Comparison of Error Measures and Machine Learning Methods for Strain-Based Structural Health Monitoring

Abstract

The development of aircraft structures requires many fatigue tests. These tests are usually carried out to validate the corresponding finite element and damage models and to prove the expected damage-tolerant behavior. Monitoring aircraft structures requires experienced staff and is very time-consuming and expensive as the recurring inspection of the structure is a tedious task. We propose a machine learning-based approach that exploits continuous load and strain measurement data to support structural health monitoring and to shift the inspection program towards predictive maintenance. The machine learning model is used for mapping loads onto local strains. With the trained model, different error measures between current measurements and the predicted values are determined. When a specific threshold value based on an error confidence level is exceeded, an alarm is set off, and appropriate actions can be taken. The approach is applied to several fatigue tests with two different types of structures and damage mechanisms. Various error measures and models are compared. The paper shows that, first, simple error measures, such as the root mean squared error, are sufficient and even outperform more sophisticated error distances for detecting cracks with continuous strain measurements. Second, the standard deviation of strain or rather the load-strain slope is a key feature to detect cracks. And third, machine learning models enable structural health monitoring with sensors that even have only small strain values.