Abstract:

Objective This study aims to tackle the critical challenge of seam slippage in worsted wool fabrics used for high-end suiting by establishing a reliable predictive model integrating structural parameters and deep learning. The persistent issue of seam integrity undermines both product quality and durability, necessitating an accurate, efficient, and intelligent solution to replace conventional trial-and-error approaches. The objective of this research is to develop a CNN-based framework capable of predicting seam slippage resistance in both warp and weft directions to support fabric development and quality control with higher precision and lower cost.

Method A total of 932 sets of experimental data were collected from worsted wool fabric samples, covering key parameters such as fabric structure, warp and weft densities, raw material, yarn fineness, fabric tightness, yarn twist, and finishing processes. Ultimately, eight core features most strongly correlated with seam slippage resistance were selected as input vectors. A CNN-MLP dual-branch neural network model was constructed, wherein the CNN branch processes image features of the fabric structure, and the MLP branch handles other process parameters. The two types of features were subsequently fused for prediction, generating predicted values for the warp and weft seam slippage resistance indicators, respectively.

Results The mean absolute percentage error (MAPE) was adopted as the core evaluation metric. The model demonstrated outstanding predictive performance on an independent validation set. The MAPE for predicting warp seam slippage resistance was as low as 7.8%, while that for weft seam slippage resistance reached 6.8%. The prediction errors for both indicators remained at a low level, indicating strong practical value for guiding production.The MAPE curves for the training and validation sets converged and remained closely aligned, with no signs of overfitting, confirming the reliability of the predictions and strong generalization capability of the model. Feature importance analysis revealed that fabric total tightness and fabric structure were the most significant factors influencing seam slippage resistance, together accounting for over 50% of the impact. This aligns well with the mechanical principles of woven fabrics. As the fabric tightness increased, both warp and weft seam slippage resistance were improved significantly. When the tightness increased without a substantial change in the weft-to-warp ratio, both warp and weft densities increased. This resulted in more interlacing points per unit length between warp and weft yarns. Consequently, the seam slippage resistance increases, leading to improved seam slippage resistance in both warp and weft directions.The influence of fabric structure is primarily reflected in the number of interlacing points between warp and weft yarns. A higher number of interlacing points would increase the resistance during seam slippage, causing the fabric less prone to slipping. Additionally, the fulling process also considerably affects seam slippage resistance. Under identical other parameters, fulled fabrics exhibited greater seam slippage resistance compared to non-fulled fabrics by virtue of the felting effect between warp and weft yarns induced by fulling, which enhances inter-yarn friction and reduces seam slippage.

Conclusion This study proposes a novel method for predicting seam slippage resistance in worsted wool fabrics. The CNN branch captures structural features of the fabric weave, while the MLP branch learns the influences of process parameters. The dual prediction heads accurately reflect anisotropy, effectively addressing the challenge of performance prediction under multi-parameter coupling. This enables precise prediction of seam slippage resistance behavior in worsted wool fabrics, providing a reliable computational tool for fabric structure design and process optimization. The model demonstrates not only excellent prediction accuracy and robustness, but also the readiness for integrating into existing textile production management systems or CAD software. It offers real-time, data-driven decision support for process optimization and quality control, indicating high industrial applicability and broad promotion prospects.Future work will focus on collecting larger-scale datasets to further enhance model stability and exploring its transfer learning capabilities across different fabric types. Additionally, efforts will be made to improve the model's adaptability under dynamic working conditions and investigate its integration into full-process smart manufacturing systems.

Key words: CNN-MLP dual-branch model, neural network, worsted wool, seam slippage resistance, performance prediction