Remaining Useful Life (RUL) prediction of rotating machines based on Deep Learning

Abstract

In this comprehensive research endeavor, we introduce a cutting-edge Robust Multi- branch Deep Learning-based system designed for the dual purpose of predicting Remaining Useful Life (RUL) and identifying Operating Conditions (OC) in rotating machines. Our proposed system integrates a sophisticated architecture with three main components, each playing a crucial role in enhancing the accuracy and robustness of the predictions. The initial component involves the implementation of an LSTM-Autoencoder meticulously crafted to denoise the vibration data. This step is pivotal in ensuring that the subsequent analysis is conducted on high-quality, noise-free data, laying a solid foundation for reliable predictions. Following this, the second component of our system focuses on feature extraction, where we employ advanced techniques to generate a diverse set of features. This includes extracting information from the time domain, frequency domain, and time-frequency domain, providing a comprehensive understanding of the underlying patterns within the denoised data. The culmination of our system lies in the third component, which introduces a novel and robust multi-branch deep learning network architecture. This architecture is specifically designed to capitalize on the diverse features extracted in the previous step. By leveraging multiple branches, our system can effectively capture intricate relationships within the data, leading to more accurate predictions of both Remaining Useful Life and Operating Conditions. To validate the efficacy of our proposed system, extensive evaluations were conducted on two benchmark datasets—XJTU-SY and PRONOSTIA. The performance of our system was rigorously compared against state-of-the-art methodologies, and the experimental results unequivocally demonstrate the superiority of our approach. Not only does our system outperform existing solutions, but it also showcases remarkable potential for real-life applications, particularly in the context of bearing machines. In our thesis, we presents a robust and innovative framework that addresses the challenges of RUL prediction and OC identification in rotating machines. The amalgamation of advanced

denoising techniques, feature extraction methods, and a multi-branch deep learning architecture positions our proposed system as a frontrunner in the field, promising enhanced accuracy and reliability for predictive maintenance applications in real-world scenarios.