Machine learning models such as artificial neural networks (ANNs) are frequently used in automation. Despite their advantages, their opacity limits their deployment in environments requiring strict functional safety and transparent behavior. This paper presents a methodology to substitute predictive ANNs with intrinsically interpretable mathematical surrogate models. The procedure employs symbolic regression to directly approximate the output of a trained ANN. This knowledge extraction process treats the original ANN as an observable system without requiring access to its internal architecture or original training data. The approach is validated using a seven-joint serial manipulator performing an assembly task. The resulting mathematical expressions approximate the original ANN with high precision while being globally and locally interpretable. These explicit equations allow engineers to apply classical system theory to analyze stability and ensure compliance with safety requirements.