High-precision glue coating prepreg production control method and production line

Abstract

This application provides a high-precision prepreg production control method and production line. The method includes: S1: collecting prepreg production process parameters and performing layered processing to obtain layered process parameters; S2: inputting the layered process parameters into an improved LSTM prediction model to obtain the predicted resin content and its prediction confidence at a future set time; S3: dynamically adjusting the weights of the feedforward control based on the prediction confidence; S4: acquiring the measured resin content online at the future set time, using the measured resin content to incrementally learn and update the improved LSTM prediction model, and performing feedback calibration on the control commands. This method improves control accuracy and real-time performance by using a layered LSTM model to predict resin content in advance and evaluate the confidence level, thereby dynamically adjusting the weights of feedforward and feedback control. Simultaneously, the model can self-update, adapting to changes in operating conditions, and improving the stability and consistency of prepreg quality.

Description

Technical Field

[0001] This application relates to the field of coated prepreg production, and in particular to a high-precision coated prepreg production control method and production line. Background Technology

[0002] In the field of high-performance composite material manufacturing, prepreg, as an intermediate substrate, directly determines the performance of the final composite parts. Among these factors, the precision of resin content control is a core indicator in prepreg production, directly affecting the areal density, fiber volume fraction, and the uniformity and reliability of the final mechanical properties of the composite material. Therefore, achieving high-precision and stable control of resin content is a key technological step in improving prepreg quality and meeting the stringent requirements of high-end fields such as aerospace.

[0003] Currently, the industry primarily relies on closed-loop feedback control methods based on detection (such as near-infrared spectroscopy) to control resin content. This method adjusts coating process parameters by measuring and comparing them with setpoints in real time, ensuring control stability to some extent. However, this type of method is inherently a lag control, with a time delay between detecting a deviation and executing an adjustment, leading to excessive resin content in the prepreg produced during this period. Furthermore, various process parameters in the production process are interconnected and dynamically changing, making it difficult for traditional control models to accurately describe their complex nonlinear relationships. Traditional models also exhibit poor adaptability to fluctuations in operating conditions, and their control effectiveness tends to deteriorate when dealing with raw material batch changes or equipment state drift. In addition, although some research has attempted to introduce predictive models for feedforward control to overcome lag, existing models often lack sufficient refinement in handling process variables across multiple time scales, lack a mechanism for evaluating their own predictive reliability, and are unable to perform safe and efficient self-updating in continuous production.

[0004] Therefore, there is an urgent need in this field for an advanced control method to solve the above-mentioned technical problems. Summary of the Invention

[0005] This application provides a high-precision prepreg production control method and production line. This method uses a hierarchical LSTM model to predict resin content in advance and assess confidence levels, dynamically adjusting the weights of feedforward and feedback control accordingly, thus achieving a shift from delayed correction to proactive prevention. This method significantly improves control accuracy and real-time performance. Simultaneously, the model can self-update during production, adapting to changes in operating conditions, and significantly improving the stability and consistency of prepreg quality.

[0006] Firstly, a method for controlling the production of high-precision coated prepregs is provided, the method comprising: S1: Real-time acquisition of production process parameters for coated prepreg, and layering of the process parameters to obtain layered process parameters. The layered process parameters include at least: fast dynamic layer parameters, slow process layer parameters, and batch characteristic layer parameters. S2: Input the layered process parameters into the improved long short-term memory (LSTM) prediction model, which is used to obtain the predicted value of resin content at a future set time and its prediction confidence. S3: Based on the predicted confidence level, dynamically adjust the weight of the feedforward control: when the confidence level is high, feedforward control is the main method, and an advance adjustment command is generated based on the predicted resin content; when the confidence level is low, the feedforward weight is reduced and the feedback control is enhanced. S4: At the set future time, acquire the measured value of resin content online, use the measured value of resin content to incrementally learn and update the improved long short-term memory network (LSTM) prediction model, and perform feedback calibration on the control command.

[0007] In conjunction with the first aspect, in some implementations of the first aspect, step S1 includes: stratifying the time scale of the influence of the process parameters on the resin content, wherein the fast dynamic layer parameters are parameters with a sampling frequency on the order of seconds, including the distance between the coating rollers and the production line speed; the slow process layer parameters are parameters with a sampling frequency on the order of minutes, including the resin temperature and the oven temperature; and the batch characteristic layer parameters include the fiber batch number and the resin type, which remain unchanged within a single production batch.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the improved Long Short-Term Memory (LSTM) prediction model includes a multi-branch input encoding layer; The multi-branch input encoding layer is configured to encode the fast dynamic layer parameters, slow process layer parameters, and batch feature layer parameters through independent embedding or fully connected networks, and then fuse the encoded feature vectors to form a unified time step feature, which is then input into the subsequent LSTM sequence layer for processing.

[0009] It should be understood that by finely stratifying process parameters at the second, minute, and batch levels, and by using a multi-branch input encoding layer to independently extract and fuse features of various parameters, the LSTM prediction model can more accurately characterize the differentiated impact of process variables at different speeds on resin content. This significantly improves the model's ability to represent complex production processes and its prediction accuracy, thus providing a more reliable foundation for high-precision front-end control.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the improved Long Short-Term Memory (LSTM) prediction model includes: during the model inference phase, enabling Monte Carlo Dropout technology to perform multiple forward propagations, using the mean of the multiple prediction results as the predicted value, and using the standard deviation or distribution interval of the predicted value as the prediction confidence level.

[0011] It should be understood that this application requires the model to not only provide predicted values ​​but also output their prediction confidence levels, which serve as the basis for intelligent decision-making in dynamically adjusting feedforward / feedback control weights. Monte Carlo Dropout technology is introduced, and during the model inference phase, Dropout is randomly activated multiple times for forward propagation. This transforms the deterministic single-point prediction into a probability distribution, converting randomness into a quantitative assessment of prediction uncertainty (such as standard deviation), thereby generating the confidence signal upon which control decisions depend.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, the specific steps of incremental learning and updating in step S4 include: S401: Combine the newly acquired process parameter sequence with the measured resin content to form a new training sample; S402: Freeze the weights of the lower network layers of the improved Long Short-Term Memory (LSTM) prediction model, and only unlock and update the weights of the top fully connected layers of the improved Long Short-Term Memory (LSTM) prediction model, and use the new samples for mini-batch training. S403: Once the loss function of the improved Long Short-Term Memory (LSTM) prediction model converges to a stable state, save the updated model parameters to complete online adaptation.

[0013] It should be understood that by freezing the weights of the LSTM model at the bottom level and only fine-tuning the parameters at the top level, the incremental learning method enables the model to quickly adapt to changes in operating conditions and equipment drift using real-time production data. At the same time, it effectively avoids forgetting the basic process rules that have been mastered due to learning new data, ensuring the stability and continuity of the predictive ability in long-term operation, and providing the core driving force for achieving high-precision adaptive control of the production line.

[0014] In conjunction with the first aspect, in some implementations of the first aspect, step S3 includes: Set a confidence threshold; when the prediction confidence is higher than the threshold, the weight coefficient of the feedforward control command is α; when the prediction confidence is lower than the threshold, the weight coefficient of the feedforward control command is reduced to β, and an abnormal operating condition warning is issued.

[0015] In conjunction with the first aspect, in some implementations of the first aspect, the dynamic adjustment of the feedforward control weights in step S3 further includes: Based on the different intervals in which the predicted confidence level is located, different fuzzy control rules are adopted to dynamically calculate the hybrid weight coefficients of feedforward control and feedback control, wherein the input variables of the fuzzy control rules are the predicted confidence level and the trend of the confidence level.

[0016] Secondly, a high-precision adhesive-coated prepreg production line is provided, the production line comprising: A control module that executes the method as described in any implementation of the first aspect and generates control commands; An execution module that adjusts the process parameters based on the control instructions; An online detection module is used to acquire the measured value of the resin content online and feed the measured value back to the control module. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a high-precision adhesive-coated prepreg production control method provided in an embodiment of this application. Detailed Implementation

[0018] The terminology used in the following embodiments is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” and “this” are intended to also include expressions such as “one or more,” unless the context clearly indicates otherwise. It should also be understood that in the following embodiments of this application, “at least one” and “one or more” refer to one, two, or more than two. The term “and / or” is used to describe the relationship between related objects, indicating that three relationships can exist; for example, A and / or B can indicate: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character “ / ” generally indicates that the preceding and following related objects are in an “or” relationship.

[0019] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0020] In the manufacturing of high-performance composite materials, the quality of prepregs has a decisive impact on the performance of the final product. The core control indicator is resin content, which directly relates to the uniformity and reliability of the material. Currently, the industry mainly uses closed-loop feedback control based on near-infrared spectroscopy detection. This method has inherent lag, and the period between detection deviation and execution adjustment can lead to product defects. Furthermore, the production process parameters are complexly coupled and dynamically changing, making it difficult for traditional models to accurately describe their nonlinear relationships. They also lack adaptability when dealing with raw material changes or equipment drift. Existing predictive models often fail to handle multi-timescale variables precisely and lack mechanisms for predictive reliability assessment and online safety updates, thus limiting further improvements in control accuracy.

[0021] This application provides a high-precision adhesive prepreg production control method and production line, which can effectively overcome the above-mentioned problems.

[0022] The technical solutions provided in the embodiments of this application will be described below with reference to the accompanying drawings.

[0023] Figure 1 This is a flowchart illustrating a high-precision adhesive-coated prepreg production control method provided in an embodiment of this application.

[0024] refer to Figure 1 In some examples, the method includes: S1: Real-time acquisition of production process parameters for coated prepreg, and layering of the process parameters to obtain layered process parameters. The layered process parameters include at least: fast dynamic layer parameters, slow process layer parameters, and batch characteristic layer parameters. S2: Input the layered process parameters into the improved long short-term memory (LSTM) prediction model, which is used to obtain the predicted value of resin content at a future set time and its prediction confidence. S3: Based on the predicted confidence level, dynamically adjust the weight of the feedforward control: when the confidence level is high, feedforward control is the main method, and an advance adjustment command is generated based on the predicted resin content; when the confidence level is low, the feedforward weight is reduced and the feedback control is enhanced. S4: At the set future time, acquire the measured value of resin content online, use the measured value of resin content to incrementally learn and update the improved long short-term memory network (LSTM) prediction model, and perform feedback calibration on the control command.

[0025] In some examples, step S1 includes: stratifying the time scale of the effect of the process parameters on the resin content, wherein the fast dynamic layer parameters are parameters with a sampling frequency on the order of seconds, including the distance between the coating rollers and the production line speed; the slow process layer parameters are parameters with a sampling frequency on the order of minutes, including the resin temperature and the oven temperature; and the batch characteristic layer parameters include the fiber batch number and the resin type, which remain unchanged within a single production batch.

[0026] In some examples, the improved Long Short-Term Memory (LSTM) prediction model includes a multi-branch input encoding layer; The multi-branch input encoding layer is configured to encode the fast dynamic layer parameters, slow process layer parameters, and batch feature layer parameters through independent embedding or fully connected networks, and then fuse the encoded feature vectors to form a unified time step feature, which is then input into the subsequent LSTM sequence layer for processing.

[0027] In one possible implementation, this embodiment constructs an input encoding layer with three independent fully connected branches at the model level, performs feature transformation and normalization on the three types of parameters respectively, and then concatenates the output feature vectors to form a fused temporal feature tensor, which is then input into the subsequent LSTM layer for sequence modeling and prediction.

[0028] In some examples, the improved Long Short-Term Memory (LSTM) prediction model includes: during the model inference phase, enabling Monte Carlo Dropout to perform multiple forward propagations, using the mean of the multiple prediction results as the predicted value, and using the standard deviation or distribution interval of the predicted value as the prediction confidence level.

[0029] In some examples, the specific steps of incremental learning update in step S4 include: S401: Combine the newly acquired process parameter sequence with the measured resin content to form a new training sample; S402: Freeze the weights of the lower network layers of the improved Long Short-Term Memory (LSTM) prediction model, and only unlock and update the weights of the top fully connected layers of the improved Long Short-Term Memory (LSTM) prediction model, and use the new samples for mini-batch training. S403: Once the loss function of the improved Long Short-Term Memory (LSTM) prediction model converges to a stable state, save the updated model parameters to complete online adaptation.

[0030] In some examples, step S3 includes: Set a confidence threshold; when the prediction confidence is higher than the threshold, the weight coefficient of the feedforward control command is α; when the prediction confidence is lower than the threshold, the weight coefficient of the feedforward control command is reduced to β, and an abnormal operating condition warning is issued.

[0031] In one possible implementation, the confidence threshold is set to 0.85. When the confidence of the model output prediction is higher than this threshold, the weight coefficient α of the feedforward control command is 0.9. When the confidence is lower than this threshold, the feedforward weight coefficient β is reduced to 0.2, and an abnormality warning is issued, thereby switching between active adjustment and robust control based on the confidence level of the prediction.

[0032] In some examples, the dynamic adjustment of the feedforward control weights in step S3 further includes: Based on the different intervals in which the predicted confidence level is located, different fuzzy control rules are adopted to dynamically calculate the hybrid weight coefficients of feedforward control and feedback control, wherein the input variables of the fuzzy control rules are the predicted confidence level and the trend of the confidence level.

[0033] In one possible implementation, three confidence intervals—"high," "medium," and "low"—and their corresponding basic weights are set, while three trends—"rising," "stable," and "falling"—are defined. The confidence value at the current moment and its trend compared to the previous moment are used as inputs to the fuzzy controller. Reasoning is performed through a preset rule base (e.g., "if the confidence is 'high' and the trend is 'stable,' then the feedforward weight is 'high'"). Finally, a specific hybrid weight coefficient between 0 and 1 is calculated through defuzzification to achieve smooth and adaptive adjustment of the control strategy.

[0034] This application embodiment also provides a high-precision adhesive-coated prepreg production line, the production line comprising: A control module that executes the method described in any of the foregoing examples and generates control commands; An execution module that adjusts the process parameters based on the control instructions; An online detection module is used to acquire the measured value of the resin content online and feed the measured value back to the control module.

[0035] 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. Any equivalent modifications or variations made by those skilled in the art based on the content disclosed in the present invention should be included within the scope of protection set forth in the claims.

Claims

1. A method for controlling the production of high-precision coated prepregs, characterized in that, The method includes: S1: Real-time acquisition of production process parameters for coated prepreg, and layering of the process parameters to obtain layered process parameters. The layered process parameters include at least: fast dynamic layer parameters, slow process layer parameters, and batch characteristic layer parameters. S2: Input the layered process parameters into the improved long short-term memory (LSTM) prediction model, which is used to obtain the predicted value of resin content at a future set time and its prediction confidence. S3: Based on the predicted confidence level, dynamically adjust the weight of the feedforward control: when the confidence level is high, feedforward control is the main method, and an advance adjustment command is generated based on the predicted resin content; when the confidence level is low, the feedforward weight is reduced and the feedback control is enhanced. S4: At the set future time, acquire the measured value of resin content online, use the measured value of resin content to incrementally learn and update the improved long short-term memory network (LSTM) prediction model, and perform feedback calibration on the control command.

2. The method according to claim 1, characterized in that, Step S1 includes: stratifying the time scale of the effect of the process parameters on the resin content, wherein the fast dynamic layer parameters are parameters with a sampling frequency in the second range, including the distance between the coating rollers and the production line speed; the slow process layer parameters are parameters with a sampling frequency in the minute range, including the resin temperature and the oven temperature; and the batch characteristic layer parameters include the fiber batch number and the resin type, which remain unchanged within a single production batch.

3. The method according to claim 2, characterized in that, The improved Long Short-Term Memory (LSTM) prediction model includes a multi-branch input encoding layer; The multi-branch input encoding layer is configured to encode the fast dynamic layer parameters, slow process layer parameters, and batch feature layer parameters through independent embedding or fully connected networks, and then fuse the encoded feature vectors to form a unified time step feature, which is then input into the subsequent LSTM sequence layer for processing.

4. The method according to claim 1, characterized in that, The improved Long Short-Term Memory (LSTM) prediction model includes: during the model inference phase, enabling Monte Carlo Dropout technology to perform multiple forward propagations, using the mean of the multiple prediction results as the predicted value, and using the standard deviation or distribution interval of the predicted value as the prediction confidence level.

5. The method according to claim 4, characterized in that, The specific steps of incremental learning and updating in step S4 include: S401: Combine the newly acquired process parameter sequence with the measured resin content to form a new training sample; S402: Freeze the weights of the lower network layers of the improved Long Short-Term Memory (LSTM) prediction model, and only unlock and update the weights of the top fully connected layers of the improved Long Short-Term Memory (LSTM) prediction model, and use the new samples for mini-batch training. S403: Once the loss function of the improved Long Short-Term Memory (LSTM) prediction model converges to a stable state, save the updated model parameters to complete online adaptation.

6. The method according to claim 1, characterized in that, Step S3 includes: Set a confidence threshold; when the prediction confidence is higher than the threshold, the weight coefficient of the feedforward control command is α; when the prediction confidence is lower than the threshold, the weight coefficient of the feedforward control command is reduced to β, and an abnormal operating condition warning is issued.

7. The method according to claim 1 or 6, characterized in that, The step S3 of dynamically adjusting the weights of the feedforward control also includes: Based on the different intervals in which the predicted confidence level is located, different fuzzy control rules are adopted to dynamically calculate the hybrid weight coefficients of feedforward control and feedback control, wherein the input variables of the fuzzy control rules are the predicted confidence level and the trend of the confidence level.

8. A high-precision adhesive-coated prepreg production line, characterized in that, The production line includes: A control module that executes the method as described in any one of claims 1 to 7 and generates control commands; An execution module that adjusts the process parameters based on the control instructions; An online detection module is used to acquire the measured value of the resin content online and feed the measured value back to the control module.

Images