Machine Learning Models for Predicting Pitting Conditions

This document is a technical report detailing an approach to predict outcomes in the 2023 PHM North America Data Challenge using machine learning models. The report outlines a methodology that incorporates both supervised and unsupervised learning techniques to address the challenge of predicting unseen labels. The approach consists of five main steps, starting with defining the problem type, which involves multi-dimensional vibration time series data from gear pitting experiments. The report describes the use of cross-validation methods to partition the training data effectively, ensuring that the model can generalize well to new data. It explains the creation of a classification model using features extracted from time series data and discusses the performance of different machine learning techniques, including gradient boosting and neural networks. The report also details the prediction process for unseen labels using a nearest neighbor clustering method, providing insights into the model's behavior and the clustering results. Overall, the document presents a comprehensive framework for tackling the data challenge using advanced machine learning techniques.