| 1. |
- Capaci, Francesca, et al.
(författare)
-
A two-step procedure for fault detection in the Tennessee Eastman Process simulator
- 2016
-
Konferensbidrag (refereegranskat)abstract
- High-technological and complex production processes and high availability and sample frequencies of data in large scale industrial processes need the concurrent development of appropriate statistical control tools and monitoring techniques. Therefore, multivariate control charts based on latent variables are essential tools to detect and isolate process faults.Several Statistical Process Control (SPC) charts have been developed for multivariate and megavariate data, such as the Hotelling T2, MCUSUM and MEWMA control charts as well as charts based on principal component analysis (PCA) and dynamic PCA (DPCA). The ability of SPC procedures based on PCA (Kourti, MacGregor 1995) or DPCA (Ku et al. 1995) to detect and isolate process disturbances for a large number of highly correlated (and time-dependent in the case of DPCA) variables has been demonstrated in the literature. However, we argue that the fault isolation capability and the fault detection rate for processes can be improved further for processes operating under feedback control loops (in closed loop).The purpose of this presentation is to illustrate a two-step method where [1] the variables are pre-classified prior to the analysis and [2] the monitoring scheme based on latent variables is implemented. Step 1 involves a structured qualitative classification of the variables to guide the choice of which variables to monitor in Step 2. We argue that the proposed method will be useful for many practitioners of SPC based on latent variables techniques in processes operating in closed loop. It will allow clearer fault isolation and detection and an easier implementation of corrective actions. A case study based on the data available from the Tennessee Eastman Process simulator under feedback control loops (Matlab) will be presented. The results from the proposed method are compared with currently available methods through simulations in R statistics software.
|
|
| 2. |
- Capaci, Francesca, et al.
(författare)
-
Simulating Experiments in Closed-Loop Control Systems
- 2016
-
Ingår i: ENBIS-16 in Sheffield.
-
Konferensbidrag (refereegranskat)abstract
- Design of Experiments (DoE) literature extensively discusses how to properly plan, conduct and analyze experiments for process and product improvement. However, it is typically assumed that the experiments are run on processes operating in open-loop: the changes in experimental factors are directly visible in process responses and are not hidden by (automatic) feedback control. Under this assumption, DoE methods have been successfully applied in process industries such as chemical, pharmaceutical and biological industries.However, the increasing instrumentation, automation and interconnectedness are changing how the processes are run. Processes often involve engineering process control as in the case of closed-loop systems. The closed-loop environment adds complexity to experimentation and analysis since the experimenter must account for the control actions that may aim to keep a response variable at its set-point value. The common approach to experimental design and analysis will likely need adjustments in the presence of closed-loop controls. Careful consideration is for instance needed when the experimental factors are chosen. Moreover, the impact of the experimental factors may not be directly visible as changes in the response variables (Hild, Sanders, & Cooper, 2001). Instead other variables may need to be used as proxies for the intended response variable(s).The purpose of this presentation is to illustrate how experiments in closed-loop system can be planned and analyzed. A case study based on the Tennessee Eastman Process simulator run with a decentralized feedback control strategy (Matlab) (Lawrence Ricker, 1996) is discussed and presented.
|
|
| 3. |
- Vanhatalo, Erik, et al.
(författare)
-
Lag Structure in Dynamic Principal Component Analysis
- 2016
-
Konferensbidrag (refereegranskat)abstract
- Purpose of this PresentationAutomatic data collection schemes and abundant availability of multivariate data increase the need for latent variable methods in statistical process control (SPC) such as SPC based on principal component analysis (PCA). However, process dynamics combined with high-frequency sampling will often cause successive observations to be autocorrelated which can have a negative impact on PCA-based SPC, see Vanhatalo and Kulahci (2015).Dynamic PCA (DPCA) proposed by Ku et al. (1995) has been suggested as the remedy ‘converting’ dynamic correlation into static correlation by adding the time-lagged variables into the original data before performing PCA. Hence an important issue in DPCA is deciding on the number of time-lagged variables to add in augmenting the data matrix; addressed by Ku et al. (1995) and Rato and Reis (2013). However, we argue that the available methods are rather complicated and lack intuitive appeal.The purpose of this presentation is to illustrate a new and simple method to determine the maximum number of lags to add in DPCA based on the structure in the original data. FindingsWe illustrate how the maximum number of lags can be determined from time-trends in the eigenvalues of the estimated lagged autocorrelation matrices of the original data. We also show the impact of the system dynamics on the number of lags to be considered through vector autoregressive (VAR) and vector moving average (VMA) processes. The proposed method is compared with currently available methods using simulated data.Research Limitations / Implications (if applicable)The method assumes that the same numbers of lags are added for all variables. Future research will focus on adapting our proposed method to accommodate the identification of individual time-lags for each variable. Practical Implications (if applicable)The visualization possibility of the proposed method will be useful for DPCA practitioners.Originality/Value of PresentationThe proposed method provides a tool to determine the number of lags in DPCA that works in a manner similar to the autocorrelation function (ACF) in the identification of univariate time series models and does not require several rounds of PCA. Design/Methodology/ApproachThe results are based on Monte Carlo simulations in R statistics software and in the Tennessee Eastman Process simulator (Matlab).
|
|
