Multimodal learning using large language models to improve transient identification of nuclear power plants

Abstract

Transients are events that cause nuclear power plants (NPPs) to transition from a normal state to an abnormal state, which can lead to severe accidents if not properly handled. Transient identification is crucial for NPPs’ safety and operation. In this paper, we propose a novel multimodal text-time series learning framework(MTTL), the first work to apply a large language model for transient identification. The MTTL consists of self-supervised learning pre-training and zero-shot classification for transient identification. During pre-training, the framework utilizes a large language model(LLM) and a time-series(TS) encoder to fully exploit the rich multimodal information available in NPPs, i.e., to obtain the embeddings of both text data and TS data. The LLM is used to capture the transient knowledge of the NPPs by learning from the text data, and the TS encoder is used to capture the temporal dependencies of the transients by encoding the TS data. Both the LLM and the TS encoder have a linear projection head to map the embeddings into a common space. The similarity between the embeddings of the text and TS data is calculated to minimize the contrastive learning loss and obtain a pre-trained model with rich transient knowledge. During the zero-shot classification, the framework utilizes a pre-trained model to effectively identify real-world NPP transients where the data is different from the pre-trained simulated data. The proposed framework is evaluated on the High-Temperature Reactor-Pebblebed Modules (HTR-PM) plant, and the results demonstrate that the MTTL outperforms several baseline methods, including Transformer, LSTM, and CNN1D. The better zero-shot transient identification capability makes it possible to perform better in real-world NPPs.