Most of the popular internet applications, e.g. social networks, streaming media and smart working platforms, lead to an exponential growth of data. The direct consequence of the network data's increase is an additional bandwidth request that forces the photonic technology to analyze new solutions, optical transceivers, in order to provide high scalability, capacity and speed rates. The QSFP28-100Gb/s optical transceivers allow to satisfy these requirements and, being very complex devices, the highest Quality of Service (QoS) is a crucial condition to be guaranteed in short time. The thesis focuses on the management of a huge quantity of data relating to the analysis of QSFP28-100G's test phases, in order to ensure the product quality and the robustness of the device. Until now, the failure identification procedure, during the testing phase, is executed by human experts, who identify root causes of the failure events and proper countermeasures in order to cope with the problem. This thesis proposes supervised and unsupervised machine learning techniques to classify and group the failed optical transceiver into multiple sub-classes, in order to propose a more specific troubleshooting and provide countermeasure procedures to fix the under-performing or non-compliant device to the specifications, in the fastest and cheapest way. Classification algorithms, e.g. Decision Trees, Logistic Regression, K-Nearest Neighbor and Support Vector Machines are implemented, and their accuracy has been calculated, showing the supremacy of the Decision Tree algorithm with 61.54% precision value. Finally, two equivalent and effective clustering techniques are implemented: K-Means and Hierarchical Clustering.
L'uso sempre più frequente di applicazioni, come ad esempio i social network, i video streaming e le piattaforme di smart-working, portano ad una continua generazione di dati. Il persistente utilizzo di servizi ed applicazioni di questo tipo, implica una ulteriore richiesta in termini di banda, che induce la fotonica a studiare nuove soluzioni, come i transceivers ottici, al fine di fornire alta scalabilità, capacità e velocità. I transceivers ottici QSFP28-100Gb/s permettono di soddisfare tali requisiti ed, essendo dispositivi molto complessi, la massima qualità del servizio (QoS) è un attributo fondamentale che deve essere garantito in breve tempo. La tesi si concentra sulla gestione di una grande quantità di dati relativi all'analisi delle fasi di test QSFP28-100G, in modo da garantire la qualità del prodotto e la robustezza del dispositivo. Fino ad ora, la procedura di identificazione degli errori è stata eseguita da esperti, che ricercano la natura degli errori e le adeguate contromisure per far fronte al problema. L'obbiettivo di questa tesi è di proporre diverse tecniche di apprendimento automatico provenienti dal settore del Machine Learning al fine di distinguere, classificare e raggruppare i transceivers ottici che non siano funzionanti o conformi alle specifiche, nel più breve tempo possibile. Sono implementati algoritmi di classificazione come Decision Trees, Logistic Regression, K-Nearest Neighbor e Support Vector Machines, ed è stata calcolata la loro accuratezza, mostrando il predominio dell'algoritmo Decision Tree, nonostante abbia una precisione del 61,54%. Infine, vengono presentate e implementate due tecniche di clustering equivalenti ed efficaci: il K-Means e lo Hierarchical Clustering.
Machine learning application to high throughput testing of optical transceivers
REOLON, ERICA
2018/2019
Abstract
Most of the popular internet applications, e.g. social networks, streaming media and smart working platforms, lead to an exponential growth of data. The direct consequence of the network data's increase is an additional bandwidth request that forces the photonic technology to analyze new solutions, optical transceivers, in order to provide high scalability, capacity and speed rates. The QSFP28-100Gb/s optical transceivers allow to satisfy these requirements and, being very complex devices, the highest Quality of Service (QoS) is a crucial condition to be guaranteed in short time. The thesis focuses on the management of a huge quantity of data relating to the analysis of QSFP28-100G's test phases, in order to ensure the product quality and the robustness of the device. Until now, the failure identification procedure, during the testing phase, is executed by human experts, who identify root causes of the failure events and proper countermeasures in order to cope with the problem. This thesis proposes supervised and unsupervised machine learning techniques to classify and group the failed optical transceiver into multiple sub-classes, in order to propose a more specific troubleshooting and provide countermeasure procedures to fix the under-performing or non-compliant device to the specifications, in the fastest and cheapest way. Classification algorithms, e.g. Decision Trees, Logistic Regression, K-Nearest Neighbor and Support Vector Machines are implemented, and their accuracy has been calculated, showing the supremacy of the Decision Tree algorithm with 61.54% precision value. Finally, two equivalent and effective clustering techniques are implemented: K-Means and Hierarchical Clustering.File | Dimensione | Formato | |
---|---|---|---|
Thesis_Reolon_Erica.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
4.05 MB
Formato
Adobe PDF
|
4.05 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/10589/165123