Session: 03-16-01: AI Integration in Mechanical Engineering and Smart Manufacturing

Paper Number: 146131

146131 - Machine-Learning Based Curing Cycle Optimization in Wind Blade Manufacturing. 

Vacuum-assisted resin infusion mold (VARIM) process has been widely employed for wind blades manufacturing due to its cost-effectiveness, high performance, and consistent results. In the current manufacturing setup, composite molds for wind blade component manufacturing and assembly have multiple heating zones embedded with heating coils.  A mold heat control system is used to set the mold heating profile and adjust the energy supplied to the heating zones based on the temperature measured at each heating zone. Heating zones are configured based on the laminate definition per product design. However, generally for blade manufacturing, for simplicity, all the heating zones of the mold are set to follow the same temperature profile regardless of the local laminate schedule being made or the amount of thermal mass of bonding adhesive applied within each heating zone.   This results in the situation where some regions reach a sufficient degree of cure (DOC) for demolding, while other regions are under-cured.  Therefore, a safety margin must be built into the cure profiles to ensure all sections of the laminate are cured, which not only induces high processing costs and longer cycles but also regional over-cure. DOC differential can cause severe product quality issues due to uneven material shrinkage leading to cracks in bonding adhesives or thermal waves, and distortion in laminate. That generates a significant amount of waste in blade manufacturing plants. There also exist quality problems that escape from blade manufacturing plants and later reveal themselves in the field. The cost of fixing that type of problem in the field is orders of magnitude higher than doing that in the blade plant.

To tackle these challenges, we propose a multi-zone heated bed setup where temperature zones are tailored to the varying blade thickness. However, determining the optimal temperature for each zone presents a computational challenge due to the complexity of the curing process and the countless possible temperature combinations. To circumvent the challenges, we introduce a novel machine-learning approach employing a digital twin of the VARIM process based on a high-fidelity multiphysics solver to capture the key mechanisms of curing process. These simulations provide training data for a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) based machine learning model. This method obviates the need for costly and time-consuming lab experiments, as the ML model effectively learns complex heating patterns and resin curing dynamics. Once trained, the CNN-LSTM model can efficiently compute temperature zone values for an optimized curing cycle. Validation of the proposed framework is carried out through a lab-scale experimental setup involving a three-zone, tapered cross-section composite. This methodology holds the potential to significantly reduce manufacturing time while ensuring consistent and reliable mechanical properties in wind turbine blades.

 

However, the current manufacturing process encounters challenges such as long curing times and potential defects like distortion/delamination due to nonuniform heating applied across the entire heterogeneous blade structure during curing. This results in varying degrees of cure along the blade length, impacting mechanical properties and performances. To tackle these challenges, we propose a multi-zone heated bed setup where temperature zones are tailored to the varying blade thickness. The effectiveness of this process relies on determining the optimal temperature for each zone. While a Multiphysics-based solver is capable of simulating the curing process, the inverse problem of determining the optimal temperature for each zone poses a computational challenge due to the complexity of the curing process and the countless possible temperature combinations. On the other hand, machine learning approaches provide quicker results but rely on large datasets requiring numerous expensive and laborious experiments that are further required to be labelled.

 

To circumvent the challenges, we introduce a novel machine-learning approach employing a digital twin of the VARIM process based on a high-fidelity Multiphysics solver to capture the key mechanisms of curing process. These simulations provide training data for a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) based machine learning model. A CNN-LSTM model is shown to be capable of understanding changing and spatial as well as temporal domain which is required to understand the complex interactions of the resin flow and its curing at high temperatures. This method obviates the need for costly and time-consuming lab experiments, as the ML model effectively learns complex heating patterns and resin curing dynamics. Once trained, the CNN-LSTM model can efficiently compute temperature zone values for an optimized curing cycle. Validation of the proposed framework is carried out through a lab-scale experimental setup involving a three-zone, tapered cross-section composite of unidirectional glass fibers sandwiching a foam core. This methodology holds the potential to significantly reduce manufacturing time while ensuring consistent and reliable mechanical properties in wind turbine blades.

Presenting Author: Sahil Kamath University of Texas at Dallas

Presenting Author Biography: Current PhD student at UT Dallas.

Authors:

Sahil Kamath University of Texas at Dallas
Guilherme Caselato Gandia The University of Texas at Dallas
Niloufar Adab The University of Texas at Dallas
Ehsan Mehrdad The University of Texas at Dallas
Dong Qian University of Texas at Dallas
Hongbing Lu The University of Texas at Dallas