工業物聯網之時間序列資料管理與詢問式異常偵測

Abstract

工業4.0推動製造業邁向數位化與智慧化，藉由虛實整合系統（Cyber-Physical Systems, CPS）、物聯網（Internet of Things, IoT）、人工智慧（AI）與邊緣計算（Edge Computing）等技術的整合，建立具即時感知與自主決策能力的智慧工廠。隨著工廠中感測裝置的大量部署，其所產生之大量高頻時間序列資料，成為設備狀態監控、品質管理、異常偵測與預測性維護的重要依據。然而，工業用戶對於將敏感資料上傳至公有雲進行分析存有疑慮，亦難以因應工業用戶對自主部署與彈性調整之實務需求。為解決上述問題，本研究基於機器學習技術，設計一套可部署於企業內部之工業物聯網系統。本研究包含以下三項貢獻：一、建構支援時間序列資料存取及異常偵測之工業物聯網平臺；二、提出結合詢問式學習之條件式變分自動編碼器的異常偵測方法，透過針對較具學習意義的高度不確定性樣本微調模型，標記數據需求量減少54.50%，微調時間縮短53.74%，同時模型效能平均提升1.43%；三、支援使用壓縮資料直接進行異常分析，免除資料還原步驟，提升傳輸效率並降低系統儲存與運算的負擔。

Industry 4.0 is propelling manufacturing toward digitalization and intelligence by integrating Cyber‑Physical Systems (CPS), the Internet of Things (IoT), Artificial Intelligence (AI), and edge computing, thus enabling smart factories with real‑time perception and autonomous decision‑making. The extensive deployment of sensors generates massive high‑frequency time‑series data that underpin equipment condition monitoring, quality control, anomaly detection, and predictive maintenance. Nevertheless, manufacturers hesitate to upload sensitive data to public clouds for analysis and struggle to satisfy practical needs for self‑hosted, easily reconfigurable deployments. To address these challenges, this study designs an on-premises Industrial IoT (IIoT) platform based on machine learning techniques. This study makes three contributions: (1) Establishing an IIoT platform that supports time series data access and anomaly detection; (2) Proposing an anomaly detection method that combines Query-based Learning with a Conditional Variational Autoencoder. By fine-tuning the model using highly uncertain samples that are more informative for learning, the approach reduces the required amount of labeled data by 54.50%, shortens fine-tuning time by 53.74%, and improves model performance by an average of 1.43%; and (3) Supporting anomaly analysis directly on compressed data, eliminating the need for decompression and thereby improving transmission efficiency and reducing system storage and computational overhead.