Banca de DEFESA: JOSÉ PAULO GONÇALVES DE OLIVEIRA
Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.STUDENT : JOSÉ PAULO GONÇALVES DE OLIVEIRA
DATE: 18/11/2022
TIME: 09:00
LOCAL: Sessão por videoconferência - Link: https://meet.google.com/toe-izuf-rzn
TITLE:
MACHINE LEARNING IN INDUSTRY 4.0: ANOMALY DETECTION IN EMBEDDED SYSTEMS AND CLASSIFICATION OF SUBSTANCES.
KEY WORDS:
Automated testing, industry 4.0; anomaly detection; classification of substances; machine learning; autoencoder.
PAGES: 158
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUMMARY:
Quality control is a critical aspect, especially in Industry 4.0 context. In addition to being extremely important to meet the functional prerequisites of a given product, quality is closely related to safety, security, and economic issues. In this work, we address two specific aspects of quality certification in the modern sense of the industry: detection of anomalies in embedded systems and classification of chemical or biological substances. The solutions addressed are based on machine learning models.
In the electronics industry, component miniaturization and the use of multi-layer boards have considerably increased the complexity of testing. Consequently, traditional forms of testing based on manual inspections have become outdated and inefficient. In addition, Industry 4.0 demands products with a high level of personalization. This imposes additional requirements, such as a high degree of flexibility in the conception, design and testing processes. Therefore, effective and flexible solutions that do not require physical contact with the tested product are needed. Our study features automated and non-invasive testing solutions.
Regarding substance testing, spectroscopy techniques are traditionally employed using spectrometers. Despite of being a very mature technique, its limitation is the cost and complexity of the equipment. We propose a simple, however efficient alternative to perform tests without the use of spectrometers.
For validation, we designed and built prototypes to carry out experiments. For anomaly detection, we built a system using development board. Six software versions have been implemented – one of which is functional, while the rest represents some kind of anomaly. Anomalies are detected by analyzing temporal signals captured from the circuit in a non-invasive way. The signals are converted into spectrographic images that are analyzed by a machine learning model, trained to detect whether the circuit behaves as specified.
For classification of substances, we present a Proof of Concept. Instead of employing a spectrometer, we use an optical transmission and reception system. The transmitted signal has a specially designed waveform for maximum performance. The detected signal is converted into a spectrographic image that is used by a machine learning model that performs the classification.
Experimental results demonstrate the effectiveness of the proposed test methods. For various electronic system test scenarios, the detection rate reaches 100%. Substance classification also performs optimally (100% accuracy) in various configurations. Additionally, we present a technique to increase performance by transforming the data used for training and validating the model. The effectiveness of the technique is experimentally proven both for detecting anomalies and for classifying substances.
COMMITTEE MEMBERS:
Externo ao Programa - 1650867 - ABEL GUILHERMINO DA SILVA FILHO - UFPEPresidente - ***.367.184-** - CARMELO JOSÉ ALBANEZ BASTOS FILHO - UFPE
Externo à Instituição - FRANCISCO MADEIRO BERNARDINO JUNIOR - UNICAP
Externo à Instituição - GUILHERME DE ALENCAR BARRETO - UFC
Externo à Instituição - JOSE ALFREDO FERREIRA COSTA - UNICAMP