Sound Machine Learning for Failure Event Prediction
Gouila, Bastien (2019)
2019
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2019121226189
https://urn.fi/URN:NBN:fi:amk-2019121226189
Tiivistelmä
The objective of this thesis was to develop a system, which can interpret live audio data monitoring industrial processes to predict an upcoming failure event of the operative parts. The customer was particularly interested in the monitoring of pyrometallurgical processing furnaces where an explosion of one of them can inflict severe damages to personnel and equipment.
The development and testing part of the study involved audio features extraction, machine learning models design and evaluation. The practical test focused on comparing the behav-ior of each implemented model while facing real data and observing how well or poorly they predicted failure events.
By combining deep neural networks with speech recognition and time series prediction methods, the produced algorithm could identify and forecast failures several minutes ahead, giving the operators the time to react and adjust accordingly. Moreover, the results did not only match the expectations, but also revealed several potential continuations such as being adapted to different types of sensors and methods.
The development and testing part of the study involved audio features extraction, machine learning models design and evaluation. The practical test focused on comparing the behav-ior of each implemented model while facing real data and observing how well or poorly they predicted failure events.
By combining deep neural networks with speech recognition and time series prediction methods, the produced algorithm could identify and forecast failures several minutes ahead, giving the operators the time to react and adjust accordingly. Moreover, the results did not only match the expectations, but also revealed several potential continuations such as being adapted to different types of sensors and methods.