隨著全球對低碳能源的需求迫切，核能作為基載電力的重要性不言可喻，核電廠控制室在面臨異常事故時，常因大量湧入的警報訊號引發「警報洪水（Alarm Flood）」現象，嚴重影響操作員的決策判斷，因此發展一套具備高即時性與高準確度的故障診斷與健康管理（PHM）系統，對於確保核安至關重要。 本研究旨在解決傳統序列模型在處理大規模警報數據時的效率瓶頸，在數據層面，本研究採用 IEEE PCTRAN 開放資料集作為基準，該資料集透過α參數範圍從0.00至1.00的調整，模擬了不同誤報率（False Alarm Rate， FAR）與漏報率（Missed Alarm Rate， MAR）設定下的多樣化故障情境，為模型評估提供了完整的實驗場域。 在方法論上，儘管循環神經網路（RNN）及其變體（LSTM，GRU）在過往研究中展現了優異的分類能力，但其「序列運算」的本質導致模型需依賴內部狀態遞迴傳遞，且對前處理階段的滑動視窗（Sliding Window）長度調參具有高度依賴，因此本研究提出以時間卷積網路（Temporal Convolutional Network， TCN）作為新一代的解決方案。 TCN 透過因果卷積與膨脹機制，不僅具備理論上的平行化運算優勢，亦能透過可設計之感受野捕捉長距離時序依賴，本研究主要聚焦於驗證其在嚴格盲測設定下之分類可行性與實務極限，並證實 TCN 於核電廠早期預警系統中之應用潛力。;With the growing global demand for low-carbon energy， nuclear power remains an essential source of baseload electricity. However， during abnormal events in nuclear power plant control rooms， a large number of alarm signals may occur simultaneously， resulting in an “alarm flood” that can seriously impair operators’ decision-making. Therefore， developing a highly real-time and accurate Prognostics and Health Management (PHM) system is crucial for ensuring nuclear safety. This study aims to address the efficiency bottlenecks of traditional sequential models when processing large-scale alarm data. At the data level， this research adopts the open IEEE PCTRAN dataset as the benchmark. By adjusting the α parameter from 0.00 to 1.00， the dataset simulates diverse fault scenarios under different false alarm rate (FAR) and missed alarm rate (MAR) settings， providing a comprehensive experimental environment for model evaluation. In terms of methodology， although recurrent neural networks (RNNs) and their variants， such as LSTM and GRU， have demonstrated strong classification performance in previous studies， their inherently sequential computation requires the recursive propagation of internal states. In addition， these models are highly dependent on the tuning of sliding window lengths during preprocessing. To overcome these limitations， this study proposes the Temporal Convolutional Network (TCN) as a next-generation solution. Through causal convolution and dilation mechanisms， TCN not only offers theoretical advantages in parallel computation but also captures long-range temporal dependencies through a configurable receptive field. This study mainly focuses on validating the classification feasibility and practical limits of TCN under a strict blind testing setting. The results further demonstrate the potential of TCN for application in early warning systems for nuclear power plants.
