本研究使用 PCTRAN 模擬產生之核電廠警報時間序列資料，資料內容包含正常狀 態與多種異常事故類別，並涵蓋不同 alpha 警報敏感度設定。研究過程中，首先進行資料整理、正規化與滑動視窗樣本建構，將原始時間序列資料轉換為固定長度之模型輸入，再以 Time Series Transformer 進行多類別分類。透過 Transformer 之自注意力機制，模型可學習不同時間點警報訊號之間的關聯性，進而辨識各類異常事件之警報變化特徵。 實驗結果顯示，不同滑動視窗長度與 alpha 訓練方式皆會影響模型分類表現。當滑 動視窗長度為 240 時，模型可取得較高的整體分類準確率；而視窗長度為 120 時，則在 Macro F1-score 上具有較佳表現，顯示其對各類別之平均辨識能力較為均衡。此外，個別 alpha 訓練模型整體表現較為穩定，適合用於已知警報敏感度條件下之分類任務；混合 alpha 訓練模型則具備跨不同 alpha 條件之應用彈性，但在高 alpha 區間仍有改善空間。;This study uses PCTRAN-simulated nuclear power plant alarm time-series data， including normal operating conditions and multiple abnormal accident scenarios， under different alpha alarm sensitivity settings. In the research process， data organization， standardization， and sliding window sample construction are first conducted to transform the original time-series data into fixed-length model inputs. The Time Series Transformer is then applied for multi-class classification. Through the self-attention mechanism of Transformer， the model can learn the relationships among alarm signals at different time points and further identify the alarm variation patterns of different abnormal events. The experimental results show that both sliding window length and alpha training strategy affect model classification performance. When the sliding window length is set to 240， the model achieves higher overall classification accuracy; when the window length is set to 120， the model performs better in terms of Macro F1-score， indicating more balanced average recognition ability across different classes. In addition， the individual-alpha training model shows more stable overall performance and is suitable for classification tasks under known alarm sensitivity conditions. In contrast， the mixed-alpha training model provides greater flexibility across different alpha conditions， but its performance still requires improvement in the high-alpha range.
