Defesa de Dissertação de Mestrado – Rafael Vendramini Savi – 22/4/2019
17/05/2019 16:09
| Defesa de Dissertação de Mestrado | |
| Aluno | Rafael Vendramini Savi |
| Orientador | Prof. Marcelo Ricardo Stemmer, Dr. – DAS/UFSC |
| Data | 22/4/2019 (segunda-feira) – 10h00
Sala PPGEAS I (piso superior) |
| Banca | Prof. Marcelo Ricardo Stemmer, Dr. – DAS/UFSC (presidente);
Prof. Rodrigo Bastos Fernandes, Dr. – EMC/UFSC; Prof. Ubirajara Franco Moreno, Dr. – DAS/UFSC; Prof. Jomi Fred Hübner, Dr. – DAS/UFSC. |
| Título | Use of Dirty and Incomplete Claim Data on the Inference of Product Reliability with Satistical and Neural Network Models |
| Abstract: This work is concerned about presenting and making use of techniques to treat very deficient datasets of field claims in order to estimate reliability models of the most important product families of Bosch Rexroth’s electrical portfolio. This a relevant issue to corporates that wish to take decisions regarding product development and market strategies based on the behavior of their products in the field. The entire process from cleansing and classifying the data until the choice of appropriate reliability model is guided by a detailed study of the state-of-the-art techniques. A constraint present in this work is that the extremely large amount of data stored in the past twenty years should be processed automatically with the least possible human iteration. This motivated the research for alternative resolutions involving machine learning methods that are not usually adopted in the field of reliability assessment based on claim data. State-of-the-art techniques for natural text processing found in the branch of sentiment analysis were adapted to the classification of claim events based on text remarks of the repairs realized. For this purpose, a soft classifier based on a logistic regression model was trained and obtained an accuracy of 86.2% when analyzing features extracted from the text fields. When comparing to the classification of the claims disregarding the text analysis, it suggested corrections to 10.9% of the labels. Since the prepared databases are incomplete due to failures that are not reported to the company, adjustment methods that are suitable for data containing suspensions are employed to the median rank estimators before generating empirical reliability functions. These mappings are used to evaluate both parametric and nonparametric models. The first approach fits the two-dimensional data to a Weibull distribution, which is considered one of the most important statistical distributions for modeling life data. A fine tuning of hyperparameters of a feed-forward artificial neural network is demonstrated before being used to fit the same data. The performance of both modeling approaches are compared utilizing all the data available in this work. The results show that, while the regression to the Weibull distribution is realized several thousand times faster than training the artificial neural networks, the latter achieved up to 212 times smaller prediction error. | |
