MACHINE OLFACTION: ADVANCED EXCITATION METHODS FOR INORGANIC CHEMORESISTORS

1
MACHINE OLFACTION ADVANCED EXCITATION METHODS
FOR INORGANIC CHEMORESISTORS

• R. Gutierrez-Osuna
• Wright State University

2
Outline

• Introduction
• The electronic nose
• Temperature modulation
• Experimental setup
• Analyte selection and exposure
• Temperature modulation profiles
• Data analysis
• Selectivity improvements
• Pattern stability
• Discussion
• Conclusions
• Future work

3
Acknowledgements

• Collaborators
• Shino Korah (Wright State University)
• Alex Perera (Univesitat de Barcelona)
• Funding
• National Science Foundation
• Universal Energy Systems, Inc.
• Wright State University

4
INTRODUCTION
5
What is an electronic nose?

• An instrument which combines
• an array of electronic chemical sensors with
  partial specificity
• an appropriate pattern-recognition system capable
  of recognizing simple or complex odours

6
Chemoresistive sensors

• Absorption of gases modifies conductivity of a
  sensing membrane
• Metal Oxide Semiconductors (MOS)
• Conducting Polymers (CP)

7
Metal-oxide chemoresistors

• Limitations of MOS sensors
• Poor selectivity
• High power consumption
• Approaches to improve the selectivity of COTS
  sensors
• Computational
• Transient response analysis
• Analytical
• Thermal desorption, chromatography, filters
• Instrumentation
• Temperature modulation, AC impedance

8
MOS chemoresistor excitation

• Isothermal operation
• Heater voltage is maintained constant during
  exposure to analytes
• Sensor resistance is measured at one temperature

• Temperature modulation
• Heater voltage is cycled during exposure to
  analytes
• Sensor resistance is measured at each temperature

9
MOS transduction principle

• In clean air
• Atmospheric oxygen chemisorbed on the surface
• Electronic carriers are tied, creating a
  potential barrier ?b
• As a result, the conductivity of the MOS
  decreases
• In the presence of reactive gases
• Oxygen reacts and is removed from the surface
• Electrons are freed
• The conductivity of the MOS increases
• Why temperature modulation?
• The stability of oxygen species (O-2, O-, O2-)
  will depend on temperature
• Different gases have different optimal reaction
  temperatures

10
EXPERIMENTAL
11
Experimental setup

• Two commercial MOS sensors
• Figaro TGS 2610
• FIS SB11A
• Static headspace analysis
• 30ml glass vial with 10 ml analytes
• Sensor inserted through a tight aperture on the
  cap
• This setup eliminates cooling effects by effluent
  flow
• Analyte database
• Blank (air)
• Vinegar (5 acetic acid)
• Ammonia
• Isopropyl Alcohol
• Acetone
• Data collection
• 10 days, 30 samples/analyte

12
Analyte concentration

• How to test for selectivity enhancements if
  analytes can be discriminated isothermally?
• Each analyte is serially diluted in water until
  the sensor response is the same for all the
  analytes
• Therefore, the concentration range is at or below
  the isothermal discrimination threshold at
  nominal temperature

13
Instrumentation

• Data acquisition
• Personal computer with a data-acquisition card
• Data generation and acquisition at 100Hz
• Measurement
• Sensitive element placed
• in a voltage divider
• Heater excitation
• Analog output generates heater voltage
• Current-boosting with a Darlington pair

14
Heater profile

• Isothermally
• Heater voltage maintained at manufacturers
  nominal value
• Temperature-modulation
• SIX segments at 0.125Hz, 0.25Hz, 0.5Hz, 1Hz, 2Hz
  and 4Hz
• TEN cycles per frequency

15
Sensor response
ISOTHERMAL
TEMPERATURE-MODULATION
16
ANALYSIS
17
Physical structure of the sensors
TGS 2610
FIS SB11A
18
Pattern recognition for the e-nose

• A classical architecture
• A variety of Pattern Recognition and Machine
  Learning techniques available for each module
• Only preprocessing module is sensor dependent

19
Pattern analysis

• Preprocessing
• Sub-sampling (down to 25 features per TM pattern)
• Dimensionality reduction
• Linear Discriminants Analysis (down to 4
  dimensions)
• Classification
• K Nearest Neighbors (kN/NC/2)
• Validation
• 10-fold cross-validation (1 fold 1 day)

20
Pattern stability

• How to measure the stability of sensor patterns
  over time?
• Increasing training data (N) allows the
  pattern-classifier to filter out the drift
• Larger time-stamp differences (D) between
  training and test data are likely to reduce
  pattern-classification rate
• Worst-case scenario
• Set N1 ? pattern-classifier is trained on data
  from a single day

21
Predictive accuracy over time
A 0.125 Hz B 0.250 Hz C 0.500 Hz D 1.000
Hz E 2.000 Hz F 4.000 Hz
22
Drift compensation

• Drift behavior
• Mostly multiplicative (gain)
• Also additive (offset)
• Compensation by normalization

23
Normalized patterns
24
Predictive accuracy over time
A 0.125 Hz B 0.250 Hz C 0.500 Hz D 1.000
Hz E 2.000 Hz F 4.000 Hz
25
DISCUSSION
26
Conclusions

• Selectivity
• Temperature modulation increases the selectivity
  well below the isothermal discrimination
  threshold
• Information content
• At low frequencies in the shape
• At high frequencies in the DC offset
• Speed
• FIS allows for faster frequencies than TGS due to
  physical dimensions
• Stability
• Both sensors are, unfortunately, also affected by
  drift
• TGS appears to be more stable than FIS
• Poisoning by acetic acid?

27
Future work

• Study pattern stability over longer periods of
  time (weeks, months)
• Study pattern repeatability across nominally
  identical sensors
• Drift compensation by normalization with respect
  to a reference gas
• Merging information from multiple frequencies
• Heater resistance control as opposed to heater
  voltage control
• Discrimination performance with mixtures and
  complex odors