Analysis of bearings defaults using machine learning techniques

Abstract

Rotating machines have become ubiquitous in contemporary industries, playing a pivotal role in various applications. The consequences of defects in these machines extend beyond mere technical issues, potentially leading to substantial economic losses and posing a significant threat to human safety. Operators often grapple with the intricacies of troubleshooting these complex systems, where a single mistake can have catastrophic consequences. One of the most critical elements in these machines is the bearings. Consequently, numerous researchers have dedicated their time and efforts to addressing this matter. While extensive studies have been conducted in this field, a common limitation is the focus on constant-speed scenarios. In reality, rotating machines typically operate under non-stationary conditions, making constant-speed techniques largely theoretical. This thesis is divided into two essential parts. The first part addresses the challenges of diagnostic resolution under time-varying conditions. Given the dynamic nature of the working environment, understanding and mitigating faults in non-stationary conditions is imperative for practical applications. Our method aims to tackle the diagnostic issue under time-varying conditions. The technique was tested on a bearing database collected under time-varying conditions, containing three types of faults. Vibrational signals are initially processed using the Empirical Wavelet Transform (EWT) to extract AM-FM modes. Subsequently, a list of features is extracted from these modes. For feature selection, the Clan-Based Cultural Algorithm (CCA) is employed, and model training utilizes the Random Forest algorithm. The results demonstrate the robustness of the diagnostic process despite varying conditions. The second part focuses on feature selection, which plays a crucial role in controlling the quality of the diagnostic system and reducing misleading factors. This area of research is increasingly attracting attention, with numerous methods developed. However, many of these techniques require in-depth domain knowledge, particularly concerning parameter tuning and result interpretation. In this work, we introduce a robust technique based on standard deviation and Random Forest methods for sequential feature selection. The method was tested on three different bearing databases, including time-varying conditions, and three signal decomposition techniques (EWT, EMD, and MODWPT). It provided promising results in terms of both quality and quantity, being user-friendly and not demanding extensive knowledge in the optimization field