Banca de DEFESA: RODRIGO DE PAULA MONTEIRO

Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.
STUDENT : RODRIGO DE PAULA MONTEIRO
DATE: 17/02/2022
TIME: 08:00
LOCAL: Sessão por Videoconferência - Google Meet
TITLE:

TRAINING A FEATURE EXTRACTOR BASED ON DEEP LEARNING FOR DETECTING ANOMALIES.

KEY WORDS:

Anomaly Detection; Deep Learning; Convolutional Neural Networks; Machine Learning.

PAGES: 142
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUMMARY:

The anomaly detection consists of identifying patterns that differ from an expected behavior. It is an important field of study, whose applicability extends to several domains, such as the security of communication networks, the detection of diseases and frauds, among others. The anomaly detection is an essential step in decision-making processes, such as planning the factory maintenance or starting the treatment of serious illnesses. Anomalous behaviors can be caused by errors in the process or by events not known by the detection system. Anomaly detection presents some challenges that makes it different from a traditional classification problem. One of them concerns the unbalance of the data available to train the anomaly detection model. The lack, or even the availability in small quantities of data belonging to the anomalous classes are quite common situations in real-world problems. It makes difficult to define a region in space that contains all possible normal behaviors without comprising the anomalies. Several techniques have been developed over the years to address this problem. However, a group of techniques has gained special attention from the academic community. Such techniques are based on the use of deep learning. They consist of processing information across multiple layers, making it possible to obtain more meaningful representations of the input information for a given classification or regression problem. Despite the advances achieved in this field, using deep learning in anomaly detection tasks still presents some difficulties, especially in obtaining characteristics capable of representing the normal class in a satisfactory way, while simultaneously distinguishing it from the anomalous class. This paper presents the development stages of an anomaly detection system based on the joint operation of deep and traditional machine learning techniques. The first approaches that we analyzed consisted of supervised trained algorithms, assuming that all anomalous classes were known, to promote a better understanding of the problem. Then, we proceeded to the approaches in which the training of the algorithm was performed only by using data belonging to the normal class. Among the techniques proposed in the thesis, the one that presented the most promising results regarded the training of the feature extractor together with a prototype selection process. The technique presented relatively high and stable mean AUC values, e.g., above 0.95, for a niche of anomaly detection problems. The trained models were evaluated with databases composed of sound and vibration signal spectrograms, collected by sensors placed on electromechanical devices.

BANKING MEMBERS:
Externo à Instituição - ANTHONY JOSÉ DA CUNHA CARNEIRO LINS - UNICAP
Externo à Instituição - BRUNO JOSE TORRES FERNANDES - UPE
Externo à Instituição - BYRON LEITE DANTAS BEZERRA - UPE
Presidente - 026.367.184-43 - CARMELO JOSÉ ALBANEZ BASTOS FILHO - UFPE
Externo à Instituição - MANASSES DANIEL DA SILVA - UPE