Similarity retrieval is one of the significant problems in Computer Science with a wide range of applications in industry. It aims to find similar instances to a query object from a large database. With the advances in deep learning, the capability and performance of retrieval systems have improved by a large margin. In this thesis, we aim to address the problem of one-dimensional signal retrieval over OTDR events by adopting deep learn- ing methods. The OTDR events represent failures along optical fiber links and in our case, there are nine different types of events that were acquired by Cisco Photonics in Italy. We attempt to adopt deep learning vision techniques and architectures in a way to be suitable for one-dimensional signals. Our contribution is to design a retrieval system where we can query similar events from a large collection of signals. We propose three deep learning architectures based on Convolutional Autoencoder, Convolutional Neural Network for classification and Triplet Semi-hard Network for different training scenarios namely supervised and unsupervised. There are two principal challenges that we need to take into consideration for designing our methods: (i) our solutions have to generalize on unknown categories of events, that are not presented in the training set (ii) and methods are required to handle varying size signals in inference time. Our solutions demonstrate considerable performance and promising results in the retrieval of OTDR events.
La ricerca di similarità mira a trovare i segnali che sono più simili a un oggetto in in- gresso in un database di grandi dimensioni. È un problema significativo in Informatica con un’ampia gamma di applicazioni industriali. Sebbene sia comunemente usato per recuper- are immagini, lo scopo di questo lavoro di tesi è affrontare il problema del recupero degli eventi OTDR, che rappresentano guasti sulle fibre ottiche. Cisco Photonics (Vimercate, Italia) è attualmente in grado di distinguere tra nove diversi tipi di eventi e ci ha fornito il set di dati OTDR utilizzato in questo studio. Il nostro contributo è progettare un sistema di recupero che ci permetta di cercare eventi simili in un grande set di dati. Proponiamo tre soluzioni basate su diverse architetture tipiche dell’Apprendimento Profondo: (i) Au- toencoder, (ii) CNN per la classificazione e (iii) Triplet semi-hard networks, per affrontare gli scenari di addestramento sia supervisionato che non supervisionato. Ci sono due sfide principali che dobbiamo affrontare durante la progettazione delle nostre soluzioni: (i) i nostri metodi devono essere in grado di generalizzare su categorie sconosciute di eventi che non sono incluse nel set di allenamento e (ii) sono necessari metodi per gestire segnali di dimensione variabile che possono presentarsi in fase di inferenza. Le nostre soluzioni dimostrano prestazioni considerevoli e risultati promettenti nel recupero di eventi OTDR.
Rare OTDR events retrieval with deep learning approaches
MOLAEI, AMIRSALAR
2021/2022
Abstract
Similarity retrieval is one of the significant problems in Computer Science with a wide range of applications in industry. It aims to find similar instances to a query object from a large database. With the advances in deep learning, the capability and performance of retrieval systems have improved by a large margin. In this thesis, we aim to address the problem of one-dimensional signal retrieval over OTDR events by adopting deep learn- ing methods. The OTDR events represent failures along optical fiber links and in our case, there are nine different types of events that were acquired by Cisco Photonics in Italy. We attempt to adopt deep learning vision techniques and architectures in a way to be suitable for one-dimensional signals. Our contribution is to design a retrieval system where we can query similar events from a large collection of signals. We propose three deep learning architectures based on Convolutional Autoencoder, Convolutional Neural Network for classification and Triplet Semi-hard Network for different training scenarios namely supervised and unsupervised. There are two principal challenges that we need to take into consideration for designing our methods: (i) our solutions have to generalize on unknown categories of events, that are not presented in the training set (ii) and methods are required to handle varying size signals in inference time. Our solutions demonstrate considerable performance and promising results in the retrieval of OTDR events.| File | Dimensione | Formato | |
|---|---|---|---|
|
Thesis.pdf
accessibile in internet per tutti
Descrizione: Final revised version
Dimensione
4 MB
Formato
Adobe PDF
|
4 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/191933