Machine learning approaches for predictive maintenance in industrial manufacturing

Abstract

This study investigates the application of machine learning techniques for predictive maintenance in manufacturing environments. We compare Random Forest, XGBoost, and LSTM-based models on vibration sensor data collected from CNC milling machines over 18 months. The LSTM model achieves 96.3% accuracy in predicting bearing failures 48 hours in advance, significantly reducing unplanned downtime and maintenance costs by an estimated 34%.

Topics

Full Text Preview

Vibration data was collected from 24 Siemens Sinumerik CNC milling machines equipped with PCB Piezotronics 352C33 accelerometers sampling at 20 kHz. Features were extracted using FFT, wavelet decomposition, and statistical measures (RMS, kurtosis, crest factor). The LSTM network with 2 layers of 128 units and dropout of 0.3 was trained on 14 months of data and validated on 4 months of held-out operational data.

Published Through