Banca de DEFESA: ALCEMY GABRIEL VITOR SEVERINO

Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.
STUDENT : ALCEMY GABRIEL VITOR SEVERINO
DATE: 20/11/2023
TIME: 09:00
LOCAL: Sala virtual Google Meet
TITLE:

Representation Based on Stacked Autoenconder Optimized by Particle Swarm

KEY WORDS:

Particle Swarm Optimization, Soft Sensors, Deep Learning, Stacked Au- toencoders, Mutual Information.

PAGES: 71
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUBÁREA: Eletrônica Industrial, Sistemas e Controles Eletrônicos
SPECIALTY: Controle de Processos Eletrônicos, Retroalimentação
SUMMARY:

Market competitiveness drives organizations to pursue technological development in order to improve product quality and reduce production costs while meeting the socio- environmental demands of consumers. However, industrial processes can pose challenges in real-time monitoring and control of critical variables. One solution to this problem is the use of soft sensors, which are algorithms capable of estimating difficult-to-measure variables based on easily measurable secondary variables. A common challenge in soft sensor projects is the lack of labeled data, making semi-supervised methods more promi- sing than traditional methods. In this context, the Stacked Autoencoder neural network architecture has been widely employed. This architecture is trained in an unsupervised manner and subsequently fine-tuned in a supervised manner. However, appropriately de- fining the hyperparameters of the Stacked Autoencoder, such as batch size, learning rate, and number of hidden features, presents a challenge. Traditional methods like Grid Se- arch and Random Search are computationally intensive and may not quickly find the best combination of hyperparameters. A more efficient alternative is the use of metaheu- ristic algorithms, such as Particle Swarm Optimization. These algorithms intelligently explore the search space and are more effective in high-dimensional spaces. A promising approach is to incorporate Mutual Information into the evaluation function of Particle Swarm Optimization, along with the Mean Squared Error. Mutual Information captures nonlinear relationships between the outputs of the Stacked Autoencoder and the actual system outputs, while the Mean Squared Error measures the difference between these outputs. In this context, the present thesis proposes the Representation Based on Particle Swarm Optimized Stacked AutoEncoder method, which utilizes Particle Swarm Optimi- zation with a modified evaluation function to optimize the hyperparameters of a Stacked Autoencoder-based soft sensor. It is expected that this approach will improve the accu- racy and representation capacity of the Stacked Autoencoder compared to conventional approaches that only utilize the Mean Squared Error. In order to evaluate the performance of the models generated by the proposed method, two widely used nonlinear processes in the industry were selected. These processes were chosen due to their relevance in virtual sensor implementation and are frequently employed in comparative analyses. The results demonstrate that the incorporation of Mutual Information in the evaluation function al- lows for a more efficient and balanced search, resulting in a Stacked Autoencoder with improved performance and representation capacity.

COMMITTEE MEMBERS:
Presidente - 1451883 - FABIO MENEGHETTI UGULINO DE ARAUJO
Interno - 1328152 - CARLOS EDUARDO TRABUCO DOREA
Externo ao Programa - 3374361 - JEAN MARIO MOREIRA DE LIMA - UFRNExterno à Instituição - LEANDRO LUTTIANE DA SILVA LINHARES - IFPB
Externo à Instituição - SERGIO NATAN SILVA