← All projects

Bearing RUL Prediction

Deep learning models that estimate the remaining useful life of rolling element bearings from run-to-failure vibration data.

Problem
Estimate how much life a bearing has left, framed as a sequence-to-value regression on vibration data.
Data
UNSW bearing run-to-failure dataset (outer-race defects), horizontal acceleration, published on Hugging Face.
Approach
An RMS health indicator per frame, smoothed and paired with a normalised linear RUL target. RNN, LSTM and WaveNet models are trained and compared.
Tools
Python, TensorFlow, Hugging Face Datasets, Librosa, scikit-learn
Vibration from two bearing run-to-failure tests: Test 1 used as training data, Test 3 used to estimate RUL from its current state
Test 1 is used as training data. For Test 3, the model estimates the remaining useful life from the current state.