Title: A multi-sensor attention-enhanced temporal convolution network for hydropower plant equipment fault prediction

Authors: Ruilin Yang; Yan Jin; Jiada Wei; Liyang Jiang; Sijia Wang

Addresses: Hunan Wuling Power Technology Co., Ltd., Changsha, 410000, China; Hunan Provincial Hydropower Intelligent Engineering, Technology Research Center, Changsha, 410000, China ' Hunan Wuling Power Technology Co., Ltd., Changsha, 410000, China; Hunan Provincial Hydropower Intelligent Engineering, Technology Research Center, Changsha, 410000, China ' Hunan Wuling Power Technology Co., Ltd., Changsha, 410000, China; Hunan Provincial Hydropower Intelligent Engineering, Technology Research Center, Changsha, 410000, China ' Hunan Wuling Power Technology Co., Ltd., Changsha, 410000, China; Hunan Provincial Hydropower Intelligent Engineering, Technology Research Center, Changsha, 410000, China ' Hunan Wuling Power Technology Co., Ltd., Changsha, 410000, China; Hunan Provincial Hydropower Intelligent Engineering, Technology Research Center, Changsha, 410000, China

Abstract: Hydropower systems depend critically on the continuous operation of generators, bearings, and control equipment, which endure complex multi-physical stresses. Accurate fault prediction is crucial for preventing catastrophic failures, yet it remains challenging due to the high dimensionality of sensor data and the presence of long-term temporal dependencies. Traditional approaches cannot model long-range patterns and dynamically weight sensor features. To overcome these challenges, we propose an attention-enhanced temporal convolutional network that integrates dilated convolutions for efficient long-sequence modelling, along with a multi-head self-attention mechanism for adaptive feature selection. The model captures multi-scale temporal features and dynamically emphasises informative sensors and time steps. Evaluated on the National Aeronautics and Space Administration Commercial Modular Aero-Propulsion System Simulation dataset, our method achieves a root mean square error of 12.56 and a prognostic score of 356, outperforming the vanilla temporal convolutional network by 8.7% and 15.3%, respectively.

Keywords: fault prediction; temporal convolutional network; TCN; attention mechanism; multi-sensor data; remaining useful life; RUL.

DOI: 10.1504/IJSNET.2026.153834

International Journal of Sensor Networks, 2026 Vol.51 No.1, pp.16 - 29

Received: 01 Sep 2025
Accepted: 14 Nov 2025
Published online: 27 May 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article