Premature Ventricular Contraction refers to irregular heartbeat and is one common symptom to several heart diseases. Currently, physiological databases are not only large in volume but also complex in dimensional aspect, so that intelligent systems that can process multi-dimensional data to detect Premature Ventricular Contraction (PVC) are highly needed. In this paper, we propose novel models of combinations of multi-leads ECG from the 12 lead ECG St. Petersburg Arrhythmias database to detect PVCs and optimize the required data pre-processing resources for Convolutional Neural Network(CNN) implemented on wearable devices. Although exhibiting fewer performances than previous works, the proposed method is able to perform automatic features extraction, reduce the CNN complexity and is scalable to be applied to 3-Lead to 16-Lead ECG systems. The combination scenarios include Wavelet fusion method and Tucker-decomposition before CNN is deployed as a classifier. The achieved accuracy to detect PVC for tensor-based feature extraction, the most optimized processing technique, is 90.84% with a sensitivity of 78.60% and a specificity of 99.86%.