- 學位論文
基於屬性空間擴展以發展生成式零樣本學習模型於未知故障的診斷
Developing Generative-based Zero-shot Learning Models Using Attribute Space Augmentation for Unknown Fault Diagnosis
摘要
許多化學工業程序的運作已經行之有年，在這期間工程師也積累了許多經驗，但還是有許多的操作問題沒有被診斷出來，過往曾發生過的故障可以被使用歷史數據所建立的模型診斷出來，但對於某些特定的故障數據卻可能只有稀少的數據量或沒有數據，因此如何及時對未發生的故障數據建立有效的故障診斷模型成為一大挑戰。本研究分別發展了兩種故障診斷模型，一種是利用充足的已知故障數據及零樣本的未知故障數據進行訓練，另一種則是在少樣本的已知故障數據及零樣本的未知故障數據進行訓練。 過往的零樣本學習提出了使用屬性代替標籤，透過物理意義所建立出的屬性可以描述故障發生的緣故，利用這個方式可以由已知故障推斷出未知故障的屬性。然而過往基於屬性建立的生成式的零樣本學習還存在一些問題。過去在訓練生成模型時，使用了已知故障的特徵與屬性進行訓練，訓練出的生成模型能夠生成已知故障數據，但不能保證生成模型生成正確的未知故障數據，也無法將未知故障數據的空間區隔開，這可能導致生成的未知故障數據與已知故障數據重疊。為了解決這個問題，本研究提出了透過三個階段的訓練來對未知故障數據空間進行拓展，第一個階段只使用了已知故障數據與屬性，建立出已知故障數據的空間，第二個階段會基於在第一個階段建立的已知故障數據空間，重新訓練模型來對未知故障數據的空間進行拓展，並且將已知故障與未知故障數據空間分開，第三個階段則是利用生成出的未知故障數據與真實的已知故障數據重新訓練原空間與特徵空間的關係，以便對未來收到的故障數據進行分類。 為了確保操作程序的安全性和正確性，在起始階段能精準地診斷故障至關重要，然而在生產與製造程序中，想要完整地收集故障數據是不可能的，而由於已知故障數據的缺乏，將難以訓練診斷模型對未知故障數據進行診斷。已知故障數據的缺乏會造成生成的未知故障數據不夠準確，過往方法運用了元學習的架構去彌補少樣本的影響，但也因此遇到元學習架構所帶來計算量過大之問題。為了解決此問題，本研究第二個主題會對故障診斷模型中的元學習進行改進，(1)首先會對元學習的架構進行改善，透過共同與個別模型來處理少樣本的問題，並同時減輕計算量；共同模型能將資訊共享給其他故障數據，個別模型則是能更精準的描述每種故障數據的空間。(2)如同第一個主題，同樣會分為三個階段對元學習架構下的未知故障數據的空間進行拓展，第一個階段建立出已知故障的共同與個別模型後，第二個階段會沿用共同模型並加入未知故障數據對應的個別模型來進行訓練，並於第三階段使用生成的未知故障數據及真實的已知故障數據重新訓練原空間與特徵空間的關係，以便對未來收到的故障數據進行分類。 最後，為了展現本研究的有效性及優點，兩個主題皆會利用數值例子和化學反應程序進行驗證對沒有故障數據的故障診斷，並與現有方法比較，展現出本研究發展出故障診斷模型的優勢。
關鍵字
零樣本並列摘要
Many chemical processes have been operating for years. Even though engineers have gained much experience, many operational problems and inefficiencies remain undiagnosed. Various improper operations that would cause the abnormalities of the operated processes can be identified from the historical data, but the collected data with the specified faults often do not exist or are sparse. How to promptly establish an effective fault diagnosis model for unseen faults in the operating process is a practical and challenging issue. The difficulty lies in the fact that very few or even no specified faults are available for reliable training of the fault diagnosis model in a process. Thus, two topics for training fault diagnosis models are separately developed. One is rich seen and zero unseen fault data; the other is limited seen and zero unseen fault data. In the past zero-shot learning introduced the use of fault attributes instead of fault labels. The attributes based on physical domain knowledge can fully describe the causes of fault occurrences. This approach can infer possible attributes for unseen faults from the attributes of seen faults. However, based on the defined attributes, several generative-based zero-shot learning methods have been proposed, but some issues are still not solved. The generative model trained with features and attributes of seen faults is used to generate unseen fault data. The generative model can generate seen fault data, but it cannot guarantee that generated unseen fault data are correct as the seen and unseen fault data are distributed in two different spaces. This certainly will cause the generated unseen fault data and seen fault data to overlap. To overcome this, a three-phase training procedure is proposed based on the defined attributes to expand the space for unseen fault data. In the first phase, the seen fault data and attributes are used to establish the seen faults in the feature and the original data spaces. The second phase, based on the seen spaces established in the first phase, retrains the established models to expand the spaces of unseen fault data but the spaces of seen and unseen fault data are independent. With the generated models, the third phase uses the generated unseen and real seen fault data to retrain the models in the original data space and the feature space to facilitate the classification of future collected fault data. To ensure the safety and formalities of the operating process, it is crucial to accurately diagnose the process faults at an early stage. However, it is impossible to fully and completely collect deficiency and improper operation in the production and manufacturing process. Due to the scarce samples of seen faults, it is difficult to train a diagnosis model for unseen faults. The lack of seen fault data can lead to inaccuracies in generating unseen fault data. Past work has proposed the framework of meta-learning to compensate for data scarcity, but this faces computational complexity. To address this issue, the second topic of this work improves the meta-learning used in the fault diagnosis with unseen fault data. Two improvements are proposed. (1) Enhancements to the meta-learning framework are proposed by jointly using common and specified models to address data scarcity. The common model shares information between different types of fault data, while the specified models provide more accurate descriptions for each type of fault data. (2) Expanding meta-learning, like the first topic, the research is divided into three phases to expand the space of unseen fault data. The first phase establishes common and specified models for the seen fault data. The second phase extends the established models with the seen faults to specified models corresponding to the unseen fault data, further training the common model. The third phase leverages both generated unseen and real seen fault data to retrain the relationship between the original data space and the feature space. The effectiveness and merits of the proposed method for two topics are verified by a numerical case and a chemical reactor process, showing significant advantages over the existing methods.