Researchers have built a single Transformer model that can read vibration and sensor data from wildly different machines and flag faults without needing to be retrained for each one.
The model, called FreqCondNorm, uses a FiLM-style frequency-conditioned normalization layer to handle signals that range from 1 Hz to roughly 100 kHz in sampling rate, a gap that normally breaks cross-machine models. It was pretrained on five public predictive-maintenance datasets (CWRU, MFPT, UOC18, PRONOSTIA, CMAPSS) using masked auto-encoding and contrastive learning, with balanced sampling across domains so no single dataset dominates training. On fault diagnosis, it hit 99.2% accuracy on CWRU, a 6.4 percentage-point jump over a CNN baseline, and 82.1% zero-shot accuracy on MFPT, meaning it had never seen that dataset's machines during training.
That zero-shot number is the real story. Predictive maintenance has long been stuck retraining bespoke models per machine because labeled failure data is scarce and sensor setups vary wildly. A model that generalizes across sampling frequencies could let a factory with mixed equipment run one system instead of a dozen.
The catch: FreqCondNorm did not improve remaining useful life prediction, the part of maintenance that actually tells you when to schedule a repair. Spotting that a bearing is failing is not the same as knowing how many days it has left, and on that harder question, this model has nothing new to offer yet.