摘要 本研究以現地型地震預警（On-site EEW）為研究背景，使用中央氣象署申購資料及公開下載之桃園地區地震加速度資料，共整理722筆地震紀錄。首先以短時間平均與長時間平均比值法進行P波自動撿拾，並透過調整P波自動撿拾流程，依序採用原始門檻、統計值門檻及補充判定條件，使自動撿拾結果更接近人工判定時間，作為後續地震訊號特徵擷取之基準。 本研究擷取P波抵達後3秒內之加速度訊號，分別建立統計特徵集、地震工程特徵集、混合特徵集及改良特徵集。其後利用隨機森林回歸模型預測最大地動加速度（Peak Ground Acceleration， PGA），並依預測PGA轉換為地震震度。模型訓練採用固定資料切分、分層K折交叉驗證及窮舉搜尋，以比較不同特徵組合與模型超參數對平均絕對誤差、平均絕對百分比誤差及震度預測正確率之影響。 研究結果顯示，地震訊號特徵的選擇對模型預測結果具有明顯影響。相較於單獨採用波形統計特徵或地震工程參數，結合兩類資訊的混合特徵集具有較佳的震度預測表現。納入本研究依加速度訊號頻率特性所建立之週期特徵後，部分混合特徵組合之震度預測正確率略有提升。另一方面，其他週期特徵與對數轉換方式並未在各類特徵組合中形成一致改善，顯示新增特徵的有效性仍與原有特徵組成及資料分布有關。 本研究透過多階段判定條件改善P波自動撿拾流程，提高地震資料之P波撿拾成功率，增加可供後續分析之資料筆數，使自動撿拾之P波抵達時間更接近人工判定結果，並有助於隨機森林模型進行PGA與震度預測。本研究成果可作為現地型地震預警系統中P波自動撿拾、特徵選擇及模型設定之參考。;Abstract This study investigates an on-site earthquake early warning (On-site EEW) framework using earthquake acceleration records from the Taoyuan region， obtained from purchased datasets and publicly available data provided by the Central Weather Administration. A total of 722 earthquake records were compiled. P-wave arrivals were first identified automatically using the short-term average to long-term average ratio method. The P-wave picking procedure was then refined by sequentially applying the original threshold， statistical thresholds， and supplementary criteria， so that the automatically determined P-wave arrival times more closely matched the manually identified results and could serve as the reference points for subsequent seismic signal feature extraction. Acceleration signals within 3 s after the P-wave arrival were extracted to construct statistical feature sets， earthquake engineering feature sets， combined feature sets， and modified feature sets. A random forest regression model was then employed to predict peak ground acceleration (PGA)， which was subsequently converted into seismic intensity. Model training was conducted using a fixed data split， stratified K-fold cross-validation， and grid search. The effects of different feature combinations and model hyperparameters were evaluated in terms of mean absolute error， mean absolute percentage error， and seismic intensity prediction accuracy. The results indicate that the selection of seismic signal features has a clear influence on model performance. Compared with the use of waveform statistical features or earthquake engineering parameters alone， the combined feature sets yielded better seismic intensity prediction performance. After incorporating a period-related feature derived in this study from the frequency characteristics of acceleration signals， the seismic intensity prediction accuracy slightly increased. In contrast， other period-related features and logarithmic transformations did not produce consistent improvements across all feature combinations， indicating that the effectiveness of newly introduced features remains dependent on the composition of the original feature set and the distribution of the data. Through the use of multiple-stage decision criteria， this study improved the automatic P-wave picking procedure， increased the P-wave picking success rate， expanded the number of earthquake records available for subsequent analysis， and reduced the discrepancy between automatically and manually identified P-wave arrival times. These improvements support the subsequent application of random forest models for PGA and seismic intensity prediction. The findings of this study may serve as a reference for P-wave automatic picking， feature selection， and model configuration in on-site earthquake early warning systems.
