Elaine Martin Centre for Process Analytics and Control Technology University of Newcastle, England www.ncl.ac.uk/cpact/ The Conjunction of Process and.

Slides:



Advertisements
Similar presentations
Location Forum 2006, 7 November, 2006 School of Surveying & Spatial Information Systems The University of New South Wales, Australia Adaptive Kalman Filtering.
Advertisements

Regression analysis Relating two data matrices/tables to each other Purpose: prediction and interpretation Y-data X-data.
Artificial Intelligence 13. Multi-Layer ANNs Course V231 Department of Computing Imperial College © Simon Colton.
Stat 112: Lecture 7 Notes Homework 2: Due next Thursday The Multiple Linear Regression model (Chapter 4.1) Inferences from multiple regression analysis.
Pattern Recognition and Machine Learning
This PowerPoint presentation shows you how to use the NRM 1.0.xls Excel Workbook to fit several popular regression models to experimental data. The models.
« هو اللطیف » By : Atefe Malek. khatabi Spring 90.
Introduction The aim the project is to analyse non real time EEG (Electroencephalogram) signal using different mathematical models in Matlab to predict.
CMPUT 466/551 Principal Source: CMU
The loss function, the normal equation,
QUALITY CONTROL OF COMPOSITION OF BLACK POLYMERES.
Multiple Criteria for Evaluating Land Cover Classification Algorithms Summary of a paper by R.S. DeFries and Jonathan Cheung-Wai Chan April, 2000 Remote.
Radial Basis Functions
PERFORMANCE OF THE DELPHI REFRACTOMETER IN MONITORING THE RICH RADIATORS A. Filippas 1, E. Fokitis 1, S. Maltezos 1, K. Patrinos 1, and M. Davenport 2.
CALIBRATION Prof.Dr.Cevdet Demir
Multivariate Data Analysis Chapter 4 – Multiple Regression.
LAB 3 AIRBAG DEPLOYMENT SENSOR PREDICTION NETWORK Warning This lab could save someone’s life!
The Application of Partial Least Squares to Non-linear Systems in the Process Industries Elaine Martin and Julian Morris Centre for Process Analytics and.
Lecture 17 Today: Start Chapter 9 Next day: More of Chapter 9.
,. Sugar measurements in soybeans using Near Infrared Spectroscopy Introduction  Soluble carbohydrates are the third compound of soybeans by weight (11%),
Face Processing System Presented by: Harvest Jang Group meeting Fall 2002.
Uncertainty analysis is a vital part of any experimental program or measurement system design. Common sources of experimental uncertainty were defined.
Neural Networks And Its Applications By Dr. Surya Chitra.
SPECTRAL AND HYPERSPECTRAL INSPECTION OF BEEF AGEING STATE FERENC FIRTHA, ANITA JASPER, LÁSZLÓ FRIEDRICH Corvinus University of Budapest, Faculty of Food.
Radial Basis Function Networks
Face Detection using the Viola-Jones Method
Colorado Center for Astrodynamics Research The University of Colorado STATISTICAL ORBIT DETERMINATION Project Report Unscented kalman Filter Information.
Statistical Methods For Engineers ChE 477 (UO Lab) Larry Baxter & Stan Harding Brigham Young University.
Attention Deficit Hyperactivity Disorder (ADHD) Student Classification Using Genetic Algorithm and Artificial Neural Network S. Yenaeng 1, S. Saelee 2.
PATTERN RECOGNITION AND MACHINE LEARNING
1 Least squares procedure Inference for least squares lines Simple Linear Regression.
Introduction to variable selection I Qi Yu. 2 Problems due to poor variable selection: Input dimension is too large; the curse of dimensionality problem.
Chapter 11 – Neural Networks COMP 540 4/17/2007 Derek Singer.
Quantification of the non- parametric continuous BBNs with expert judgment Iwona Jagielska Msc. Applied Mathematics.
Pattern Classification All materials in these slides were taken from Pattern Classification (2nd ed) by R. O. Duda, P. E. Hart and D. G. Stork, John Wiley.
Time Series Data Analysis - I Yaji Sripada. Dept. of Computing Science, University of Aberdeen2 In this lecture you learn What are Time Series? How to.
Neural Networks for Protein Structure Prediction Brown, JMB 1999 CS 466 Saurabh Sinha.
1 RECENT DEVELOPMENTS IN MULTILAYER PERCEPTRON NEURAL NETWORKS Walter H. Delashmit Lockheed Martin Missiles and Fire Control Dallas, TX 75265
Current and Future Applications of the Generic Statistical Business Process Model at Statistics Canada Laurie Reedman and Claude Julien May 5, 2010.
CSC321: 2011 Introduction to Neural Networks and Machine Learning Lecture 9: Ways of speeding up the learning and preventing overfitting Geoffrey Hinton.
Time Series Analysis and Forecasting
Digital Media Lab 1 Data Mining Applied To Fault Detection Shinho Jeong Jaewon Shim Hyunsoo Lee {cinooco, poohut,
A comparison of the ability of artificial neural network and polynomial fitting was carried out in order to model the horizontal deformation field. It.
Metabolomics Metabolome Reflects the State of the Cell, Organ or Organism Change in the metabolome is a direct consequence of protein activity changes.
Use of spectral preprocessing to obtain a common basis for robust regression 5 spectral preprocessing combinations gave significantly higher RPDs (α =
Solution of a Partial Differential Equations using the Method of Lines
A n = c 1 a n-1 + c2an-2 + … + c d a n-d d= degree and t= the number of training data (notes) The assumption is that the notes in the piece are generated.
Prediction of NMR Chemical Shifts. A Chemometrical Approach К.А. Blinov, Y.D. Smurnyy, Т.S. Churanova, М.Е. Elyashberg Advanced Chemistry Development (ACD)
Chimiometrie 2009 Proposed model for Challenge2009 Patrícia Valderrama
Comparison of PLS regression and Artificial Neural Network for the processing of the Electronic Tongue data from fermentation growth media monitoring Alisa.
Time Series Analysis and Forecasting. Introduction to Time Series Analysis A time-series is a set of observations on a quantitative variable collected.
Feature Selection and Weighting using Genetic Algorithm for Off-line Character Recognition Systems Faten Hussein Presented by The University of British.
1 January 24, 2016Data Mining: Concepts and Techniques 1 Data Mining: Concepts and Techniques — Chapter 7 — Classification Ensemble Learning.
Classification and Prediction: Ensemble Methods Bamshad Mobasher DePaul University Bamshad Mobasher DePaul University.
QUANTITATIVE ANALYSIS OF POLYMORPHIC MIXTURES USING INFRARED SPECTROSCOPY IR Spectroscopy Calibration –Homogeneous Solid-State Mixtures –Multivariate Calibration.
REU 2009-Traffic Analysis of IP Networks Daniel S. Allen, Mentor: Dr. Rahul Tripathi Department of Computer Science & Engineering Data Streams Data streams.
Standardization of NIR Instruments: How Useful Are the Existing Techniques? Benoit Igne Glen R. Rippke Charles.
A “Peak” at the Algorithm Behind “Peaklet Analysis” Software Bruce Kessler Western Kentucky University for KY MAA Annual Meeting March 26, 2011 * This.
COMPARATIVE STUDY BETWEEN NEAR- INFRARED(NIR) SPECTROMETERS IN THE MEASUREMENT OF SUCROSE CONCENTRATION.
Mustafa Gokce Baydogan, George Runger and Eugene Tuv INFORMS Annual Meeting 2011, Charlotte A Bag-of-Features Framework for Time Series Classification.
Studies on the feasibility of using chemometric modeling of spectral data for the determination of post-mortem interval of skeletal remains. Kenneth W.
Combining Models Foundations of Algorithms and Machine Learning (CS60020), IIT KGP, 2017: Indrajit Bhattacharya.
Deep Feedforward Networks
第 3 章 神经网络.
Interval selection complexity
Diagnostics and Transformation for SLR
Soft Error Detection for Iterative Applications Using Offline Training
The loss function, the normal equation,
Diagnostics and Transformation for SLR
Presentation transcript:

Elaine Martin Centre for Process Analytics and Control Technology University of Newcastle, England The Conjunction of Process and Spectral Data for Enhanced Fault Detection

Motivation n It is conjectured that there may be factors relating specifically to a process that cannot be identified from the spectroscopic measurements that could be described by the process data or vice versa. n Consequently one way to enhancing prediction accuracy and process performance and fault detection is through the integration of process and spectral data. n The aim of the subsequent studies was to investigate the combined power of spectral and process data.

Overview n Process Modelling l Fermentation Process Spectral Data Spectral and Process Data n Process Monitoring and Fault Detection l Polymer-resin Manufacturing Process Data Process and Spectral Data

Challenges in the Monitoring of Fermentation Processes n Fermentation is a process in which micro-organisms convert chemical species to products of higher value. n On-line information relating to the progression of the process is not easily attained. n Near Infrared and Mid Infrared spectroscopy have been applied for the monitoring of fermentation processes. n The successful implementation of these spectroscopic approaches necessitates the application of appropriate multivariate data analysis techniques, such as partial least squares (PLS).

Experimental Data Set n The industrial pilot-plant scale Streptomyces fermentation process involves two stages: l Seed stage l Final stage n The seed stage materialises in the generation of biomass. l The starting ingredients include carbohydrate, soya protein, vegetable oil and trace elements in water. n The biomass is transferred to the final stage for the production of the desired product. l The final stage is a fed batch process lasting approximately 140hrs. n NIR measurements were collected for the final stage of the process.

Spectra Data Acquisition n The NIR spectral data were recorded using a Zeiss Corona 45

Description of the Data Set n Final stage data from 7 standard batches and 7 Design of Experiment batches form the basis of the subsequent analysis. n Data collected included on-line process data, off-line data, biochemical and NIR measurements.

Methodological Summary n Pre-processing of the spectral data set l First derivates l Splining n Segmented wavelength region selection n Global modelling – Linear PLS, Neural Network PLS, Quadratic PLS n Local modelling - Linear PLS, Neural Network PLS, Quadratic PLS n Bagging of the models l Linear partial least squares l Averaging

Data Pre-processing n The NIR data (Zeiss Corona NIR) were recorded every 15 minutes and the first derivatives were taken. n Since only ten values of titre were recorded, a spline was fitted to the data. l The splined titre values were aligned to the 550 spectral values for each batch. n The range utilised for both the spectral and quality data was to 125 log hours.

Data Pre-processing

NIR Data and First Derivatives NIR Data First Derivative

Spectral Window Selection Algorithm N Select training and validation batches Mean centre and take derivatives of the spectral data Generate random centres and widths Build model ‘input’ matrix eliminating common data. Generate PLS model Calculate RMS errors Generate random changes to centres and widths Apply the random changes to the current centres and widths Build new input matrix, generate model and calculate RMS errors Has the RMS on training data decreased? Has number of iterations been exceeded and there are more models to build ? Present the final bagged model N Y Y

Spectral Window Selection Algorithm Centre Width Generate random increment in centre and width Centre Width Update the centre and width Take another step with the Centre and Width increment Step too far. The prediction error has increased. Go back to where we were. Generate a new increment in centre and width and continue search Has the prediction error decreased? Yes, then a step in the right direction

Benefits of the SWS Algorithm n SWS offers the opportunity to consider not only the extremes of a single wavelength and the full set but also restricts selection to multiple sub-sets of the full set. n Finds the ‘best’ possible models for the product concentration and the biochemical components. n Finds the ‘best’ wavelength range from which these models can be built.

Bagging n SWS does not provide a unique model. n To obtain a more robust model, bagging is implemented. n ‘Resample and Combine’ method or ‘bagging’ is an algorithm that helps improve the robustness of models by combining predictions from different models.

Bagging of Models n 30 models were generated by changing the initial random seed of the wavelength selection algorithm. n Bagging was applied to the 30 models: l The average value was calculated from the output of the 30 models. l A PLS model was fitted between the real and fitted values to give a weighted average.

Global and Local Modelling

2 critical points at 70 and 100 hours were identified from plots of the biochemical data Local Modelling

First Time IntervalSecond Time IntervalThird Time Interval

Local Modelling Approach n Three time regions for both the spectra and the quality variable values (titre) were selected.  Samples up to 70 log hours, i.e sample points.  From 70 log hours to 100 log hours, i.e sample points.  From 100 log hours up to the end of the chosen window, i.e sample points.

Local Modelling Approach Region 2 Region 3

Results : Time Interval 1

n The RMS of the training set for models 1, 7 and 29 is large. n The RMS of the validation data set for models 1, 7 and 29 is small. n The RMS error for PLS Bagging is smaller than the error of each individual model RMS error after PLS Bagging

Linear PLS – Region 1 (Wavelength Selection) Training Data Set Validation Data Set

Results : Time Interval 1 The wavelengths between 30 and 40 are selected most frequently.

Neural Network PLS – Region 2 (Wavelength Selection) Training Data Set Validation Data Set

Polynomial PLS – Region 3 (Wavelength Selection) Training Data Set Validation Data Set

Local Modelling : Training Data Set Global Modelling Local Modelling Global Modelling predictions Local Modelling predictions for time intervals 1, 2 and 3

Local Modelling : Validation Data Set 1rst Time Interval 2nd Time Interval 3rd Time Interval

Genetic Algorithm Results Genetic algorithms provide the possibility of selecting individual wavelengths but potentially does not predict future samples well. SWS Genetic Algorithms

GA Results – Region 2 SWS Averaging Ga’S Averaging RMS of Validation - SWS: GAs:0.069

Genetic Algorithm Results Time Interval 1Time Interval 2Time Interval 3 PLS Bagging Average Bagging PLS Bagging Average Bagging PLS Bagging Average Bagging SWS with Linear PLS GAs with Linear PLS TRAINING Time Interval 1Time Interval 2Time Interval 3 PLS Bagging Average Bagging PLS Bagging Average Bagging PLS Bagging Average Bagging SWS with Linear PLS GAs with Linear PLS VALIDATION RESULTS

Summary of Results n GAs produced slightly better predictions for the training data set resulting in overfitting. n In the validation model, SWS combination with bagging for local modelling gave better results than the GA in combination with bagging. n Local modelling gives better results than global modelling. n SWS with bagging gives better results compared with the purported ‘one-shot wonder’ models.

Design of Experiment Data Integration of Process and Spectral Data

Conjunction of Process and Spectral Data n In the later stages of the fermentation, the error in the calibration models was observed to be greater with offsets being present. n During this time, significant changes in the fermentation broth concentrations occur. n The offset can potentially be modelled by utilising other process information such as off-gas measurements.

Data Set and Aim n The aim is to infer product concentration and the biochemical components from the spectral data. n Working on the off-line, biochemical and NIR data for the design of experiment batches. n Changing conditions in experimental design: Temperature (°C) pH Sugar feed (gh -1 ) Oil feed (%)

Conjunction of Process and Spectral Data MODEL Spectral Σ + Biochemical Concentration - Calibration spectral residuals MODEL Process Data Σ + Calibration Spectral Residuals - Innovations First Step: Calculation of the calibration spectral residuals. Second Step: Modelling of the calibration spectral residuals from the process data and the generation of the innovations. Σ Biochemical Concentration Predictions by Spectra Residuals Prediction by Process Data Final Product Concentrations Final Step: Prediction of the product concentration

Conjunction of Process and Spectral Data CER CO 2 Total pH OUR Temperature 5 variables were considered to be the most important for the prediction of product concentration Time Series Plot 5 pH Time Series Plot 2 CER Time Series Plot 3 CO2 Total Time Series Plot 9 OUR

Predicted train values Conjunction of Process and Spectral Data Predictions Residuals Residuals for training data set Predicted valid values

Final predictions of the product Real values, Predicted values and Final predicted values for valid New residuals The off-set is reduced The residuals exhibit less structure and reflect noise Conjunction of Process and Spectral Data

Conclusions n A Spectral Window Selection (SWS) algorithm has been proposed to select a window of wave numbers. n Multiple models are ‘bagged’ to produce a more robust model. n SWS produces better results than when the complete wavelength region is included. n Process data was combined with spectral data to eliminate offsets. n The wavelength selection-bagging approach in combination with the process data is now under investigation. n The results to date are promising.