To address the insufficient accuracy of traditional damage accumulation models and the limited generalization capability of purely data-driven models in high–low cycle combined fatigue life prediction under small-sample conditions, a high–low cycle combined fatigue life prediction method based on physics residual correction is proposed. First, the Miner linear damage accumulation model is used as the physical baseline, and its residual characteristics in high–low cycle combined fatigue life prediction are analyzed. Second, based on the low-cycle damage ratio, high-cycle damage ratio, and life-scale difference between high-cycle and low-cycle fatigue, a high–low cycle coupled damage parameter ΠLH is constructed. Combined with material category information, physics-informed damage features are established. Then, taking the logarithmic residual of the Miner model as the learning target, a physics-informed residual correction model is developed using support vector regression to predict the combined fatigue life. The proposed model is validated by leave-one-out cross-validation using 43 high–low cycle combined fatigue test samples from seven materials. The results show that the proposed model achieves an RMSElog10 of 0.0895, which is lower than those of the Miner model, Manson-Halford model, SVR model, and XGBoost model. Compared with the better-performing SVR model, the RMSElog10 is reduced by approximately 46.2%. Meanwhile, the prediction error factor of the proposed model ranges from 0.6775 to 1.4950, and all samples fall within the factor-of-1.5 error band. The ablation experiments indicate that the residual correction mechanism, the high–low cycle coupled damage parameter, and material information all contribute to the prediction results on the present dataset. The results indicate that the proposed method achieves favorable empirical prediction performance on the present dataset of 43 samples from seven materials under the LOOCV setting, while its applicability to unseen materials and more complex test conditions still requires further validation.