Banca de QUALIFICAÇÃO: TITO CACO CURIMBABA SPADINI

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : TITO CACO CURIMBABA SPADINI
DATE: 16/12/2024
TIME: 14:00
LOCAL: https://conferenciaweb.rnp.br/sala/ricardo-33
TITLE:

IT DOESN’T SOUND GOOD: Sound-based Intelligent Machinery Fault Diagnosis

PAGES: 103
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUMMARY:

This study investigates intelligent fault diagnosis in rotary machinery using exclusively sound signals captured by a single microphone. The analyzed faults include vertical misalignment, horizontal misalignment, load imbalance, bearing cage defects, outer raceway faults, and ball bearing failures. Each fault type exhibits multiple intensity levels, resulting in 42 independent classes in the study. The dataset, imbalanced in terms of sample distribution per class, contains sound signals originally recorded at 50 kHz, 24-bit resolution, with a duration of 5 seconds.
During preprocessing, the signals were resampled and quantized, normalized, had silences removed, and underwent denoising using a Wiener filter. Following segmentation, time-domain and frequency-domain features were extracted, forming a feature matrix. The samples were scaled, divided into training and testing sets, and the training process was augmented with oversampling of minority classes. The model, based on ensemble learning utilizing eXtreme Gradient Boosting (XGBoost), was optimized through an initial broad grid search, followed by iterative fine-tuning refinements.
The final model, evaluated on the test set, demonstrated high performance. Using all features, it achieved 99.54% accuracy and 99.52% F$_\beta$ score. When restricted to Mel-Frequency Cepstral Coefficients (MFCCs) and their first- and second-order Deltas, the performance yielded 97.83% accuracy and 97.74% F$_\beta$ score. A forward selection wrapper, employing the 50 most relevant features among MFCCs and Deltas, achieved 97.90% accuracy and 99.32% F$_\beta$ score. The MFCC Deltas proved more significant in importance compared to the MFCCs themselves.

Finally, a detailed table of parameters contributing to the best results is presented.

COMMITTEE MEMBERS:
Presidente - Interno ao Programa - 1761107 - RICARDO SUYAMA
Membro Titular - Examinador(a) Interno ao Programa - 1544392 - ALINE DE OLIVEIRA NEVES PANAZIO
Membro Titular - Examinador(a) Interno ao Programa - 1761105 - MURILO BELLEZONI LOIOLA
Membro Suplente - Examinador(a) Interno ao Programa - 2334927 - ANDRE KAZUO TAKAHATA