1. University of Kentucky, Auburn UniversitySlide 1 System Level Design of Chemical Sensing Microsystems D.M. Wilson University of Kentucky, Electrical Engineering T. Roppel and M.L. Padgett Auburn University, Electrical Engineering April 2, 1998 2nd Southeastern Workshop Mixed-Signal VLSI and Monolithic Sensors

2. University of Kentucky, Auburn UniversitySlide 2 Outline b b Project Goals b b System Architecture b b System Analysis b b Results of (sample) System Analysis b b Modularization of sensing solution b b Front-end Processing b b Back-end Processing b b Summary

3. University of Kentucky, Auburn UniversitySlide 3 Project Goals w w Develop a Chemical Sensing Framework for adaptability to a variety of low-cost or modular chemical sensing applications Characteristics – – Good reproducibility among batch -fabricated sensors – – High sensitivity through low noise transduction of sensory signal – – Reduction in communication bandwidth via local signal processing – – Resistance to drift via adaptable pattern recognition engine Phases – – Sensing Technology Development – – Sensory Plane Signal Processing Design and Implementation – – Base Station Processing and Interactive Feedback

4. University of Kentucky, Auburn UniversitySlide 4 System Architecture Sensing Nodes: Sensing Technology Large Arrays of Sensors Local Signal Processing Smart A/D Conversion Features Communicated over Standardized Protocol Base Station Processing stimulus

5. University of Kentucky, Auburn UniversitySlide 5 System Architecture Feedforward Objectives Raw Sensor Output Robust Aggregate Output Feature Extraction Conversion to Data Transfer Protocol Pattern Recognition and Interpretation

6. University of Kentucky, Auburn UniversitySlide 6 System Architecture Feedback Objectives Sensor Control Feedback Distribution of Parameters Conversion to Data Transfer Protocol Definition of Array Parameters and Constraints Conversion to Data Transfer Protocol Distribution of Parameters

7. University of Kentucky, Auburn UniversitySlide 7 Components of System Architecture Design b b Analyze the problem b b Determine a block-level solution b b Modularize the solution b b Establish communication protocols between modules b b Build and characterize modules b b Integrate modules b b Test System

8. University of Kentucky, Auburn UniversitySlide 8 Analyzing the Chemical Sensing Problem b b Arrays of discrete sensors (tin-oxide powder) b b Initial Data Collection wide range of array characteristics (temperature, dopant type) representative set of chemicals b b Use of science to determine initial set of features b b Clustering and Analysis of raw data b b Determination of optimal array size b b Principal Component Analysis of Steady-state features b b Principal Component Analysis of Temporal Features

9. University of Kentucky, Auburn UniversitySlide 9 Initial Architecture for Feature Analysis b b Size: 15 sensors b b Type: Tin-oxide powder TGS822: alcohol sensitivity TGS880: ammonia sensitivity TGS813: carbon monoxide sensitivity b b Array Dimensions Three types of sensors, Five operating temperatures Operating temperatures from (320-420 deg C) b b Six Representative Chemicals: Acetone, butanol, ethanol, methanol, propanol, xylene

10. University of Kentucky, Auburn UniversitySlide 10 Principal Component and Feature Analysis b b Raw data b b Steady-State Features median of array baseline-immune response saturated slope b b Temporal (transient) Features Time to threshold Mean of first d erivative Initial first derivative (beginning of transient response) Initial saturated output (end of transient response)

11. University of Kentucky, Auburn UniversitySlide 11 Results: Raw Data Sensor Circuit Repsonse, V Time, x 600ms

12. University of Kentucky, Auburn UniversitySlide 12 Results: Deep Saturation PC 1 For clarity, only acetone, methanol, and ethanol clusters are shown. One possible outlier is indicated by ANN Result: 95% correct discrimination Avg. of 100 trials Two different BP models Feature (d) PCA

13. University of Kentucky, Auburn UniversitySlide 13 Results: Initial Saturation ANN Result: 82% correct discrimination Avg. of 100 trials Two different BP models Feature (c) PCA

14. University of Kentucky, Auburn UniversitySlide 14 Results: Transient Slope ANN Result: 65% correct discrimination Avg. of 100 trials Two different BP models Feature (b) PCA

15. University of Kentucky, Auburn UniversitySlide 15 Results: Time-to-Threshold a? m? ANN Result: 42% correct discrimination Avg. of 100 trials Two different BP models Feature (a) PCA

16. University of Kentucky, Auburn UniversitySlide 16 Transient Results- New Data Feature: Initial slope Provides coarse distinction between “fast” and “slow responses, and some additional clustering. Potentially useful as one element of a hierarchical classifier.

17. University of Kentucky, Auburn UniversitySlide 17 Homogeneous Processing b b Averaging over sensors reduces sensor noise b b Averaging over time reduces ambient noise b b Example: Effect of averaging over time

18. University of Kentucky, Auburn UniversitySlide 18 Homogeneous Processing b b Effect of Averaging over Sensors 24 element array of TGS822 Tin-Oxide Sensors All sensors operate at same temperature in any given experiment Temperature is varied from experiment to experiment

19. University of Kentucky, Auburn UniversitySlide 19 Homogeneous Processing b b Effect of Averaging over Sensors 24 element array of TGS822 Tin-Oxide Sensors All sensors operate at same temperature in any given experiment Temperature is varied from experiment to experiment

20. University of Kentucky, Auburn UniversitySlide 20 Homogeneous Processing b b Circuits for Averaging over Sensors: Voltage Mode - + V in_n V out_n V bias - +V in_n V out_n V mean voltage averagingoutlier removal

21. University of Kentucky, Auburn UniversitySlide 21 Homogeneous Processing b b Circuits for Averaging over Sensors: Current Mode V in_n V mean current averagingoutlier removal Calculation of Outlier Current (Adjustable) V out_n

22. University of Kentucky, Auburn UniversitySlide 22 Heterogeneous Processing b b Heterogeneous Processing Extract features from sensor arrays consisting of: – –different types of sensors – –different operating conditions (temperature) Example: 15 sensor array – –3 types of sensors – –5 operating temperatures – –Extracted feature: median value of array – –Feature presentation: binary with respect to median

23. University of Kentucky, Auburn UniversitySlide 23 Heterogeneous Processing b b Heterogeneous Processing Circuits for extracting median from 15 sensor array

24. University of Kentucky, Auburn UniversitySlide 24 Heterogeneous Processing b b Heterogeneous Processing Experimental Results for median thresholding of array AcetoneEthanolMethanol

25. University of Kentucky, Auburn UniversitySlide 25 Summary b b Completed work analytically established features appropriate for extraction from arrays of metal-oxide chemical sensors proof-of-concept for homogeneous processing of such arrays CMOS circuits designed and fabricated for first stage of homogenous processing CMOS circuits designed for first stage of feature extraction b b Next Step additional homogeneous processing and feature extraction circuits repeat experiments for discrete, thin film, smaller sensors to establish benefits of miniaturizing long-term: extend to integrated, metal-oxide sensor arrays