Generative models have been designed to discover and learn the latent structure of the input data in order to generate new samples based on the regularities discovered in the data. Starting from the first simplest models such as the Restricted Boltzmann Machines up to the Variational Autoencoders and Generative Adversarial Networks (or GAN), these models have experienced a surprising development in generating data as similar to reality as possible. The potential of these models, especially in Deep Learning, has led to the most disparate applications: generation of images, videos and music, image-to-image translations, text-to-image translation and conversion of low resolution images to high resolution, to name a few. In this thesis work, carried out during the internship period of the Master's Degree, the main focus is on GANs, a generative model that makes use of the principles of supervised training through the use of two competing "sub models": a generator, trained to produce new realistic samples, and a discriminator, which tries to distinguish between real and generated data. Usually, when this model is employed, the focus is mainly on the role of the generator used to produce new data. In this case, however, the idea is to use the discriminator as a binary classifier in the context of Predictive Maintenance of a packaging machine. In other words, the discriminator obtained as a result of GAN training is used to classify the state of the machine as either normal or critical. After an initial pre-processing and exploration of the datasets, the results obtained are compared with other classifiers. Finally, the limits and possible developments of this approach are discussed.

Generative models have been designed to discover and learn the latent structure of the input data in order to generate new samples based on the regularities discovered in the data. Starting from the first simplest models such as the Restricted Boltzmann Machines up to the Variational Autoencoders and Generative Adversarial Networks (or GAN), these models have experienced a surprising development in generating data as similar to reality as possible. The potential of these models, especially in Deep Learning, has led to the most disparate applications: generation of images, videos and music, image-to-image translations, text-to-image translation and conversion of low resolution images to high resolution, to name a few. In this thesis work, carried out during the internship period of the Master's Degree, the main focus is on GANs, a generative model that makes use of the principles of supervised training through the use of two competing "sub models": a generator, trained to produce new realistic samples, and a discriminator, which tries to distinguish between real and generated data. Usually, when this model is employed, the focus is mainly on the role of the generator used to produce new data. In this case, however, the idea is to use the discriminator as a binary classifier in the context of Predictive Maintenance of a packaging machine. In other words, the discriminator obtained as a result of GAN training is used to classify the state of the machine as either normal or critical. After an initial pre-processing and exploration of the datasets, the results obtained are compared with other classifiers. Finally, the limits and possible developments of this approach are discussed.

Generative adversarial network for predictive maintenance of a packaging machine

RASETTA, ADRIANO
2021/2022

Abstract

Generative models have been designed to discover and learn the latent structure of the input data in order to generate new samples based on the regularities discovered in the data. Starting from the first simplest models such as the Restricted Boltzmann Machines up to the Variational Autoencoders and Generative Adversarial Networks (or GAN), these models have experienced a surprising development in generating data as similar to reality as possible. The potential of these models, especially in Deep Learning, has led to the most disparate applications: generation of images, videos and music, image-to-image translations, text-to-image translation and conversion of low resolution images to high resolution, to name a few. In this thesis work, carried out during the internship period of the Master's Degree, the main focus is on GANs, a generative model that makes use of the principles of supervised training through the use of two competing "sub models": a generator, trained to produce new realistic samples, and a discriminator, which tries to distinguish between real and generated data. Usually, when this model is employed, the focus is mainly on the role of the generator used to produce new data. In this case, however, the idea is to use the discriminator as a binary classifier in the context of Predictive Maintenance of a packaging machine. In other words, the discriminator obtained as a result of GAN training is used to classify the state of the machine as either normal or critical. After an initial pre-processing and exploration of the datasets, the results obtained are compared with other classifiers. Finally, the limits and possible developments of this approach are discussed.
2021
Generative adversarial network for predictive maintenance of a packaging machine
Generative models have been designed to discover and learn the latent structure of the input data in order to generate new samples based on the regularities discovered in the data. Starting from the first simplest models such as the Restricted Boltzmann Machines up to the Variational Autoencoders and Generative Adversarial Networks (or GAN), these models have experienced a surprising development in generating data as similar to reality as possible. The potential of these models, especially in Deep Learning, has led to the most disparate applications: generation of images, videos and music, image-to-image translations, text-to-image translation and conversion of low resolution images to high resolution, to name a few. In this thesis work, carried out during the internship period of the Master's Degree, the main focus is on GANs, a generative model that makes use of the principles of supervised training through the use of two competing "sub models": a generator, trained to produce new realistic samples, and a discriminator, which tries to distinguish between real and generated data. Usually, when this model is employed, the focus is mainly on the role of the generator used to produce new data. In this case, however, the idea is to use the discriminator as a binary classifier in the context of Predictive Maintenance of a packaging machine. In other words, the discriminator obtained as a result of GAN training is used to classify the state of the machine as either normal or critical. After an initial pre-processing and exploration of the datasets, the results obtained are compared with other classifiers. Finally, the limits and possible developments of this approach are discussed.
generative network
anomaly detection
classification
File in questo prodotto:
File Dimensione Formato  
Rasetta_Adriano.pdf

accesso aperto

Dimensione 2.49 MB
Formato Adobe PDF
2.49 MB Adobe PDF Visualizza/Apri

The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/31830