An Ensemble Learning Methodology for Predicting Medical Micro-robot Degradation Classes

The new generation of medical devices for surgical operation involves to develop equipment on the scale from micrometers to millimeters in order to perform more precise microsurgical procedures. Such kinds of devices must fulfill several tests according to medical standards to be commercialized. Hence, it is necessary to model the micro-robot degradation in order to ensure the optimal performance limits during surgical acts. The work aims to predict a micro-robot degradation class using a machine-learning-based methodology and consists of classifying the degradation into three classes: healthy, degraded, and out of service. Firstly, the degraded data are collected by using a four-bar complaint mechanism. This mechanism allows obtaining relevant attributes for the micro-robot degradation behavior. Secondly, a data preprocessing analysis and feature engineering are conducted to generate representative attributes that provide a better learning representation for the machine learning (ML) algorithm. Then, non-linear supervised learning algorithms are trained to construct the prediction. Random forest outperforms other algorithms in terms of predicting the remaining useful life (RUL) while gradient boosting generates the optimal decision boundary for classification using the RUL and features generated by autoencoders in presence of noise. Finally, a pipeline for the classification of the micro-robot degradation state is provided. This methodology ensures a procedure that evaluates whether or not the ML model can represent the underlying system in presence of noise.