Abordagens com aprendizagem on-line e off-line para detecção, classificação e estimação de falhas em sistemas dinâmicos

Abstract

This study addresses methods for detecting, classifying, and assessing the severity of failures in dynamic systems, in response to the need for effective monitoring in complex industrial environments. Aiming to mitigate human errors and identify failures in real-time, machine learning approaches are employed, both off-line and on-line. The first part of the work uses an off-line learning methodology, employing features selected for their relevance based on information extracted from an Explainable Artificial Intelligence (XAI) technique, with the goal of developing effective models. The Support Vector Machine (SVM) was utilized at all stages of this approach. The second part of the study focused on an on-line learning approach, using evolutionary algorithms throughout all phases. Two data preprocessing approaches were tested: one based on the relevance of features obtained off-line and the other using temporal windowing on sensor data. Additionally, a modification to the Typicality and Eccentricity Data Analysis (TEDA) algorithm was proposed for failure detection and classification, comparing two versions of the algorithm to identify the most effective one. In the final on-line phase, the AutoCloud algorithm was employed to identify the severity of the failures. A common aspect between the off-line and on-line learning approaches is the sequential criterion, where previously identified faulty data are used in failure classification, while data for each type of failure are used separately in severity assessment. For validation of the proposals, the benchmark from Case Western Reserve University (CWRU) for bearing faults was used. In the off-line learning approach, satisfactory results were obtained with a reduced number of features, demonstrating the efficiency and effectiveness of the proposed model. Results from the on-line learning approach showed that the Modified TEDA algorithm achieved superior evaluation metrics compared to the Original TEDA in detecting failures, regardless of the preprocessing approach used. However, classification capability was more satisfactory when using the windowed data preprocessing approach in conjunction with the Original TEDA. Regarding severity assessment, the first approach yielded satisfactory results, especially for failures of a specific type, while the second approach faced difficulties, resulting in lower evaluation metrics. Comparing the on-line and off-line learning approaches, both showed similar effectiveness in failure detection and classification, but severity identification was more precise with the off-line learning approach. It is concluded that both proposals are promising.