使用主動式數據選擇改進潛變數的軟測量模型

Improving Quality of Latent Variable Soft Sensor Models Using Active Sequential Data Selection

吳慶洋(Qing-Yang Wu)

指導教授 ： 陳榮輝

中原大學/工學院/化學工程研究所/碩士(2016年)

https://doi.org/10.6840/cycu201600901

摘要

軟測量(Soft sensor)為一個現代工業替代實體感測器的技術，軟測量的概念為使用系統的知識和可用的測量，以便改進的感興趣的變量的測量值的可靠性，精度和成本。數據驅動的軟測量方法是仰賴於數據品質的好壞，若數據選取的不好勢必會使得軟測量的預測效果是糟糕的。然而大部分數據驅動的軟測量方法都是假設數據是良好的狀況，鮮少考慮數據品質的問題。因此本研究的目標是找出軟測量模型哪些區域不精確，並且將新的數據點放入模型以至於預測改善。但是用人工的方式去選擇數據是沒有效率的，所以此研究利用一個主動式的架構進行改善，使得軟測量模型能自行選擇需要的數據。在此架構中包含了幾個元素：訓練數據集、軟測量模型、數據選擇池、主動的協助員。主動的協助員將會協助軟測量模型來選擇數據，而整個主動式循環需要一個選擇數據的準則，以及一個學習收斂的準則。此研究以潛變數模型作為的軟測量技術，高斯過程作為主動式的輔助模型。利用潛變數做為高斯模型的輸出，得出潛變數模型的不確定性。藉由不確定性以及預測值的梯度訊息當作選擇數據的準則。此主動式學習過程將分成靜態與動態程序分別進行討論，並且實際使用中油的數據進行主動式的軟測量建模。

關鍵字

高斯過程模型,主動式學習,軟測量,數據選擇

並列摘要

Soft Sensor is an alternative to modern industrial hardware sensor technology. In soft sensor applications, the modeling data are collected from the historical data in an existing process, and the selected models are built upon the collected data. However, one never knows in advance if training data provide sufficient information. This often causes soft sensor models to inevitably have some unreliable predictions in some regions in their industrial applications. This study aims to develop a sequentially active learning method which selects a set of significant data to enhance soft sensor models. The selection of the significant and representation data from the samples is a crucial task in the active learning method. In this study, the latent variable models (LVMs) are adopted to train the soft sensors because it is widely reported in literature that highly correlated variables may lead to numerical problems during the modeling step.

The objective of this work is to proposed LVMs with an auxiliary Gaussian process model for an uncertainty selection criterion that takes the process and quality variables into account. An uncertainty index of LVM is presented. It contains the variances of the predicted outputs and the changes of the predicted outputs per unit change in the designed inputs. Under the active learning framework, two different soft sensor models, static and dynamic models, are developed respectively. In static models, the process is in static states with a small portion of process data adopted and the quality variables are dependent on the process variables and sampled at the particular time interval. Without any prior knowledge of the processes, the uncertainty index is sequentially utilized to find out from which regions the new data should be collected to enhance the model quality. However, the static assumption of the soft sensors is incapable of handling the dynamic of the processes, because in an actual process the quality variables are in transition and modeling the quality variables in using a steady state model can result in an unsatisfactory performance. In dynamic models, extension to formulate the dynamic soft sensors is also proposed. This is accomplished by taking into account the quality variables related to a number of the past process variable data from the historical data, and the resultant soft sensor is able to extract dynamic information for better performance when dealing with dynamic processes. Based on the future predictions, the time to update the model can be carried out without waiting for the process to measure the quality variables. The uncertainty index for the online quality variable is continuously checked and updated as it is necessary to maintain the performance of the dynamic soft sensor.

The proposed methods can be applied to any types of LVMs. The effectiveness and promising results of the static and the dynamic soft sensors are respectively demonstrated using two numerical examples and an industrial process with multiple outputs in Taiwan.

並列關鍵字

Gaussian process regression model, Active learning, Soft sensor, Data selection

參考文獻

[1] Aguado, D., Ferrer, A., Seco, A., & Ferrer, J. Comparison of different predictive models for nutrient estimation in a sequencing batch reactor for wastewater treatment. Chemometrics and intelligent laboratory systems, 2006, 84(1), 75-81.

[2] Bosca, S., Barresi, A. A., & Fissore, D. Use of a soft sensor for the fast estimation of dried cake resistance during a freeze-drying cycle. International journal of pharmaceutics, 2013, 451(1), 23-33.

[3] Ge, Z. Active learning strategy for smart soft sensor development under a small number of labeled data samples. Journal of process control, 2014, 24(9), 1454-1461.

[4] Ge, Z. Active probabilistic sample selection for intelligent soft sensing of industrial processes. Chemometrics and intelligent laboratory systems, 2016, 151, 181-189.

[5] Joseph, B., & Brosilow, C. B. Inferential control of processes: Part I. steady state analysis and design. American institute of chemical engineers journal, 1978, 24(3), 485–492.