Introduction
As the global energy landscape shifts toward hydrogen, the safe and efficient storage of high-pressure gas has become a critical engineering priority. Type IV composite pressure vessels (CPVs)—comprising a plastic liner reinforced with carbon fiber—are the standard for Fuel Cell Electric Vehicles. However, the complexities of the filament winding process introduce manufacturing "noise" that traditional Finite Element Analysis (FEA) struggles to model. Consequently, engineers often rely on a conservative safety margin of 2.25 times the design pressure, which increases the weight and cost of storage systems. This paper presents a machine learning framework designed to predict burst strength directly from manufacturing data, potentially enabling more optimized vessel designs.
Methodology
The study utilized a dataset of 24 carbon fiber reinforced polymer (CFRP) vessels manufactured under 11 different configurations. Parameters such as tow tension, winding speed, and curing pressure were recorded and compressed into average values for model training.
To ensure robust performance despite the limited sample size ($n=24$), the researchers employed Leave-One-Out (LOO) cross-validation. In this approach, the model is trained on $n-1$ samples and tested on the remaining single sample, with the process repeated for every data point in the set.
The team evaluated several architectures, including standard linear regression, a multi-layer Neural Network (NN), Kernel Principal Component Analysis (KPCA) combined with Lasso regression, and Gaussian Process Regression (GPR). An Ensemble model was also created by averaging the predictions of the KPCA-Lasso and GPR models.
Key Insights: The Challenge of Small Data
A primary finding was the relative failure of the Neural Network. Despite its sophisticated architecture (visualized below), the NN performed poorly compared to simpler kernel-based methods. This highlights a common pitfall in engineering ML: complex models often overfit or fail to generalize when training data is scarce.
Conversely, GPR and the Ensemble model proved highly effective. GPR’s inherent ability to handle non-linearities and provide uncertainty estimates made it the strongest individual performer. Additionally, analysis revealed a non-linear "Goldilocks" relationship for fiber tension: while sufficient tension is required to avoid slack, excessive tension pre-strains the fibers, leading to premature failure.
Results
The Ensemble model emerged as the most accurate, achieving a Mean Absolute Percentage Error (MAPE) of 1.46%, a significant improvement over the baseline dummy model's 2.42%. Statistical significance was confirmed via the Kolmogorov-Smirnov test, with the GPR and Ensemble models yielding p-values below 3%, indicating their predictions were significantly better than random chance.
Conclusion
This study demonstrates that machine learning can successfully identify hidden correlations within manufacturing logs to predict the structural limits of composite pressure vessels. While the current model relies on averaged signals from a small dataset, it provides a foundation for real-time monitoring and more efficient hydrogen storage solutions. Future research will focus on utilizing full time-series data to identify localized manufacturing defects.
Summary updated