- 學位論文
發展深度模型的圖像/影像監控系統於工業上的應用
Development of Deep Monitoring Models for Industrial Image/Video Data
摘要
在化學工業中，生產需維持安全和品質的穩定，許多多變量統計方法被用來監控程序異常。隨著儀器與硬體的進步，許多基於圖像的深度學習方法被提出。對於操作程序而言，傳統的多變量量測僅為局部的信息，而圖像量測可包含較全局的信息。本研究在利用深度學習為基礎的模型，發展工業程序圖像的自動監控系統。本研究的監控系統分成兩個部分，靜態圖像與動態圖像的監控系統。在靜態圖像的監控系統由於高維度的圖像，利用卷積的變分自動編碼神經網路(Convolutional Neural Network - Variational AutoEncoder, CNN-VAE)的編碼器(Encoder)用以抽取潛變量(Latent Variable)及去編碼器(Decoder)用以重構圖像 (Reconstruction)。當模型學習完成後，再由核密度函數的映射以建立高斯分布的潛變量及預測誤差於潛變量和殘差空間的管制指標，並用火焰燃燒圖像驗證CNN-VAE的監控效能。在動態圖像的監控系統中，由於大型的化學工業程序呈現緩慢的動態行為，當異常操作發生時，並不會立即影響程序，但可逐漸地體現於圖像量測的幾何分布變化上。為了在異常徵兆發生的初期，即可檢測出程序異常，本研究延伸靜態的CNN-VAE方法，發展出CNN-KRVAE(Kalman Recurrent Variational AutoEncoder)。CNN-KRVAE的編碼器和去編碼器結構，分別由非線性的馬可夫暫態分佈(Transition Distribution)與擴散分布(Emission Distribution)來描述模型。當模型學習完成後，同樣地由核密度函數的方式對時序的潛變量和殘差發展T2和SPE管制指標。最後，用非侵入式的紅外線熱像儀紀錄熱交換反應器的溫度分布影像，以驗證CNN-KRVAE的即時監控效能。
並列摘要
In industries, multivariable statistical analysis is often used to monitor operating processes to maintain the safety and quality stability of the processes. With the advancement of instruments and hardware, many image-based deep learning algorithms have been developed. Traditionally measured variables can only represent local information of process units, but a single image can easily capture the whole information of the operating units. This research aims to develop an image-based deep network model for the automatic monitoring of industrial processes. Based on the image sampling frequency, two different types of image-based deep learning models are proposed. One is for monitoring static image data; the other, for monitoring dynamic video data. In the static image data, the convolutional neural network-variational autoencoder (CNN-VAE) is used. The encoder is used to extract the latent variables (LVs) and the decoder is used to reconstruct the image. Once the model is well trained, the non-Gaussian control indices are calculated by the kernel density estimation (KDE) in the latent and residual spaces, respectively. The case study of flame combustion is used to demonstrate the effective monitoring of the proposed CNN-VAE. In the dynamic video data, the static CNN-VAE method is extended to dynamic Kalman recurrent variational autoencoder (CNN-KRVAE). Because of the slow dynamic behaviors in large-scale chemical plants, the dynamic process changes transitionally, not abruptly, when the abnormal operation (fault) occurs. It is certain that the geometric distribution of image data gradually changes. To detect the fault at the early stage of operation, the encoder and de-encoder of CNN-KRVAE are constructed by using the nonlinear Markov transient distribution and emission distribution respectively. Once the CNN-KRVAE model is well trained, the and control limits of time-series image data in the latent and residual spaces are defined by KDE. The non-invasive measurement through infrared thermal image experiments is used in the heat exchanger reactor to show the proposed CNN-KRVAE can effectively monitor the temperature distribution of the recorded video in real time.