Rotating machinery is widely employed in various industries, including petroleum, automotive, food processing, etc. To facilitate the smooth rotational movement of various subcomponents, bearings are commonly employed in such machinery. However, due to factors such as fluctuating speeds, excessive loads, and prolonged periods of operation, bearings are susceptible to wear and degradation. To prevent bearing failures, enhance equipment reliability, and reduce maintenance costs, real-time monitoring and diagnostic techniques for bearings are essential. Predictive Maintenance (PdM) is a widely adopted strategy for ensuring consistent operational conditions of machinery by monitoring their health status through sensor data. However, the high diversity and massive volume of sensor data present significant challenges in the analysis of fault signals. To address this challenge, we propose an Exponential Power Entropy (EPE) based feature extraction method for extracting salient features from sensor data and feeding them to a Neural Network (NN) for further NN-based fault diagnosis. Additionally, we propose a Decision Threshold (DT) approach to enhance the prediction accuracy of the NN. These approaches not only ensure the quality of the fault diagnosis but also significantly reduce the computation time. In comparison to the traditional feature extraction method, the proposed EPE-based feature extraction method demonstrates an accuracy improvement from 2.8% to 28.7% and reduces the number of neurons by 54.5% to 94.7% in bearing fault diagnosis while reducing model complexity by 98.3% compared to traditional Deep Neural Network (DNN) approaches.