ETD

Archivio digitale delle tesi discusse presso l'Università di Pisa

Tesi etd-09082020-120744


Tipo di tesi
Tesi di laurea magistrale
Autore
POCHIERO, AMEDEO
URN
etd-09082020-120744
Titolo
An interpretable multi channel feature learner based on deep autoencoder for predictive maintenance
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
COMPUTER ENGINEERING
Relatori
relatore Prof. Cimino, Mario Giovanni Cosimo Antonio
relatore Prof.ssa Vaglini, Gigliola
tutor Ing. Mazzuca, Stefano
correlatore Dott. Alfeo, Antonio Luca
Parole chiave
  • predictive maintenance
  • deep learning
  • autoencoder
Data inizio appello
25/09/2020
Consultabilità
Non consultabile
Data di rilascio
25/09/2090
Riassunto
The industry is going through a fourth period of revolution in which everything is aimed at digitizing and connecting existing machines to acquire data necessary to support business decisions. Predictive Maintenance aims to provide an accurate prediction of the state of degradation of the machine, estimating the moment of failure. It can reduce the time intervals in which the machine is idle, thus increasing the hours of actual production and also it reduces maintenance costs since it is no longer an extraordinary event, but foreseen by the system, so it is possible to prepare the maintenance at an earlier stage. In this work has been designed a Predictive Maintenance system based on Deep Learning able to predict the various degradation states of the machine. The architecture consists of a pre-processing phase in which data from different sensors are processed and transformed in different matematical domains to ensure flexibility to the type of input data. Then, an Autoencoder is trained to extract the input data features and through an evaluation method based on clustering and Adjusted Rand Index the best features are selected, thus providing also global interpretability. The latter allow to train a classifier with the task of providing the machine degradation state, i.e. Normal State, Degradation State, Critical State. Several studies have been conducted to ensure the optimal parameters to be used and to achieve an accuracy that in some cases reaches as much as 99%.
File