TY - GEN
T1 - A Federated Label Anchoring Method for Fault Diagnosis of Rotating Machinery with Data Privacy
AU - Gao, Zengguang
AU - Ma, Xintong
AU - Nie, Zelin
AU - Du, Shuhong
AU - Cheng, Wei
AU - Miao, Zhuang
AU - Li, Li
AU - Chen, Xuefeng
N1 - Publisher Copyright:
© 2025 IEEE.
PY - 2025
Y1 - 2025
N2 - In practical industrial environments, data silos remain a persistent challenge in rotating machinery fault diagnosis, primarily due to concerns surrounding data protection and security. In industrial settings, the quantity and quality of monitor data available at an individual client level are often insufficient to support the development of robust diagnostic models. Enter Federated Learning (FL)-a distributed training paradigm that combines locally trained model updates to create a comprehensive shared model without compromising raw data. However, traditional FL methods often suffer from a significant drop in accuracy when faced with the non-uniform and heterogeneous nature of client data commonly observed in industrial scenarios. To tackle this problem, we propose FedAnchor, a federated label anchoring approach tailored for heterogeneous monitor datasets distributed across multiple clients. The feature extractor built by this method captures the feature representations of the data, while the label encoder constructed based on a Large Language Model (LLM) generates label embeddings. These label embeddings are used to anchor the data, addressing the inherent heterogeneity of client data. This method can also assign pseudo-labels to unlabeled data, increasing the available data for model training. The proposed method's efficacy is confirmed through rigorous testing on two benchmark datasets: the Drivetrain Dynamics Simulator (DDS) and nuclear power pump systems. Comparative experiments demonstrate that this approach consistently surpasses state-of-the-art federated learning techniques across diverse data distribution scenarios, proving its superior performance.
AB - In practical industrial environments, data silos remain a persistent challenge in rotating machinery fault diagnosis, primarily due to concerns surrounding data protection and security. In industrial settings, the quantity and quality of monitor data available at an individual client level are often insufficient to support the development of robust diagnostic models. Enter Federated Learning (FL)-a distributed training paradigm that combines locally trained model updates to create a comprehensive shared model without compromising raw data. However, traditional FL methods often suffer from a significant drop in accuracy when faced with the non-uniform and heterogeneous nature of client data commonly observed in industrial scenarios. To tackle this problem, we propose FedAnchor, a federated label anchoring approach tailored for heterogeneous monitor datasets distributed across multiple clients. The feature extractor built by this method captures the feature representations of the data, while the label encoder constructed based on a Large Language Model (LLM) generates label embeddings. These label embeddings are used to anchor the data, addressing the inherent heterogeneity of client data. This method can also assign pseudo-labels to unlabeled data, increasing the available data for model training. The proposed method's efficacy is confirmed through rigorous testing on two benchmark datasets: the Drivetrain Dynamics Simulator (DDS) and nuclear power pump systems. Comparative experiments demonstrate that this approach consistently surpasses state-of-the-art federated learning techniques across diverse data distribution scenarios, proving its superior performance.
KW - Data Heterogeneity
KW - Fault Diagnosis
KW - Federated Learning
KW - Large Language Model
UR - https://www.scopus.com/pages/publications/105015948842
U2 - 10.1109/CACRE66141.2025.11119549
DO - 10.1109/CACRE66141.2025.11119549
M3 - 会议稿件
AN - SCOPUS:105015948842
T3 - Proceedings - 2025 10th International Conference on Automation, Control and Robotics Engineering, CACRE 2025
SP - 332
EP - 337
BT - Proceedings - 2025 10th International Conference on Automation, Control and Robotics Engineering, CACRE 2025
A2 - Zhang, Fuming
A2 - Chai, Li
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 10th International Conference on Automation, Control and Robotics Engineering, CACRE 2025
Y2 - 16 July 2025 through 19 July 2025
ER -