Remaining useful life estimation of critical industrial equipment by using a data driven approaches

Abstract

In the face of fierce competition in the industrial sector, manufacturing technologies have evolved considerably, from simple to complex and sophisticated systems. Machining processes cover a wide range of manufacturing systems and play a crucial role in production. Monitoring the condition of these systems through an effective maintenance strategy is therefore essential for guaranteeing production reliability and quality; however, due to their complexity, focusing on monitoring critical components is generally more practical and efficient, as their degradation has a serious impact on the entire system. Maintenance practices have undergone significant evolution, giving rise to the emergence of the concept of Prognostics and Health Management (PHM), which introduces a predictive approach in Condition-Based Maintenance (CBM). Estimating components Remaining Useful Lives (RULs) is one of the most important aspects of PHM to track their degradation and predict how long they can operate before failing. However, to achieve accurate RUL estimations, it is essential to process the raw data efficiently. Sensor-derived data often contains noise, irrelevant information, and inconsistencies that can mask essential insights. Therefore, it is crucial to conduct appropriate data processing to extract the relevant features, or Health Indicators (HIs), that reflect the system's behavior. This data refinement not only improves the accuracy of predictive models but also increases their robustness, enabling manufacturers and maintenance managers to schedule replacements, minimize unplanned downtime, and extend machine life cycles. This is the background to the present study, which proposes a novel methodology for estimating the RUL. In the context of PHM data-based methods, also known as "data-driven methods," the main objective of this thesis is to design relevant health indicators HIs capable of reflecting the degradation behavior of critical components and estimating their RULs. Initially, a new time-frequency analysis technique called Empirical Wavelet Packet Decomposition (EWPD) was introduced. This method uses a new segmentation of the signal's Fourier spectrum, which is spread over several levels to improve the structural performance of conventional methods. Subsequently, a novel health indicator HI, is developed upon the basis of an innovative selection of time-domain features for each frequency band at each level. Lastly, the RULs of the monitored components are estimated using the Long Short-Term Memory (LSTM) network. The proposed methodology is implemented on a selection of CNC milling cutters from the "Prognostics Data Challenge 2010" database