Many industrial rotating machinery operate in harsh environments, and due to the limitations of existing data acquisition systems, collecting large volumes of labeled data for training intelligent fault diagnosis models in real-world operational scenarios is often impractical. To address this challenge of data scarcity, this study takes pumps as a typical example and proposes an intelligent fault diagnosis model based on digital twin-assisted deep transfer learning. The model utilizes patch time series Transformer (patchTST) to extract features from simulated one-dimensional time-series data and combines it with limited real-world data. Through domain adversarial adaptation and fine-tuning strategies, a fault diagnosis model is constructed. The practicality of the model has been verified to effectively reduce the reliance on real labeled data required by conventional deep learning diagnosis models using the Kumar centrifugal pump dataset.

Digital twin,Transformer,Domain adversarial,Fine tune,Fault diagnosis