This study aims to address the limitations inherent in conventional signal processing methodologies and Rectified Linear Unit (ReLU) activation functions for extracting discriminative features from non-stationary vibration signals in the context of incipient fault diagnosis in rolling element bearings. It proposes a Convolutional Neural Network (CNN) framework integrated with a Segmented Exponential Linear (SEL) activation function. The SEL mechanism performs nonlinear transformations across positive and negative signal intervals, thereby preserving critical information embedded in weak fault signatures while maintaining computational efficiency. The comprehensive experimental validation conducted using the Case Western Reserve University (CWRU) bearing dataset demonstrates that within limited training epochs, the model achieved 100% classification accuracy for both the source and target domains in the best test accuracy scenario, with no overfitting observed. These results validate the effectiveness of the proposed method in domain adaptation tasks.