Banca de DEFESA: HIAGO HENRIQUE GOMES DE ARAUJO

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : HIAGO HENRIQUE GOMES DE ARAUJO
DATE: 16/08/2022
TIME: 14:00
LOCAL: Online
TITLE:
KEY WORDS:

Anomaly Detection, Prognostic and Health Management, Condition Based Maintenance, Vibration Analysis, Neural Network.

PAGES: 81
BIG AREA: Engenharias
AREA: Engenharia de Produção
SUMMARY:

Advances in automation technologies, and advances in production machinery make maintenance costs increasingly relevant in relation to total costs in manufacturing systems. In this context, the technological upgrade and cost reduction of sensors structure, enables the development of information systems capable of collecting, storing, and monitoring evidence about the health of equipment. In a Condition Based Maintenance (CBM) context, the processing and analysis of the monitoring data can guide maintenance policies. In this paper, we address the anomaly detection activity, present in CBM, in which we identify whether the equipment is about to fail or not, without going into the classification of causes and failure modes that lead to anomalous behavior. Anomaly detection can be facilitated by means of machine learning algorithms. These models can be categorized into supervised and unsupervised.In the first case, data (e.g., vibration, temperature, pressure) is required to be associated with binary or categorical information about the health state of the equipment. However, in maintenance and industrial equipment monitoring system contexts, its usual that the data is not available with the values of these labels, or is not fully labeled, or there is insufficient data from the equipment in a faulty state. In these cases, unsupervised or semi-supervised algorithms can be used as alternatives, since they do not require a large amount of pre-labeled data. Besides the problem related to the amount of data labeled as faulty, sensor malfunction is a recurring problem in real-life condition-based maintenance contexts. The need then arises for the construction of robust anomaly detection models capable of detecting faults even in the presence of problems in the data coming from monitoring, such as missing data and the presence of noise. In this work, an approach involving adversarial networks and variational autoencoders is proposed. Four distinct categories of neural network structures have been used: Autoencoder, Variational Autoencoder, Adversarial Autoencoder, and Adversarial Generative Networks. The proposed models are tested, validated and compared from three different datasets with vibration signals commonly used as benchmarking in the literature, the results show that the Variational Autoencoder models present the best results for the models in which there is a lower presence of missing data and handle better the cases of noise inserted in the database. For the cases with higher uncertainty regarding missing data, the Generative Adversarial Network and Variational Autoencoder present superior performance when compared to the other deep learning models tested

BANKING MEMBERS:
Externo à Instituição - ERICK GIOVANI SPERANDIO NASCIMENTO
Presidente - 2732514 - ISIS DIDIER LINS
Interno - 2766188 - MARCIO JOSE DAS CHAGAS MOURA