本研究明確呼應國家「智慧機械」產業創新政策，針對目前製造業面臨之數位轉型需求，以 PET 二步法吹瓶機為研究標的，探討如何透過人工智慧（AI）與工業物聯網（IIoT）技術，解決傳統產線高度依賴人工肉眼檢測且缺乏量化數據庫的痛點。本研究開發一套「生產數據收集全端資通訊服務系統」，自動接取吹瓶機之製程參數、感測數據與視覺單元之瓶底影像，並基於此數據庫建立兩階段之 AI 瓶心辨識與瓶底偏心預測模型。 在第一階段，本研究提出基於 YOLOv8s 深度學習架構之「瓶心辨識模型」，能精確定位瓶底影像中之瓶胚中心與模具中心，其測試集效能表現優異，Precision 達 99.25%、Recall 達 95.45%、mAP50-95 達 81.97%。此模型為偏心距離的量化計算提供了可靠的座標基礎，並成功為後續預測模型生成標籤數據。 在第二階段，本研究提出「瓶底偏心預測模型」，比較 Logistic Regression、Random Forest、XGBoost 及 LightGBM 等機器學習演算法之效能，並進一步導入結合注意力機制之多層感知機（MLP with Attention）進行延伸評估。實驗結果顯示，當偏心閥值設定為 20px 時，模型之各項預測指標表現最為穩定均衡。此外，特徵重要性分析與 Attention 權重交叉比對指出，「加熱箱溫度」、「高壓吹氣位置」及「瓶胚溫度」為影響瓶底偏心之前三大關鍵因子，且「環境溫度」對於捕捉微小變異亦具高度重要性，此結果與實務製程經驗法則高度契合。本研究成果不僅證實 AI 模型能有效取代人工進行自動化監測，其量化評估結果更為吹瓶機未來邁向智慧化自適應調參（Self-adaptive）奠定了重要的實務基礎。;This research responds to the national "Smart Machinery" industrial innovation policy. Addressing the digital transformation needs of the manufacturing industry， this study focuses on the PET two-stage blow molding machine to explore how Artificial Intelligence (AI) and the Industrial Internet of Things (IIoT) can resolve the pain points of traditional production lines， which heavily rely on manual visual inspection and lack quantitative databases. This study developed a full-stack ICT service system for production data collection， automatically acquiring process parameters， sensor data， and bottle base images. Based on this database， a two-stage AI eccentricity identification and prediction model was established. In the first stage， an "Bottle Base Eccentricity Identification Model" based on the YOLOv8s deep learning architecture was proposed. It accurately localizes the preform center and mold center in bottle base images. The testing results demonstrated excellent performance， with a Precision of 99.25%， Recall of 95.45%， and an mAP50-95 of 81.97%. This model provides a reliable coordinate foundation for quantifying eccentricity distance and successfully generates ground-truth labels for the subsequent prediction model. In the second stage， a "Bottle Base Eccentricity Prediction Model" was proposed， comparing machine learning algorithms including Logistic Regression， Random Forest， XGBoost， and LightGBM， while further introducing a Multilayer Perceptron with Attention mechanism (MLP with Attention) for an extended evaluation. Experimental results indicated that setting the eccentricity threshold to 20px yielded the most stable and balanced predictive performance. Furthermore， feature importance analysis and Attention weight cross-validation revealed that "oven temperature，" "high-pressure blowing position，" and "preform temperature" are the top three critical factors affecting bottle base eccentricity， with "ambient temperature" also showing high importance in capturing subtle process variations. These findings highly correlate with practical manufacturing experience. The results of this study not only confirm that AI models can effectively replace manual monitoring， but its quantitative evaluation also establishes a solid practical foundation for the future implementation of self-adaptive tuning in blow molding machines.
