ENBIS: European Network for Business and Industrial Statistics
Forgotten your password?
Not yet a member? Please register
ENBIS-15 in Prague
6 – 10 September 2015; Prague, Czech Republic Abstract submission: 1 February – 3 July 2015The following abstracts have been accepted for this event:
-
A Novel Method to Deal with Latent Ability Testing and Evaluation
Authors: Emil Bashkansky (ORT Braude College of Engineering), Vladimir Turetsky (ORT Braude College of Engineering)
Primary area of focus / application: Design and analysis of experiments
Secondary area of focus / application: Modelling
Keywords: Maximum likelihood, Testing, Latent ability, Difficulty, Item response
-
Empiricism and the Root Cause Analysis Helix
Authors: Matthew Barsalou (BorgWarner Turbo Systems Engineering GmbH)
Primary area of focus / application: Quality
Secondary area of focus / application: Quality
Keywords: Root Cause Analysis, Exploratory Data Analysis, Empiricism, Quality
-
Confirmation - The Final DOE Step
Authors: Pat Whitcomb (Stat-Ease, Inc.), Martin Bezener (Stat-Ease, Inc.), Wayne Adams (Stat-Ease, Inc.)
Primary area of focus / application: Design and analysis of experiments
Keywords: Design of Experiments (DOE), Confirmation, Verification, Predictive model
Simple confirmation,
Concurrent confirmation,
Verification DOE.
The presentation begins by spelling the importance of confirming results from a designed experiment. (Not confirming is not a recommended strategy.) It then provides practical advice and examples of three alternative approaches to implement a confirmation strategy. Attendees will come away with valuable tools to confirm their models as the final step in their DOE process. -
Comparing Process Data to PCA-Based Contribution Plots for Model-Based Fault Identification in Batch Processes
Authors: Sam Wuyts (KU Leuven), Geert Gins (KU Leuven), Pieter Van den Kerkhof (KU Leuven), Jan Van Impe (KU Leuven)
Primary area of focus / application: Modelling
Secondary area of focus / application: Mining
Keywords: Fault detection, Fault diagnosis, Chemical batch processes, Model-based classification, Process data mining
The (bio)chemical industry relies strongly on batch processes, certainly for products with high added value such as pharmaceuticals or specialty chemicals. However, their dynamic characteristics and the difficulty to measure product quality online inherently limit their controllability. This presents major challenges for monitoring, control, fault detection and diagnosis in batch processes.
Statistical Process Control (SPC) is typically applied to tackle these challenges. The development of SPC is supported by the availability of large historical databases in chemical plants, containing all available online measurements of a large set of sensors. These databases hold tremendous information regarding process operation, which is exploited by SPC to establish a fast and reliable monitoring procedure.
Problem statement
Once an abnormal situation is detected, the underlying root cause still needs identification. Experts and operators typically use contribution plots [3] to interpret and diagnose process upsets. Contribution plots require no prior knowledge about the underlying disturbances but do not always unequivocally indicate the variable(s) responsible for the fault. Hence, expert interpretation is always required for reliable fault identification.
The expert interpretation of the contribution pattern is bypassed when all possible faults are known. An automated classification model, pinpointing the class most likely causing the fault, is trained on contribution patterns of past faults. Automation of the diagnosis significantly reduces time between fault detection and corrective actions. However, reliability is impacted by fault smearing, which negatively influences the accuracy of the variables’ contributions [4,5]. Therefore, it might be beneficial to consider alternative data patterns (i.e., patterns not subject to fault smearing) as input to the classification model.
This paper compares automatic classification based on contributions to classification using sensor data.
Results
Datasets representing a benchmark penicillin fermentation Pensim [1] are simulated in RAYMOND [2]. Two cases are studied: the first focuses on the intrinsic diagnosability of upsets by only considering basic measurement noise on a limited number of sensor variables while the second study extends the conclusions towards datasets including more complex measurement noise on multiple sensors as well as biological variability. Different pretreatments of both contributions as well as sensor data are employed to maximize performance. These manipulations are chosen based on the intrinsic nature of the faults and significantly improve performance.
The two studies yield some guidelines about the appropriate pretreatment for faults of different nature. Normalizing data around the average batch trajectory prior to classification is advisable for faults with varying starting times. Classification for faults with both positive and negative deviations from the average trajectory is improved by taking absolute values. Taking into account time windows helps to increase distinction between drift faults and step or drop faults. For both studies, better classification is achieved using pretreated sensor measurements rather than variables’ contributions since the latter are subject to the negative influence of fault smearing.
References
[1] Comput. Chem. Eng., 26(11):1553–1565, 2002.
[2] Comput. Chem. Eng., 69:110–118, 2014.
[3] ISA Trans., 37(1):41–59, 1996
[4] Chem. Eng. Sci., 104:285–293, 2013.
[5] Chemometr. Intell. Lab., 51(1):95–114, 2000. -
Business Strategy Embraces Data Analytics to Meet New Challenges in the Shipping Industry
Authors: Shirley Coleman (ISRU, Newcastle University), Ibna Zaman (ISRU, Newcastle University), Rose Norman (MAST, Newcastle University), Kayvan Pazouki (MAST, Newcastle University)
Primary area of focus / application: Business
Secondary area of focus / application: Mining
Keywords: Emissions, Monitoring, Visualisation, Consumption, Offshore
-
Behavior Based Price Enabled by Predictive Modelling
Authors: Andrea Ahlemeyer-Stubbe (ahlemeyer-stubbe)
Primary area of focus / application: Mining
Secondary area of focus / application: Business
Keywords: Behavior based price, Predictive modelling, Big Data, Automation
Submitted at 11-Mar-2015 19:43 by Andrea Ahlemeyer-Stubbe
Accepted
To detect fast changes in customer behavior or to react in as focused a manner as possible, predictive modeling must be done in good quality to get effective predictions of customer price affinities and it has to be done fast to be relevant under business aspects. Modeling speed is of great importance in industry as time is a crucial factor. This necessity requires a different technical set up for model development to fulfill both needs: quality and development speed. Today most companies like to develop their models individually with the help of specialists. But for a lot of companies, this way takes too long; even though the models are excellent, the time to develop them sometimes kills the advantages of a better prediction. This article describes the general structure and ideas how to implement industry-focused model production that will help to react quickly to changing behavior with relevant prices.