A Foundation Model for Sensor Data with Prompt-Based Adaptation Across Machines and Plants

Abstract

The proliferation of Industrial Internet of Things (IIoT) devices has generated vast quantities of high-frequency sensor data, yet the effective utilization of this data remains hindered by the heterogeneity of machinery and the variability of operating environments. Traditional deep learning approaches typically require training specialized models for specific assets, a process that is computationally expensive and scales poorly across diverse manufacturing plants. This paper introduces Sensor-FM, a foundation model architecture designed for general-purpose representation learning on industrial time-series data. Unlike conventional transfer learning methods that rely on extensive fine-tuning of model weights, Sensor-FM utilizes a prompt-based adaptation mechanism. By injecting learnable, context-specific vectors into the frozen pre-trained transformer latent space, the model adapts to novel machines and distinct plant environments with minimal data requirements. We demonstrate that this parameter-efficient approach achieves state-of-the-art performance in anomaly detection and remaining useful life (RUL) estimation tasks. Experimental results indicate that prompt tuning requires less than 1% of the trainable parameters compared to full model fine-tuning while exhibiting superior robustness against domain shifts caused by operational discrepancies. Our findings suggest a paradigm shift in industrial AI, moving from bespoke modeling to a centralized, adaptable foundation approach.