發展由單步至多步的動態概率品質預測模型

Developing Dynamic, Probabilistic, Quality Predictive Models: from One-Step Prediction to Multi-Step Predictions

黃世傑(SAI KIT OOI)

指導教授 ： 陳榮煇

中原大學/工學院/化學工程學系/碩士(2022年)

https://doi.org/10.6840/cycu202201317

摘要

在競爭激烈的市場中，爲了保證產品的品質，工廠的品質監控是迫切且必須的。大型化工廠具有複雜的特性如：非線性、慢動態和不確定性的特點。過去，已經開發了很多品質監控模型，但它們只提供當下時刻品質質量。由於化工程序規模較大，過程的內在動態特性非常明顯。當故障發生時，它不會瞬間影響過程，但會在幾個時間點後做出反應。因此，無法通過測量或預測當前時間的品質來識干擾的影響。待產品生產完畢後才發現次產品時，再修復流程已為時已晚。品質監控的核心目的是希望得以儘早發現當前運行過程的任何問題。然而，傳統方法並不關心在任何干擾顯著影響過程之前的早期檢測。此外，為了使產品的品質更加一致和穩定，預測模型從中扮演者非常重要的角色，為反饋控制提供潛在問題的預警，以達到更好的控制性能。然而，在化工程序中，標記數據的獲取通常既昂貴又耗時，通常只能在透過實驗室分析才得以獲取品質變量。因此，品質數據的稀缺導致難以建立可靠的預測模型。針對品質稀缺問題，本文針對不同的情況分別提出了兩種解決方案。

在第一個主題中，爲了滿足動態市場的需求，製造商傾向於生產多級產品。由於多級產品可能來自同一條生產線，它們可能具有相似的過程內在信息。因此，提出了一種基於遷移學習的半監督動態非線性，超前預測模型，稱為半監督潛在高斯混合非線性狀態空間模型（Semi-supervised Latent Gaussian Mixture Nonlinear State Space Model, S2-LGMNSSM）。通過利用來自源域的知識轉移至目標域以增强目標域的信息進而來提高模型的韌性和精度。為了處理丟失的品質數據，經過設計與訓練的插補網絡將用於估算丟失的數據，為在線預測期間提供品質估計。 另外，S2-LGMNSSM 可以從潛在空間中的源動態數據和目標動態數據中提取共同特徵；並將源域的特徵轉移到目標域。與普通的變分自編碼器不同，S2-LGMNSSM的先驗分佈設置為高斯混合模型而不是單位高斯分佈，以便區分目標和源數據的特殊動態特性。

在第二個主題中，基於歷史正常和異常數據，提出了一種預測監測品質的多步預測系統。多步預測系統由動態模型（Multi-step Nonlinear State Space Model, Multi-NSSM）和轉換模型（Regression Variational Autoencoder, Reg-VAE）組成。 Multi-NSSM是通過容易測量的變量所建立的，以學習操作變量、過程輸入變量和過程輸出變量之間的動態關係。並利用Multi-NSSM預測過程輸出變量的未來趨勢。 Reg-VAE則是由難以測量的主要變量建立，用於捕捉過程輸出變量與主要品質變量之間的關係；它可以通過轉換預測的過程輸出變量來推斷主要品質變量的未來趨勢。程序的內在信息可以透過將過去的輸入過程數據和輸出過程數據從觀察空間映射到潛在空間獲得並保留其動態信息用於預測品質的未來趨勢，從而實現預警。

最後，數值例子與工業例子將會用於驗證S2-LGMNSSM 和多步預測系統 (Multi-NSSM & Reg-VAE) 的有效性。

關鍵字

單步 ； 多步 ； 動態 ； 概率 ； 品質 ； 預測模型

並列摘要

In the competitive market, monitoring quality in plants can ensure the quality of products, but the complexities of the large-scaled chemical plant are characterized by strong nonlinearities, slow dynamics, and uncertainties. In the past, a lot of quality monitoring models have been developed, but they only provide the current status of quality. As the scale chemical process is large, the intrinsic dynamic properties of the process are very obvious. When the fault occurs, it will not influence the process instantaneously, but it will react after a few time points. The impact of disturbances cannot be identified by measuring or predicting the quality at the current time. After all the products are inspected, it is too late to fix the process. The purpose of quality monitoring is to detect any problem with the currently running process as early as possible. However, conventional approaches do not care about early detection before any disturbance significantly affects the process. In addition, to make the quality of the product more consistent and stable, the predictive model is highly demanded to provide the early warning of potential problems for feedback control to achieve better control performance. Nevertheless, labeled data is often expensive and time-consuming to acquire as the quality variables are usually analyzed in the laboratory. Therefore, there is often insufficient to establish a reliable predictive model. To solve the quality scarcity problem, two solutions to different situations are proposed in this thesis, respectively.
In the first topic, as manufacturers tend to produce products with multiple grades to meet the dynamic and time-to-market demands, the multi-grade products may come from the same production line. Thus, they may have similar intrinsic information of the process. Therefore, a novel transfer learning-based semi-supervised dynamic nonlinear soft sensor with a step-ahead prediction model, which is called a semi-supervised latent Gaussian mixture nonlinear state-space model (S2-LGMNSSM), is proposed to enlarge the information of the target domain by leveraging the knowledge from the source domain to enhance the model robustness and precision. To handle the missing quality data, the trained imputation network would be used to impute missing data to provide quality estimates during on-line prediction. S2-LGMNSSM can extract the common features from both source and target dynamic data in the latent space; then the features from the source grade can be transferred to the target grade. Unlike an ordinary variational autoencoder, a Gaussian mixture model instead of a unit Gaussian distribution is set as the prior distribution of LGMNSSM so that the special dynamic characteristics of the target and source data can be distinguished.
In the second topic, based on the historical normal and abnormal data, a multi-step prediction system for predictive monitoring quality is proposed. The multi-step prediction system consists of a dynamic model (Multi-NSSM, which is short for Multi-step Nonlinear State Space Model), and a conversion model, (Reg-VAE, which is short for Regression Variational Autoencoder). Multi-NSSM is established by easily measured secondary variables to learn the dynamic relationship among manipulated, process input variables and process output variables as well as forecast the process output variables. Reg-VAE is established by hard-to-measure primary variables for capturing the relationships between process output variables and primary variables; it can infer the future trend of primary variables by converting the forecasted process output variables. The past input and output process data can be mapped from the observation space into the latent space to acquire the intrinsic properties. They preserve the dynamic information for the future multi-step prediction of quality so that early warning can be achieved.
The effectiveness of S2-LGMNSSM and multi-step prediction system (Multi-NSSM & Reg-VAE) in future prediction are presented in a numerical case and an industrial case.

並列關鍵字

Dynamic ； Probabilistic ； Quality Predictive Models ； One-Step Prediction ； Multi-Step Predictions

參考文獻

[1] W. F. Massy, “Principal components regression in exploratory statistical research”, Journal of the American Statistical Association. 60 (1965) 234–256. https://doi.org/10.1080/01621459.1965.10480787.

[2] P. Geladi, B. R. Kowalski, “Partial least-square regression: a tutorial”, Elsevier Science Publishers B.V. (1986)

[3] R. Rosipal, “Kernel partial least squares regression in reproducing kernel hilbert space”, Journal of Machine Learning Research. (2001) 97-123.

[4] X. Lu, Y. Tsao, S. Matsuda, C. Hori, “Speech enhancement based on deep denoising autoencoder”, Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH. (2013) 436–440.

[5] K. Wang, M. G. Forbes, B. Gopaluni, J. Chen, Z. Song, “Systematic development of a new variational autoencoder model based on uncertain data for monitoring nonlinear processes”, IEEE Access. 7 (2019) 22554–22565. https://doi.org/10.1109/ACCESS.2019.2894764.