Abstract: Glass fiber reinforced polymer (GFRP) composites widely used in offshore wind turbine blades are highly susceptible to low-velocity impacts, which often induce internal damage and accelerate fatigue residual strength degradation. To efficiently predict the fatigue residual strength of GFRP composites with initial impact damage, a series of drop-weight impact, tension-tension fatigue, and quasi-static tensile tests are systematically conducted, and a machine learning-based prediction method is proposed. Specimens with six layup sequences and three fiber volume fractions are designed, and a fatigue residual strength dataset is established based on the experimental results. Degradation models are derived for two types of impact-damaged laminates with stacking sequences of 45°/−45°_2s and 0°/90°_2s, showing that both stacking sequence and fiber content significantly affect residual strength evolution. Furthermore, four machine learning-based prediction models are constructed, namely Support Vector Machine (SVM), Random Forest (RF), Back-Propagation Neural Network (BPNN), and Bayesian Neural Networks (BNNs). To mitigate overfitting under small-sample conditions and to meet the need for uncertainty quantification and probabilistic modeling in engineering applications, a BNNs-based prediction method is particularly developed for predicting post-impact fatigue residual strength. To enable the machine learning models to accurately recognize parameters that are discrete and mutually independent, feature engineering is carried out by employing one-hot encoding to effectively represent the layup sequence, and applying Z-score normalization to continuous input features such as impact energy. To accurately evaluate the fatigue residual strength of the glass fiber reinforced composite laminates with initial impact damage, impact energy, layup sequence, glass fiber content, and fatigue stress level are selected as input features for the machine learning models to predict the fatigue residual strength after impact. The analysis results indicate that the prediction errors of residual strength for all machine learning models are generally controlled within 20%, among which the BNNs-based model performs best in terms of prediction accuracy and stability, with an average coefficient of determination R² = 0.9565.