multivariate data analysis: from chemometrics modeling to ... · discover how data-driven tools can...
TRANSCRIPT
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
Multivariate data analysis: from chemometrics modeling to process implementation
S Roussel, J Lallemand, S Preys
1
Discover how data-driven tools can transform your
business from R&D to production
Dr. Sylvie ROUSSEL, Ph.D. - Ondalys CEO
[email protected] – Tel : +33 (0)4 67 67 97 87
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
Making sense of your data…
based on Machine Learning
for industrial measurements
Donnez du sens à vos données
Sylvie ROUSSEL, Ph.D.
Ondalys CEO
Pharmaceutical
BiotechnologicalChemicals
Petrochemicals
Agriculture
Food Industry
Cosmetics
Environment
Public researchTechnical centers
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your dataANAYSE DE DONNEES MULTIVARIEES
Massive data available with Industry 4.0 ➔ Data Analytics for Big Data or Smart Data?
Analytical Chemistry
OMICS
HPLC – GC – MS…
RGB & Hyperspectral
Imaging
NIR
FTIR
Raman
Spectroscopy
UV – Visible – NIR – FTIR
Raman – Fluo – LIBS…
Time series
Process parameters
Industrial IoT - Inline analyzers
Time pH
Temp
(°C) …
ID01 4.8 25 …
ID02 5.2 27 …
ID03 5.0 30 …
… … … …
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
Data Analytics Methodology
Audit
Consulting
Proof of Concept
Model development
Software Implementation
Model maintenance
Training
Coaching
Audit
✓ Acquisition optimization
✓ DoE / QbD
✓ ML testing
✓ Data fusion
✓ MSPC/BSPC
✓ Scale-up
✓ Transfer
✓ Update
✓ Knowledge Transfer
✓ Open-courses
✓ In-house courses
Audit
✓ Robustness
✓ Validation
2 application cases
A. Maturity Control of plums based on portable NIR spectroscopy
B. Batch process monitoring using BSPC
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
Application case
7
Sugar and acidity content prediction
in plums based on NIR spectroscopy
Agen Prunes
Plums
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
Objective: predict sugar and acidity contents of plums directly in the field to optimize the harvest time
Proof of Concept
Model development
Model Implementation & maintenance
Model Transfer
1. Proof Of Concept:
✓ PLS model development in the lab with
✓ One year
✓ Orchard impact ?
Agen Prunes
Plums
2. Robust PLS model
✓ Lab measurements
✓ Several years
✓ Several orchards
4. Model Transfer
✓ From one lab
instrument to another
of the same brand
✓ From lab instrument
to field instrument
3. Model Maintenance & Update
✓ 4 years of data
✓ More than 5 600 fruits
✓ 15 different orchards
Indico Pro
software
Study performed for Technical Center for Prunes (BIP)
Proof of Concept
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
1. Proof of concept
Data acquisition
⚫ NIR instrument : LabSpec 4 (ASD Inc.)
• Range 350-2500 nm - Step 1 nm
• Reflectance spectra on plums in the laboratory
• 1 Year: 2013
⚫ Ref: sugar and acidity measurement with wet chemistry
Data import and preparation
⚫ Average of several measurement points
➔ Obtain the best representativity of the fruit
Model development on one year
⚫ Exploratory analysis (PCA)
• Detection of outliers
• Impact of orchard / date on the signal
⚫ Model development (PLS) for sugar and acidity prediction
• One orchard
• Several orchards
Task/ Transform/ Reduce (Average)
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
1. Proof of concept
⚫ Exploratory analysis
• No orchard difference ➔ no need for a model by orchard = one global model
• Maturity effect (plum color)
No orchard differenceMaturity effect (plum color)
Orchard :
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
1. Proof of concept
⚫ First PLS prediction model for sugar and acidity
• One global model for one year on all orchards
Sugar content (°Brix) in cross-validation
R² = 0.88
RMSECV = 1.59
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
Objective: predict sugar and acidity contents of plums directly in the field to optimize the harvest time
Proof of Concept
Model development
Model Implementation & maintenance
Model Transfer
1. Proof Of Concept:
✓ PLS model development in the lab with
✓ One year
✓ Orchard impact ?
Agen Prunes
Plums
2. Robust PLS model
✓ Lab measurements
✓ Several years
✓ Several orchards
4. Model Transfer
✓ From one lab
instrument to another
of the same brand
✓ From lab instrument
to field instrument
3. Model Maintenance & Update
✓ 4 years of data
✓ More than 5 600 fruits
✓ 15 different orchards
Indico Pro
software
Study performed for Technical Center for Prunes (BIP)
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
2. Model development and implementation
Test of the 1st PLS model on the next year
⚫ Not robust model ➔ Model update required
Test on 2014Model on 2013
R² = 0.76
RMSEP = 2.49
R² = 0.88
RMSECV = 1.59
➔ For agronomic data : Need to include several years to have a robust model
Example with sugar content (°Brix)
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
2. Robust Model
⚫ Model development on several years
2. Model development and implementation
Test on 2016Model on 2013+2014+2015
R² = 0.91
RMSEP = 1.32 ☺
R² = 0.84
RMSECV = 1.89
Example with sugar content (°Brix)
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
Objective: predict sugar and acidity contents of plums directly in the field to optimize the harvest time
Proof of Concept
Model development
Model Implementation & maintenance
Model Transfer
1. Proof Of Concept:
✓ PLS model development in the lab with
✓ One year
✓ Orchard impact ?
Agen Prunes
Plums
2. Robust PLS model
✓ Lab measurements
✓ Several years
✓ Several orchards
4. Model Transfer
✓ From one lab
instrument to another
of the same brand
✓ From lab instrument
to field instrument
3. Model Maintenance & Update
✓ 4 years of data
✓ More than 5 600 fruits
✓ 15 different orchards
Indico Pro
software
Study performed for Technical Center for Prunes (BIP)
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
3. Model implementation, Maintenance and update
16
• Final models = 4 years 2013-2016, developed on The Unscrambler X®
• Exported to IndicoPro® (LabSpec 4 software)
➔ Implemented for real time prediction
➔ compatibility
Sugar content (obrix) Acidity (Meq)
R² = 0.87
RMSECV = 1.67
R² = 0.74
RMSECV = 1.56
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
Objective: predict sugar and acidity contents of plums directly in the field to optimize the harvest time
Proof of Concept
Model development
Model Implementation & maintenance
Model Transfer
1. Proof Of Concept:
✓ PLS model development in the lab with
✓ One year
✓ Orchard impact ?
Agen Prunes
Plums
2. Robust PLS model
✓ Lab measurements
✓ Several years
✓ Several orchards
4. Model Transfer
✓ From one lab
instrument to another
of the same brand
✓ From lab instrument
to field instrument
3. Model Maintenance & Update
✓ 4 years of data
✓ More than 5 600 fruits
✓ 15 different orchards
Indico Pro
software
Study performed for Technical Center for Prunes (BIP)
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
4. Model transfer
• NIR instruments
• Lab instrument: LabSpec 4 (ASD Inc.)
• Range 350-2500 nm
• Step 1 nm
• Reflectance spectra
• Years: 2013-2016
• Field instrument: MicroNIRTM OnSite (VIAVI Solutions Inc.)
• Range 800-1676 nm
• Step ~6 nm
• Absorbance spectra
• Year : 2016
18
Transfer needed !
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
4. Model transfer
19
4. Model transfer
• Transfer from lab (LabSpec 4) to field (MicroNIRTM OnSite)
• Lab instrument :
conversion from reflectance to absorbance
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
4. Model transfer
20
• Transfer from lab (LabSpec 4) to field (MicroNIRTM OnSite)
• Lab instrument :
reduction of the spectral range and interpolation to cover MicroNIR range
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
4. Model transfer
21
4. Model transfer
• Transfer from lab (LabSpec 4) to field (MicroNIRTM OnSite)
• Comparison of spectra before transfer
LabSpec 4transformed in absorbance
and interpolated across MicroNIR range
MicroNIR
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
4. Model transfer
22
4. Model transfer
• Transfer from lab (LabSpec 4) to field (MicroNIRTM OnSite)
• Transfer trials• Piecewise Direct Standardization (PDS)
• Mean difference spectrum
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
4. Model transfer
23
4. Model transfer
• Transfer from lab (LabSpec 4) to field (MicroNIRTM OnSite)
• Comparison of spectra after transfer
LabSpec4 spectraAfter transfer
MicroNIR spectra
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
4. Model transfer
24
• MicroNIRTM OnSite models
• Concatenation of the LabSpec transferred historical database [2013-2016] with MicroNIRTM spectra measured in 2016
• Redevelopment of models with The Unscrambler X®
• Possible export to MicroNIRTM OnSite Software
➔Results on an independent
test set (2017) are quite
promising
➔Need to add progressively
new “real” spectra to
improve models, especially
for acidity
➔ Compatibility
Sugar content (Brix)
Slope Offset RMSEPR-
Square
LabSpec 4 0.87 2.75 1.7 0.87
LabSpec 4 transf+ MicroNIR
0.83 3.59 1.9 0.83
Acidity (Meq)
Slope Offset RMSEPR-
Square
LabSpec 4 0.74 24.5 15.6 0.74
LabSpec 4 transf+ MicroNIR
0.51 46.5 22.8 0.48
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
Industrial Application case withThe Unscrambler® X - Process Pulse II
25
2019 Analytics Solution Conference – MN, June 18-20
CONSORTIUM OF INDUSTRIALS INDUSTRIAL ANALYSIS PLATFORM
SENSORS & CHEMOMETRICS
FOR PROCESS MONITORING
USING PROCESS PULSE II
DATA ANALYTICS & PROCESS MONITORING SOFTWARE PROVIDER
DATA ANALYTICS & MACHINE LEARNING
CONSULTING AND TRAINING
2019 Analytics Solution Conference – MN, June 18-20
AXEL’ONE ANALYSIS CONSORTIUM OF INDUSTRIALS
27
R&D & IMPLEMENTATION TOPICS
The platform works on 3 themes :
1. Separation techniques
2. Spectral techniques and chemometrics
3. Physical and chemical sensors
2019 Analytics Solution Conference – MN, June 18-20 Making sense of your data
Main objective: process monitoring using at-line/online/inline analyzers
1. Proof of Concept
2. Model development
3. Software Implementation
4. Model maintenance
✓ Data fusion (several analyzers)
✓ MSPC
✓ BSPC
✓ Scale-up
✓ Transfer
✓ Update
✓ Robustness
✓ Validation
✓ Continuous process:
Multivariate Statistical Process Control (MSPC)
✓ Batch process:
Batch Statistical process control (BSPC)
1. Proof of Concept
2019 Analytics Solution Conference – MN, June 18-20
EXEMPLE: BATCH STATISTICAL PROCESS CONTROL (BSPC)
29
Process
Proof of concept & Model development
Process parameters: temperature, pressure,
speed, etc.
Univariate sensors: pH, viscosity, etc…
Inline spectroscopy:
Raman, NIR, UV-VIS, etc…
BSPC model: Golden batch (PCA)
Concatenated datawith Maturity index
Under-control trajectory
New batch monitoring ➔ real-time drift
detection
2019 Analytics Solution Conference – MN, June 18-2030
Model implementation
EXEMPLE: BATCH STATISTICAL PROCESS CONTROL (BSPC)
2019 Analytics Solution Conference – MN, June 18-20
THANK YOU
FOR YOUR ATTENTION!
Sylvie ROUSSEL, Ph.D.
[email protected] – Tel: +33(0)4 67 67 97 87
Minneapolis