Predictive maintenance

Problem Statement Rotating machinery within our installations can fail unexpectedly due to bearing wear, imbalance, or misalignment, while current monitoring remains largely reactive and underutilizes the rich information contained in existing sensor data (flow, pressure, vibration). There is no data-driven, generalizable method to detect failure indicators early, applicable both at the short-term signal level and the long-term degradation-trend level, that can be practically validated and integrated into the EVOLV software stack. EVOLV = Own software digital twins on IBM's Node-Red.

Proposed Approach High-frequency sensor data could be collected but should be investigated first, on our internal pilot setups and converted via FFT into spectral features, which serve as input for Machine Learning models (ranging from classical classifiers to LSTM autoencoders) for anomaly detection and failure prediction. Models are first trained and validated on the pilot setups—with specific attention to class imbalance, feature selection, and transfer learning across machine types—before designing an end-to-end pipeline (including drift detection and explainable AI) for integration and real-time application within EVOLV.

Expected Outcome A validated, generalizable ML model that detects early failure signals in rotating equipment based on pilot data, along with a concrete design for a scalable, interpretable data and inference pipeline within EVOLV that enables the transition from pilot setup to field deployment.