Automatic detection of premature ventricular contractions (PVCs) is essential for early identification of cardiovascular abnormalities and reduction of clinical workload. As the most prevalent arrhythmia, PVCs can cause cardiac failure or sudden death. The difficulty resides in extracting features that effectively reflect the electrocardiogram (ECG) signals. Transition networks (TNs), which represent the transition relationships between various phases of a time series, are advantageous for capturing temporal dynamics. Therefore, in order to recognize PVCs, each heartbeat was first split into serval segments; then, their statistical properties were calculated for the sequence construction; and finally, network topology-related features were extracted from TN constructed by these sequences of statistical properties and input into decision tree-based Gentleboost for PVC recognition. The algorithm was trained on the MIT-BIH arrhythmia (MIT-BIH-AR) database and tested on the St. Petersburg Institute of Cardiological Technics (INCART) 12-lead arrhythmia database, wearable ECG (WECG) database, and noise stress test database by four evaluation metrics: sensitivity, positive predictivity, $F1$ -score ( $F1$ ), and area under the curve (AUC). The proposed algorithm achieved an average $F1$ of 0.9784 and AUC of 0.9975 on MIT-BIH-AR and proved good generalization ability on INCART and WECG with $F1$ = 0.9633 and 0.9467 and AUC = 0.9887 and 0.9755, respectively. The algorithm also exhibited robustness and noise immunity as evidenced by tests on sensitivity of R-wave peak offset and noise, and real-world daily life conditions. Overall, the proposed PVC detection algorithm based on TN theory offered high classification accuracy, strong robustness, and good generalization ability, with great potential for wearable mobile applications.