针对化纤卷绕机长套加工过程中多工序、多源偏差非线性耦合导致质量控制困难的问题，构建了异构加权偏差传递网络模型（heterogeneous weighted variation propagation network, HWVPN）和关键质量特征（key quality characteristics, KQCs）识别方法。首先，基于实际工艺流程解析多源偏差传递机理，构建融合质量与工况特征的HWVPN拓扑模型；其次，针对工业现场小样本数据约束，提出弹性网络（elastic net, EN）与贝叶斯回归（Bayesian regression, BayR）相结合的两阶段求解方法（EN-BayR），通过稀疏特征选择与概率后验推断，精准量化节点间的偏差传递权重；进而，设计综合偏差放大与抵消效应的Net-Weighted-NodeRank指标，实现关键质量特征的精准识别。工程实例验证表明，该方法拟合精度（R²）达98.1%，显著优于传统回归及集成学习模型，有效揭示了长套加工偏差的演化规律与关键控制环节。

 To address the challenges of quality control caused by the nonlinear coupling of multi-stage and multi-source variations in the machining of long sleeves for chemical fiber winders, this paper proposes a method for identifying key quality characteristics (KQCs) based on a heterogeneous weighted variation propagation network (HWVPN). First, based on the analysis of the multi-source variation propagation mechanism within the actual manufacturing process, a topological model of HWVPN is constructed to integrate both quality characteristics and operating conditions. Second, to handle the constraint of small sample sizes typical in industrial settings, a two-stage algorithm combining Elastic Net (EN) and Bayesian Regression (BayR), termed EN-BayR, is proposed. This algorithm accurately quantifies the variation propagation weights between nodes via sparse feature selection and probabilistic posterior inference. Furthermore, a Net-Weighted-NodeRank indicator is designed to account for both variation amplification and cancellation effects, facilitating the precise localization of KQCs. An industrial case study validates that the proposed method achieves a fitting accuracy (R²) of 98.1%, significantly outperforming traditional regression and ensemble learning models. The results effectively reveal the evolution laws of machining variations and identify critical control points for the long sleeve.