Μη επιβλεπόμενη ανίχνευση ανωμαλιών με χρήση βαθιών νευρωνικών δικτύων και εφαρμογή στη ναυτιλία

Abstract

This thesis explores with outlier detection techniques and their application in the real world problem of identifying anomalous data, in relation to the mechanical system of a merchant boat. The aim is to develop a predictive maintenance system by detecting anomalous behaviour and assess the correlation to historical engine damages. Three different deep neural network architectures were developed for the aforementioned purpose: an autoencoder, an adversarial autoencoder and a recurrent neural network that were compared to their ability to detect anomalies. We start by utilizing the autoencoder to identify anomalies when the reconstruction error exceeds a empirically defined threshold. Due to issues created by the high number of anomalous data, it was rendered difficult to create a statistically defined threshold. Furthermore we implemented the recurrent neural network, which served to compare it with the results of the autoencoder. In order to utilize the latent dimension on anomaly detection we integrated an adversarial autoencoder. Finally, we investigated the ability of adversarial autoencoder to identify anomalies in a clustering set-up with a categorical imposed probability and we concluded that besides the positive results, the training process was unstable when we imposed an imbalanced categorical distribution.