A powerful decision-tree approach to multivariate modeling of complex data - STATCON · A powerful...
Transcript of A powerful decision-tree approach to multivariate modeling of complex data - STATCON · A powerful...
A powerful decision-tree approach to multivariate modeling of complex data
Hierarchical modeling solutions
Make modeling of non-linear data more efficient and robust
Join Classification and Prediction models using multi-level modeling
Use Hierarchical Models on-line with instrument integration
Bring data to life
Hierarchical ModelDevelopment Module
Hierarchical ModelEngine
02 CAMO Hierarchical Model Development Module & Engine
Unscrambler® X Hierarchical Model Development Module & Hierarchical Model Engine
When analyzing process or other complex data, it is often necessary
to refine models based on the output of initial investigations. In most
cases, this is done manually and in many steps, becoming a laborious
and time consuming process which is prone to error.
Hierarchical Modeling (HM) is the joining of a number of multivariate
models using logic statements in order to arrive at a single, unique
result. This is the classic logic tree or decision tree approach to
systematically solve problems, where the analysis in one step is guided
by the previous step.
With most data, it is very difficult to make a global prediction or
classification model that predicts well in every area. In many real-world
applications non-linear models are needed to understand the data.
Hierarchical models address these issues by focusing, or ‘zooming in’,
on smaller areas of a data set more precisely.
Hierarchical models are ideal for applications such as:
Fermentation analysis of batches in biotech, food, brewing, wine
Managing products with hard to control or unwanted additives
Gasoline blending in petroleum refining
Reaction monitoring in chemical manufacturing
Classifying and characterizing pharmaceutical raw materials for
better campaigning into batches
Early stage development of drug formulations and manufacturing
systems used to develop them
Manually developing models with 3 or 4 levels is very time consuming, but with hierarchical models you only need to set it up once and validate once, then you can use the model as much as required in run time applications.
Once set up, validated models can be applied across different sites and processes for use by non-experts, enabling efficient process management and technology transfer.
Hierarchical models are ideal for at-line process analysis. They allow non-normal situations to be identified and alarmed so that third-party control systems can make quality decisions.
In the example above, the material is first classified by the model, and then the correct prediction meth-ods are assigned automatically. Alarms are shown by the red cells, where the model has detected that the samples are new or unknown.
We offer a range of tools so you can use the power of hierarchical
models on your data in the way that suits you.
Projection methods (PCA, PCR, PLSR)
Classification methods (SIMCA, LDA, SVM)
Regression methods (MLR, PCR, PLSR)
Accepts data in wide range of formats
Auto-pretreatment options for data
Alarms and warnings can be set
Output in tabular format with cell colours according to
alarm state
Can be intergated with 3rd party systems including OPC UA
Hierarchical Model Development Module & Engine 03 CAMO
Hierarchical models are very useful for complex classification problems such as raw material analysis. If a global classification model cannot uniquely identify all the groups, a hierarchical model can be developed to do the classification in a stepwise manner. This is a convenient way to join multivariate models into one versatile tool.
The example above shows a global model of different raw material spectra developed and plotted in The Unscrambler® X software. Hierarchical modeling is performed by joining together up to 10 levels of models in The Unscrambler® X.
Hierarchical models are useful if dealing with non-linear regression problems over large concentration ranges. The first predictor lets you define where you are, then the second model gives a more precise analysis of the area in question. In this manner a set of robust and locally validated models can replace highly non-linear models that are prone to over-fitting.
The example above shows how a non-linear response can be modeled by two linear models, each applied to different concentration ranges.
Hierarchical Modeling tools Key features
Hierarchical Model Development Module: Plug-in module for The Unscrambler® X software for off-line data
analysis or to develop hierarchical models for on-line use.
Hierarchical Model Engine: Apply hierarchical models in real-time by integrating them into
scientific instruments or process equipment.
Full customization options: Unscrambler® X Hierarchical Models can be integrated into process
monitoring or control systems with customization as required.
www.camo.com
NORWAY
Nedre Vollgate 8,
N-0158
Oslo
Tel: (+47) 223 963 00
Fax: (+47) 223 963 22
USA
One Woodbridge Center
Suite 319, Woodbridge
NJ 07095
Tel: (+1) 732 726 9200
Fax: (+1) 973 556 1229
INDIA
14 & 15, Krishna Reddy
Colony, Domlur Layout
Bangalore - 560 071
Tel: (+91) 80 4125 4242
Fax: (+91) 80 4125 4181
JAPAN
Shibuya 3-chome Square Bldg 2F
3-5-16 Shibuya Shibuya-ku
Tokyo, 150-0002
Tel: (+81) 3 6868 7669
Fax: (+81) 3 6730 9539
AUSTRALIA
PO Box 97
St Peters
NSW, 2044
Tel: (+61) 4 0888 2007
© 2013 CAMO Software AS
Find out more Connect with CAMOFor more information please contact your regional CAMO office or email [email protected]
http://www.linkedin.com/company/camo-software
https://twitter.com/camosoftware
https://www.facebook.com/camosoftware
http://dataanalysis.blog.camo.com/
http://www.youtube.com/camoseo
Apply 1st level(Global)
classification model
Class 1Report Class 1ID + PredictedValues and/orDiagnostics
LEVEL 1
Example workflow of a Hierarchical Model
LEVEL 2 LEVEL 3
Report Class 2ID + PredictedValues and/orDiagnostics
Report Class 3ID + PredictedValues and/orDiagnostics
Report Class 4ID + PredictedValues and/orDiagnostics
PLS Model 1
PLS Model 2
PLS Model 3
PCR Model 1
PLS Model 4
MLR Model 1
PCR Model 2
Class 2
Class 3
Class 4
Hierarchical models allow a number of multivariate models to be joined using logic statements in order to arrive at a single, unique result.
Most of the powerful classification, prediction or projection methods in The Unscrambler® X can be used as the building blocks.
The hierarchical model is built in a top-down manner, with a global model in level 1
Conditional statements about the outcome at one level define the actions to be taken on the next level
Outcomes such as classification, prediction, deviation, leverage, etc. can be tested
Up to 10 levels can be defined
Events and alarms are set up to highlight suspect samples or stop the hierarchy