An AI Framework for Predictive Maintenance with a Foundation Rooted in Physics-Based Principals

Abstract

In the field of production in general, the physical models that govern degradation are only effective in particular cases, but they are highly explanatory. On the other hand, machine learning models, although working under all conditions with high accuracy, remain unexplained due to their complexity, posing a challenge for engineers in the field. This article presents a symbolic approach to modeling failure modes from data. This approach integrates causal analysis, performance level estimation and risk analysis based on operational safety parameters evaluated according to the criteria of severity, frequency and probability. Using techniques such as deep learning (artificial intelligence), parsimony-driven model selection and symbolic regression, the aim is to minimize a variance function using defined operators. The end result is a simple, accurate function that works under all conditions, while remaining explainable to domain engineers and preserving the predictive capabilities of the neural network.