefit of the actual concentration being kept within the target range and the cost of the diluent consumption as the reward to generate a real-time diluent compensation proportioning strategy. An adaptive mixing control module, connected to the intelligent mixing decision module, is used to generate a mixing control law based on a compensation mixing strategy, and adjust the dynamic mixing ratio of diluent and adhesive base material to offset the concentration drift caused by evaporation. The online closed-loop correction module is connected to the glue state perception module, the evaporation trend prediction module, and the intelligent proportioning decision module, respectively. It acquires the glue concentration detection value after mixing in real time, calculates the residual between the concentration detection value and the target concentration, and uses the residual to perform incremental transfer learning correction on the weights of the temporal convolutional attention network prediction model. At the same time, it adjusts the exploration strategy of the deep reinforcement learning agent to maintain the long-term stability of the glue concentration in a global closed-loop manner.

2. The automatic proportioning and stabilization control system for hot stamping adhesive according to claim 1, characterized in that: When the glue state sensing module is used to acquire real-time concentration data, it first collects viscosity, density and refractive index signals, and then performs online fusion regression on the signals through a soft measurement model based on weighted least squares support vector machine to output real-time concentration data.

3. The automatic proportioning and stabilization control system for hot stamping adhesive according to claim 1, characterized in that: The multidimensional evaporation influencing factors also include the hot stamping equipment operating cycle signal, glue circulation flow rate, and glue container opening geometry factor. The real-time concentration data is obtained by multi-sensor fusion calculation of viscosity, density, and refractive index.

4. The automatic proportioning and stabilization control system for hot stamping adhesive according to claim 1, characterized in that: The temporal convolutional attention network prediction model consists of dilated causal convolutional layers, multi-head self-attention mechanisms, and gated residual connections. The multidimensional evaporation influence factor, after being mapped by an independent embedding layer, is concatenated with the temporal features output by the dilated causal convolutional layer in the channel dimension and input into the multi-head self-attention mechanism.

5. The automatic proportioning and stabilization control system for hot stamping adhesive according to claim 1, characterized in that: The deep reinforcement learning agent employs the maximum entropy deep reinforcement learning algorithm.

6. The automatic proportioning and stabilization control system for hot stamping adhesive according to claim 5, characterized in that: The phased replenishment flow curve of the dilution aid output by the deep reinforcement learning agent is parameterized as a time-series pulse sequence, and the amplitude and start time of each pulse in the time-series pulse sequence are determined by the agent's actions.

7. The automatic proportioning and stabilization control system for hot stamping adhesive according to claim 1, characterized in that: The adaptive allocation control module includes a feedforward calculation unit and a feedback fine-tuning unit.

8. The automatic proportioning and stabilization control system for hot stamping adhesive according to claim 7, characterized in that: The feedforward calculation unit generates a baseline replenishment flow rate based on the predicted curve of the evaporation consumption of the diluent. The feedback fine-tuning unit collects the concentration detection value of the mixed glue and calculates its deviation from the target concentration, and generates the flow correction amount using a fuzzy adaptive PID algorithm.

9. The automatic proportioning and stabilization control system for hot stamping adhesive according to claim 1, characterized in that: When the online closed-loop correction module performs incremental transfer learning correction on the weights of the temporal convolutional attention network prediction model, it uses the elastic weight consolidation mechanism and the importance of each network parameter to the Fisher information matrix of the historical prediction task to differentiate the parameter update step size.

10. The automatic proportioning and stabilization control system for hot stamping adhesive according to claim 1, characterized in that: The intelligent mixing decision module is connected to the glue lifecycle management module. The glue lifecycle management module receives the current glue balance and the evaporation consumption prediction curve of the diluent, calculates the remaining glue time, and outputs a signal that the glue needs to be replaced or replenished when the remaining time is lower than the safety threshold and the current compensation mixing strategy cannot maintain the concentration within the target process concentration range.

Images