Fault diagnosis based on gated recurrent unit network with attention mechanism and transfer learning under few samples in nuclear power plants

Abstract

Fault diagnosis (FD) of rotating machines is critical to the safety and economic operation of nuclear power plants (NPPs). Gated Recurrent Unit (GRU) is a gating mechanism in recurrent neural network and is a deep learning model that excels in processing sequential information and can be used to learn potential fault features in the condition monitoring data for FD. However, lack of sufficient fault samples (i.e., few samples) in NPPs prevents the GRU network from being adequately trained, resulting in poor performance. This study proposes a new GRU network combined with attention mechanism (AM) and transfer learning (TL), called GRU-AM-TL method. The attention layer is introduced to adaptively assign different weights to the extracted features for discrepant processing and enhancing focus on valuable information. The TL strategy tries to make full use of diagnosis knowledge learned from relevant fault datasets under different operating conditions, different machines or different fault severity for improving new diagnosis task under few samples. The specific FD target is to identify the fault nature (location, size or severity) by pattern recognition. Bearing, gearbox and NPP simulated fault datasets are used to validate the proposed method. Case study shows that the AM and TL strategy can help GRU network improve diagnosis accuracy under few sample scenarios. Moreover, the proposed GRU-AM-TL method can achieve the best performance in all test cases compared with GRU-based methods and other classical methods, such as convolutional neural network, support vector machine and random forest, showing good FD advantage in NPPs.