Intellectual Property India Publishes Patent Application for Federated Learning-Based Bearing Fault Detection In Centrifugal Pumps

MUMBAI, India, July 24 -- Intellectual Property India has published a patent application (202641085349 A) filed by Sri Eshwar College Of Engineering on July 11, 2026, for Federated Learning-Based Bearing Fault Detection In Centrifugal Pumps.

Inventors include T Suruthikrishna; and L. Priya.

The application for the patent was published on July 17, 2026, under issue no. 29/2026.

Abstract: The present invention concerns a privacy preserving fault diagnosis system for centrifugal pump bearings, based on Federated Learning (FL) and vibration signal analysis. In industrial applications, centrifugal pumps are extensively used, and if the following defects are not found in the initial stages, it will cause loss of efficiency and unexpected failures: Bearing wear, misalignment, imbalance. The conventional fault diagnosis systems gather vibration information from various pumps and store it in a central server resulting in high communication cost, storage load and privacy problems. To tackle these challenges, the proposed system introduces Federated Learning, a distributed learning approach in which the vibration information stays at local monitoring nodes, while the model parameters are transmitted to a central aggregation server. In the proposed system, it is proposed to collect the vibration signals from the centrifugal pump bearings and distribute the vibration signals among multiple local clients. A local fault diagnosis model is built for each client and updated with the data of the client and then sent to a central server for aggregation to a global model. This work compares three federated learning algorithms: Federated Averaging (FedAvg), Federated Proximal (FedProx), and SCAFFOLD in the centrifugal pump bearing fault diagnosis. The experimental results demonstrate that FedAvg had a training accuracy of 94.27% and a testing accuracy of 70.97%, whereas FedProx had a training accuracy of 98.53% and a testing accuracy of 85.71%, and SCAFFOLD had a training accuracy of 98.61% and a testing accuracy of 77.03%. The overall performance of the three methods was best for FedProx, which had superior convergence stability and handled non-IID data better. Thus, the proposed system is an efficient and privacy-preserving approach for predictive maintenance and fault diagnosis of centrifugal pumps in industrial and IIoT applications.

Disclaimer: Curated by HT Syndication.