Banca de DEFESA: MAILSON RIBEIRO SANTOS

Online and Offline Approaches to Fault Detection, Classification and Estimation in Dynamical Systems

Detection, classification, and severity of faults; Offline and online approaches; Support Vector Machine; TEDA; AutoCloud.

This study addresses methods for detecting, classifying, and assessing the severity of faults in dynamic systems, responding to the need for effective monitoring in complex industrial environments. Aiming to mitigate human errors and identify faults in real time, machine learning approaches, both offline and online, are employed. In the offline learning methodology, features are selected for their relevance based on information extracted from an Explainable Artificial Intelligence (XAI) technique, with the goal of developing effective and efficient models. The Support Vector Machine (SVM) was used at all stages of this approach. The second part of the study focused on an online learning approach, employing evolutionary algorithms in all phases. Two data preprocessing approaches were tested: one based on offline feature relevance results and another using windowed sensor data. Additionally, a modification to the Typicality and Eccentricity Data Analysis (TEDA) algorithm was proposed for fault detection and classification, comparing two versions to identify the most effective. In the final online phase, the AutoCloud algorithm was employed to identify fault severity. A shared aspect between offline and online learning approaches is the sequential criterion, where data previously identified as faulty is used in fault classification, while data for each fault type is separately used in severity identification. To validate the proposals, the Case Western Reserve University (CWRU) benchmark for bearing faults was used. In the offline learning approach, satisfactory results were obtained with a reduced number of features, demonstrating the efficiency and effectiveness of the proposed model. Results from the online learning approach showed that the Modified TEDA consistently outperformed the Original TEDA in fault detection, regardless of the preprocessing approach adopted. However, classification capability was more satisfactory when the second preprocessing approach was used in conjunction with the Original TEDA. Regarding fault severity identification, the first approach yielded satisfactory results, especially for specific fault types, while the second approach encountered difficulties, resulting in lower evaluation metrics. Comparing the online and offline learning approaches, both showed similar effectiveness in fault detection and classification, but severity identification was more accurate in the offline learning approach. It is concluded that both proposals are promising, with their utilization determined based on the characteristics of the dynamic system.