Abstract
To improve the efficiency and accuracy of thermal and moisture comfort prediction for underwear, this study developed a hybrid model combining Principal Component Analysis (PCA) and a Genetic Algorithm-optimized Back Propagation (GA-BP) neural network. PCA was utilized to reduce data dimensionality and eliminate multi-collinearity among variables, while the Genetic Algorithm optimized the initial weights and thresholds of the BP network to prevent local minima and slow convergence. The resulting PCA-GA-BP model demonstrates superior accuracy, stability, and computational efficiency compared to traditional BP, SVM, and K-means models, offering a robust tool for digital comfort simulation in garment design.
Introduction
Predicting clothing comfort is a complex challenge because thermal and moisture factors are non-linearly coupled. For instance, as moisture evaporates, it carries away heat, fundamentally changing the garment's thermal profile. Historically, researchers relied on sweat thermal manikins or simple fabric tests, which often fail to account for the "wearing state"—how a specific human body shape interacts with the fabric. Standard Back Propagation (BP) neural networks used for this task often suffer from data redundancy due to multi-collinearity and are prone to getting "stuck" in suboptimal mathematical solutions (local minima).
Methodology
The researchers proposed a three-stage pipeline called PCA-GA-BP to address these issues. First, Principal Component Analysis (PCA) was used to compress 12 fabric parameters and several body shape indexes into five principal components, retaining over 84% of the original information while eliminating redundancy. Second, a Genetic Algorithm (GA) was employed to optimize the initial weights and thresholds of the neural network, ensuring it starts in a "good neighborhood" for training. Finally, a standard BP Neural Network was trained on these optimized starting points to produce the final comfort prediction.
Experimental Data and Datasets
The study utilized a comprehensive dataset comprising fabric properties and human physiological measurements:
- Fabric Data: 50 different brands of men's underwear were analyzed. The researchers measured 12 physical properties for various fabric blends, including Bamboo/Cotton and Modal/Spandex.
- Human Data: Five healthy young men provided detailed physiological measurements, including BMI and waist-hip ratios, to account for the impact of body shape on comfort.
- Subjective Trials: Ground truth was established through trials where testers rated underwear on a 5-level scale in a controlled climate chamber. High consistency among testers (Kendall coefficient of 0.924) validated the subjective labels.
Results and Discussion
The PCA-GA-BP model demonstrated clear superiority across all performance metrics:
- Accuracy: The maximum absolute error was only 0.0764, significantly lower than the 0.4884 observed in standard BP networks.
- Efficiency: The hybrid model reached convergence in 122 epochs (57 seconds), compared to 625 epochs (172 seconds) for the standard BP network.
- Stability: The model showed high adaptability, accurately predicting comfort for fabric types not encountered during training.
Conclusion
The integration of PCA and GA significantly enhances the predictive power of neural networks for clothing comfort. While the study was limited to a specific demographic and environmental condition, the PCA-GA-BP model provides a robust framework for manufacturers to simulate comfort digitally. Future work should expand the demographic scope and test the model under extreme environmental conditions or with high-performance "smart fabrics."
Summary updated