本研究旨在結合生存分析與 SHAP 解釋技術，探討製造過程中不同產品製 程參數對機器異常風險的影響，並建立具可解釋性的預測模型。資料採用網路 上公開之預測型維護資料集以及 A 公司所提供之塗佈機資料集。研究首先比較 Cox 比例風險模型、隨機生存森林（Random Survival Forest, RSF）、XGBoost Cox 與 DeepSurv 等四種生存分析方法，並以 C-Index 與整合 Brier 分數 （IBS）評估模型效能，最終選擇表現最佳的隨機生存森林作為核心模型。 模型建構後，進一步應用 SHAP 與其時間依賴版本 SurvSHAP(t)，量化各 產品製程參數對異常發生風險的影響以及生存機率變化的影響與其隨時間的變 動趨勢，辨識出加速機器劣化的關鍵製程參數。在此基礎上，透過隨機生存森 林預測之生存機率門檻進行產品製程風險分類，對未來在生產製造維護策略上 可提供須優先維護觀察之產品，以降低產品在生產製造上發生異常的可能性。 結果顯示，本方法不僅能有效預測產品生產時的生存機率，亦能針對關鍵 製程參數與高風險時點提供具體解釋，協助工程師優化維護策略與製程設定。 相較於傳統僅預測是否故障的分類型預測方法，本研究所提出之方法可同時考 量風險隨時間變化的動態特性，並具備透明度與實務應用價值。;This study aims to integrate survival analysis with SHAP interpretability techniques to investigate how different product process parameters influence the risk of machine failure during manufacturing. The analysis utilizes both a publicly available predictive maintenance dataset and a coating machine dataset provided by Company A. Four survival analysis methods are compared—Cox Proportional Hazards Model, Random Survival Forest (RSF), XGBoost Cox, and DeepSurv— using the C-index and Integrated Brier Score (IBS) to evaluate model performance. The Random Survival Forest, which demonstrated the best performance, is selected as the core model. After model construction, SHAP and its time-dependent extension SurvSHAP(t) are applied to quantify the impact of each process parameter on failure risk and survival probability over time, identifying key factors that accelerate machine degradation. Based on these insights, process risk classification is performed using survival probability thresholds predicted by RSF. This allows for the identification of high-risk products that require prioritized monitoring and maintenance, thereby reducing the likelihood of anomalies during production. Results show that the proposed approach not only accurately predicts the survival probability of products during manufacturing but also provides concrete explanations for key parameters and high-risk time points. This supports engineers in optimizing maintenance strategies and process settings. Compared to traditional classification models that only predict whether failure will occur, the proposed method captures the dynamic nature of risk over time while offering greater transparency and practical value.
