CBL - Campus del Baix Llobregat

Projecte llegit

Títol: Harnessing Data-Driven Mechanisms for Shape Engineering

Estudiants que han llegit aquest projecte:

Director/a: RYZHAKOV, PAVEL

Departament: DECA

Títol: Harnessing Data-Driven Mechanisms for Shape Engineering

Data inici oferta: 13-01-2025     Data finalització oferta: 13-09-2025


Estudis d'assignació del projecte:
    GR ENG SIS TELECOMUN
    GR ENG SIST AEROESP
    GR ENG TELEMÀTICA
Tipus: Individual
 
Lloc de realització:
UPC
    Departament: DECA
 
Segon director/a extern: Angela Ares de Parga
 
Paraules clau:
Data science, neural networks, machine learning
 
Descripció del contingut i pla d'activitats:
Numerous devices operate using parameters that define voltage signals in different branches of engineering. Depending on the specific application, such as in the design of utensil components, proper manipulation of these signals leads to the desired geometric shapes more easily and accurately. In this sense, this proposal seeks to implement data-based machine learning models, which allow obtaining the best configuration of a signal to obtain outputs of particular shapes. The techniques to be studied and developed can be applied to different engineering areas, such as aerospace and telecommunications, among others.

Objectives and Tasks
Learn relevant literature on machine learning predictive models.
Understand the fundamentals of data-driven approaches , such as correlation analysis and clustering methods, to effectively clean, analyze and interpret data sets.
Get familiar with libraries like TensorFlow and PyTorch.
Develop an initial model to predict shape characteristics that can be implemented in future applications in different branches of engineering.

Responsibilities

This project will be supervised by Professor Pavel Ryzhakov, who will guide and oversee the progress of the tasks.

Note: The proposed plan is subject to adjustments and refinements based on the availability of resources and consultation with the supervisor.

The TFG will be supervised by Prof. P. Ryzhakov and Dr. A. Ares de Parga.
 
Overview (resum en anglès):
In current engineering applications, it is common to encounter a wide range of devices driven by voltage signals. Among these, many generate outputs that can be interpreted as geometrical components, whose morphology is strongly influenced by the characteristics of the voltage signal controlling the actuator. Throughout this thesis, this scenario is examined by applying methodologies aimed at facilitating the optimization of voltage signal inputs to achieve a desired output.
This objective is pursued through several specific goals, including acquiring a solid understanding of modern predictive modelling, comprehending the fundamentals of data-driven approaches and developing an initial model capable of predicting shape-related characteristics.
The specific case study of this project focuses on a publicly available repository database that links a parameterized voltage signal-composed of three main pulses, each defined by its duration, amplitude, and inter-pulse delay-as the input, with the resulting geometric parameters derived from the characterization of droplets produced by a piezoelectric dispenser driven by that voltage signal as the output. First, the database was prepared through a cleaning process in which cluster analysis was applied to refine the data for subsequent use in constructing predictive models based on neural networks. Using the processed database, a neural network was then trained to predict two output parameters-defining a geometrical object simplified as an ellipse-from eight input parameters that characterize the voltage signal. Subsequently, the most and least influential input variables in the output formation were identified through correlation analysis.
Based on this information, two additional predictive models were developed to estimate the geometrical outputs describing the ellipse once again: one trained using the three most influential inputs, and another trained with six inputs after removing the two least influential variables. This project made it possible to explore whether predictive models based on voltage signals could be effectively constructed using a reduced number of inputs. The comparison between the reduced-input models and the full-input model demonstrated that the predictive capability remained comparable in terms of precision, validating the efficiency of the reduced-input approach.

© CBLTIC Campus del Baix Llobregat - UPC