Is this project an undergraduate, graduate, or faculty project?
Undergraduate
Project Type
group
Campus
Daytona Beach
Authors' Class Standing
Anayah Smith, Senior Victoria Gaibor, Senior
Lead Presenter's Name
Anayah Smith
Lead Presenter's College
DB College of Arts and Sciences
Faculty Mentor Name
Dumindu Samaraweera
Abstract
Accurate prediction of Remaining Useful Life (RUL) is critical for enabling predictive maintenance, improving system reliability, and reducing operational costs in degrading systems. This project addresses the problem of modeling and predicting RUL using multivariate time-series sensor data from the NASA CMAPSS turbofan engine dataset, with a focus on understanding how predictive performance changes across datasets of varying complexity. The objective is to develop a reproducible machine learning pipeline that captures degradation patterns and produces reliable time-to-failure predictions. The approach includes data preprocessing, exploratory data analysis, feature engineering, dimensionality reduction, and model evaluation. RUL values are computed and capped to stabilize regression, and feature selection is performed using Principal Component Analysis (PCA) and Recursive Feature Elimination (RFE) to create a reduced sensor set. A group-aware data splitting strategy (60% training, 20% validation, 20% test) is implemented using GroupShuffleSplit to prevent data leakage across engine units. A Ridge Regression model is used as a baseline with hyperparameter tuning via GridSearchCV and GroupKFold cross-validation, and additional models, including Gradient Boosting, XGBoost, Support Vector Regression, Neural Networks, and a Hidden Markov Model, are evaluated for comparison. Preliminary results indicate that the reduced feature set maintains comparable performance to the full sensor set, while nonlinear models improve predictive accuracy over the linear baseline. This ongoing work contributes a structured, reproducible framework for RUL prediction and supports analysis of model performance, feature selection, and system complexity in multivariate time-series data.
Did this research project receive funding support (Spark, SURF, Research Abroad, Student Internal Grants, Collaborative, Climbing, or Ignite Grants) from the Office of Undergraduate Research?
No
Predicting Remaining Useful Life Using Multivariate Time-Series Data
Accurate prediction of Remaining Useful Life (RUL) is critical for enabling predictive maintenance, improving system reliability, and reducing operational costs in degrading systems. This project addresses the problem of modeling and predicting RUL using multivariate time-series sensor data from the NASA CMAPSS turbofan engine dataset, with a focus on understanding how predictive performance changes across datasets of varying complexity. The objective is to develop a reproducible machine learning pipeline that captures degradation patterns and produces reliable time-to-failure predictions. The approach includes data preprocessing, exploratory data analysis, feature engineering, dimensionality reduction, and model evaluation. RUL values are computed and capped to stabilize regression, and feature selection is performed using Principal Component Analysis (PCA) and Recursive Feature Elimination (RFE) to create a reduced sensor set. A group-aware data splitting strategy (60% training, 20% validation, 20% test) is implemented using GroupShuffleSplit to prevent data leakage across engine units. A Ridge Regression model is used as a baseline with hyperparameter tuning via GridSearchCV and GroupKFold cross-validation, and additional models, including Gradient Boosting, XGBoost, Support Vector Regression, Neural Networks, and a Hidden Markov Model, are evaluated for comparison. Preliminary results indicate that the reduced feature set maintains comparable performance to the full sensor set, while nonlinear models improve predictive accuracy over the linear baseline. This ongoing work contributes a structured, reproducible framework for RUL prediction and supports analysis of model performance, feature selection, and system complexity in multivariate time-series data.