[1F07-11] Advanced Technology for Inspection and Maintenance of Nuclear Reactors

[1F07] Weakly Supervised Time Series Analysis for Fault Diagnosis in Nuclear Power Plants

*Feiyan Dong¹, Shi Chen¹, Kazuyuki Demachi¹ (1. UTokyo)

Keywords:Nuclear power plants, Nuclear safety, Fault diagnosis, Weakly supervised learning, Time series analysis

Fault diagnosis in rotating machinery systems is essential to management and maintenance of Nuclear Power Plants (NPPs). Nowadays, deep learning algorithms has been introduced to diagnose faults through the analysis of rotational signals. However, faults are difficult to define, sparsely occurring, and along with variable noise labels, which affect the performance of supervised and unsupervised learning. Moreover, when processing time series data, the prevailing problems such as losing temporal features and gradient vanishing in general deep learning models also increase the difficulty of fault diagnosis. To alleviate these issues, we propose a weakly supervised time series analysis framework for fault diagnosis of NPPs. The performance of the proposed model was experimentally evaluated on the benchmark dataset and the experimental results demonstrate its effectiveness on fault diagnosis tasks.