Architecting Asset Specific Time Series Foundation Model and its Applications for Asset Performance Management

This paper presents a time series foundation model (TSFM) designed for asset performance management (APM) of industrial equipment such as pumps and compressors. Leveraging transformer-based architecture and a discrete tokenization scheme for multivariate sensor data, the TSFM is pretrained on a broad corpus of operational and simulated time series data. It is then fine-tuned on asset-specific data to adapt to local conditions with minimal effort. The model supports multiple tasks (e.g., forecasting and anomaly detection) using task-specific heads atop a shared representation. Experimental results on electric submersible pump (ESP) data demonstrate improved performance over traditional LSTM and threshold-based models, with a 10% reduction in RMSE and 8% reduction in MAPE for forecasting, and an F1-score improvement from 0.72 to 0.80 for anomaly detection. A field case study illustrates early failure detection, enabling proactive maintenance. This work establishes a versatile, scalable, and interpretable FM-based framework for APM, reducing the need for tailored models per asset.