A new benchmark shows the AI models that predict sensor readings in industrial and vehicle systems fall apart the moment a real sensor glitches.
Researchers built MuViS-C, the first multi-domain benchmark testing robustness of virtual sensing models against sensor failures. Virtual sensing predicts hard-to-measure values, like a pressure or temperature reading, from other available sensor data, which matters when installing a physical sensor isn't practical. The benchmark introduces ten failure modes, from subtle signal drift to total dropouts, tested at multiple severities across nine datasets spanning six domains. The team evaluated several machine-learning approaches, including gradient-boosted trees alongside convolutional, recurrent, attention-based, and MLP-mixing sequence models, then tested three fixes specifically on the attention-based model.
Virtual sensing already runs inside control systems for vehicles, factories, and power grids, where a bad estimate can cascade into a real failure. The study found every model tested got worse than a simple naive guess in at least one corruption scenario, and gradient-boosted trees, not the fancier deep-learning architectures, held up best overall.
The fixes that helped the attention-based model catch up all came at a cost to its everyday accuracy, which is the trade-off nobody puts on the marketing slide.