Health Index Estimation and RUL Prediction on Gear Pumps

Abstract

This thesis explores innovative machine learning and signal processing techniques to accurately estimate the wear by constructing Health Indexes and predict the Remaining Useful Life (RUL) of gear pumps used in ABB Robotics’ car painting robots. The study investigates pressure sensor data obtained from the gear pumps and identifies changes in the data as wear progresses.

Two methods are proposed for Health Index estimation: One leveraging Fast Fourier Transform (FFT) values in a weighted sum of the significant frequencies and another utilizing Convolutional Neural Network Autoencoders (CNN-AE) supplemented with an outlier detection model (One Class SVM / Isolation Forest). The constructed Health indexes for both methods successfully represent the degradation observed on the pumps.

The proposed RUL models utilize these constructed Health Indexes to make predictions, resulting in improved predictions compared to using only the current lifetime as input. Using the feature vector from a CNN-AE network supplied with FFT values as input to a One-Class SVM model for Health Index estimation and then utilizing these in a Bidirectional LSTM model yields the best RUL predictions. However, the other proposed methods are also performing satisfactorily.

This study contributes to the field of predictive maintenance for gear pumps and demonstrate the effectiveness of data-driven techniques in estimating wear and predicting RUL in this context.