05 February 2024
09:00 Doctoral defense IC3 Auditorium
Theme
Machine Learning Algorithms to Improve Data Quality and Inference Performance in Monitoring Applications
Student
Juliane Regina de Oliveira
Advisor / Teacher
Lucas Francisco Wanner - Co-supervisor: Eduardo Rodrigues de Lima
Brief summary
Electric power transmission lines make up a complex energy supply system around the world. Transmission lines are made up of guyed towers. The guyed towers contain a central "V" shaped mast and four anchored cables responsible for supporting the structure. Loosening of cables and displacement of the structure occur due to various environmental factors and even theft of part of the structure. Changes in the tower's parameters can lead to the structure collapsing and the transmission system collapsing. Early detection of faults allows the maintenance of the structure and can avoid catastrophic situations in electrical power transmission lines. Current sensing technologies monitor structural parameters such as stay cable tension and vibration signals and machine learning algorithms for fault detection and estimation of tower parameters. The sensors and algorithms make up the real-time structural monitoring system and the early detection of faults such as cable relaxation and structure displacement make transmission lines more robust and fault tolerant. The pre-processing of vibration signals into features in the time and frequency domain helps models detect structural failures and increases the dimensionality of the data set. In addition to dealing with structural monitoring, sensors are devices that can present component failures, be influenced by environmental changes and malicious injection of faults. These conditions result in noisy sensor readings and other faults such as bias, drift, outlier, stuck-at, and missed readings. The consequences of sensor failures range from incorrect inference of the environment to even waste of resources. Techniques to improve the quality of sensor readings are data fusion and machine learning algorithms. The scientific contributions of the thesis include improving the quality of sensing data in monitoring applications. In addition, fusion of monitoring data from guyed towers and improving the quality of inference using dimensionality reduction and data explainability strategies. The experiments conducted show the relevance of strategies composed of data fusion, feature selection techniques and machine learning algorithms for the good performance of inferences for monitoring applications. The results show that feature selection using explainability metrics was robust to noisy features and resulted in performance close to noiseless features.
Examination Board
Headlines:
| Lucas Francisco Wanner | IC / UNICAMP |
| Claudio Miceli de Farias | COPPE/UFRJ |
| Roberto Milton Scheffel | UTFPR |
| Juliana Freitag Borin | IC / UNICAMP |
| Gustavo Fraidenraich | FEEC / UNICAMP |
Substitutes:
| Allan Mariano de Souza | IC / UNICAMP |
| Giovani Gracioli | UFSC |
| Lisane Brisolara de Brisolara | UFPEL |