在建築管理系統中，冰水主機為冰水系統的核心，亦是能耗最高且運作最為複雜的子系統。隨著智慧建築與節能技術的發展，冰水主機的故障診斷偵測（Fault Detection and Diagnostics， FDD）技術已成為提升能源效率與設備可靠性的關鍵環節。然而，實務上仍存在兩大挑戰亟待解決：其一為部分關鍵變量（如供油壓力等）受限於成本與實體感測器配置而難以直接量測，導致模型可觀測性不足；其二為工業場域中真實故障樣本取得困難，極端不平衡的數據分佈使得監督式學習模型易陷入多數類別的決策盲區，引發嚴重的初期故障漏報問題。 有鑒於此，本研究提出一套整合混合式虛擬感測器與條件式潛在擴散模型（Conditional Latent Diffusion Model， CLDM）的創新故障診斷架構（VS-CLDM）。運用基於壓縮機運行機理與曲線擬合技術建立虛擬供油壓力感測器，補足隱含的物理變化特徵，確立分類器的診斷錨點。利用 CLDM 擴散模型具備的馬可夫去噪（Markovian Denoising）與流形插值能力，在極端小樣本條件下生成高保真度的合成故障數據。此方法不僅能發揮非線性濾波作用以平滑化感測雜訊，更有效拓寬了少數故障類別的特徵決策邊界，從根本上解決了資料不平衡的問題。 實驗結果顯示，本研究提出之 VS-CLDM 架構於冰水主機的故障診斷偵測任務中展現了卓越的實效。透過主成分分析（Principal Component Analysis ， PCA）的特徵視覺化與多項量化指標評估，證實該架構能有效解耦複雜的物理特徵，並大幅提升各類別故障的分類準確度。特別是在極端數據失衡與初期微弱故障的嚴苛條件下，評估模型依然展現出極高的識別強健性與決策穩定度。綜合而言，本研究成功驗證了「以軟體運算替代硬體感測器」之可行性，證明該架構不僅能克服實體設備配置的成本限制，更在工業廠房的預測性維護中具備極高的現場實施與落地價值。;In building management systems (BMS)， chillers serve as the core of the chilled water sys-tem， representing the subsystem with the highest energy consumption and the most complex operational mechanisms. With the advancement of smart buildings and energy-saving technolo-gies， Fault Detection and Diagnosis (FDD) for chillers has become a crucial link in enhancing energy efficiency and equipment reliability. However， two major challenges persist in practical applications: first， certain key variables， such as oil feed pressure， are difficult to measure di-rectly due to cost constraints and physical sensor layouts， leading to insufficient model observa-bility; second， collecting real-world fault samples in industrial settings is extremely difficult， re-sulting in highly imbalanced data distributions that cause supervised learning models to fail within the decision blind spots of majority classes， thereby triggering severe underreporting of incipient faults. To address these pain points， this study proposes an innovative fault diagnosis framework that integrates a hybrid virtual sensor with a Conditional Latent Diffusion Model (VS-CLDM). First， a virtual oil feed pressure sensor is developed based on compressor operating principles and curve-fitting techniques， which supplements implicit physical variation characteristics and establishes a key diagnostic anchor for the classifier. Next， leveraging the Markovian denoising and manifold interpolation capabilities of the CLDM， high-fidelity synthetic fault data are gen-erated under extreme few-shot conditions. This methodology not only performs non-linear fil-tering to smooth sensor noise but also effectively widens the decision boundaries of minority fault classes， thereby fundamentally resolving the data imbalance issue. Experimental results indicate that the proposed VS-CLDM framework demonstrates out-standing effectiveness in chiller FDD tasks. Through feature visualization via Principal Com-ponent Analysis (PCA) and evaluation across multiple quantitative metrics， it is verified that the framework can effectively decouple complex physical features and substantially improve the classification accuracy of various fault types. Particularly under the stringent conditions of ex-treme data imbalance and incipient weak faults， the evaluation model maintains high identifica-tion robustness and decision stability. Conclusively， this study successfully validates the feasibil-ity of "replacing hardware sensors with software algorithms，" proving that the architecture not only overcomes the cost limitations of physical sensor deployment but also possesses highly practical on-site implementation value for predictive maintenance in industrial facilities.
