Deep Learning for Predictive Maintenance: Optimizing Dynamic Time-Dependent Data Streams with Cost Function Analysis

Abstract

This thesis delves into the transformative role of deep learning techniques in predictive maintenance, with a focused investigation on the cost functions used in the evaluation and optimization of predictive models. Both linear and nonlinear forms of the cost function are explored to enhance the performance and precision of machine learning models in predictive maintenance scenarios. The study demonstrates how these cost functions can be tailored to effectively predict equipment failure, whether through binary classification for failure detection or more complex multi-class classification tasks. The research underscores the importance of cost function selection in balancing accuracy and computational efficiency, offering practical insights for industries reliant on continuous operations. By improving early detection of failures, this work aims to minimize downtime and prolong the operational life of machinery, ultimately reducing maintenance costs and increasing overall system reliability.