© All Rights Reserved, Robi Polikar, Electrical and Computer Eng. Rowan University, Glassboro, NJ 08028 Did you ever measure a smell? Can you tell whether one smell is just twice strong as another? Can you measure the difference between one kind of smell and another? It is very obvious that we have very many different kinds of smells, all the way from the odor of violets and roses to asafetida. But until you can measure their likeness and differences, you can have no science of odor. If you are ambitious to find a new science, measure a smell. Alexander Graham Bell (1914) The Department of presents… Emerging Interdisciplinary Challenges Robi Polikar October 16, 2002 Outline Introduction: emerging interdisciplinary challenges Motivation and background The mammalian olfactory system vs. the electronic nose Commercially available electronic nose systems Quartz crystal microbalances Experimental setup Identification of volatile organic compounds (VOCs) An uncooperative database / sensitivity / selectivity issues Dealing with an uncooperative database Automated Identification Neural Networks Conclusions Questions, comments and suggestions Introduction: Emerging Interdisciplinary Challenges Organic Chemistry Olfactory Physiology Electronic Nose Signal Processing . . . . . Pattern Recognition Chemical Sensors / Analytical Chemistry Computational Learning Introduction Motivation & Background Many industries, institutions and organizations can benefit from a device capable of identifying odors: Food industries: detection of food quality / wholesomeness Airport security: drug smuggling, detection of explosives Anti-personnel land-mine detection Detection of household chemicals Detection of hazardous gases VX, CO, radon, etc Detection of volatile organic compounds Wastewater odor control Selectivity & Sensitivity Issues • Humans can identify 10000 types of odors at varying sensitivity levels. • 10000 odors are considered to be combination of a few basic types of odors: floral, musky, camphorous, pepperminty, ethereal, pungent (stinging), and putrid (rotten). • Another group of researchers believe that this number is actually around 50. • More recently, it has been suggested that there are actually over 1000 smell genes in the nose, each of which encodes a unique receptor protein. • Sensitivity: 5.83 mg/L of ethyl ether, 3.30 mg/L of chloroform, 0.0000004 mg/L of methyl mercaptan (1/25 trillionth of a gram) Mammalian Nose Vs. Electronic Nose Mammalian Nose Receptor neuron Odorant binding protein 10000000 receptors Sens. Electronic Nose Sensor / transducer Coating 6-30 sensors (array) Glomeruli Signal processing module Brain Pattern recognition module 1 part per trillion Selec. 10000~20000 odors 1 part per million <50 odors Electronic Nose Systems Odorant Molecules Chemical Information CIM Coated Sensor Raw electrical signal Sensor Processor Processed Electrical signal Signal Processor Denoised Electrical signal Feature Extractor Relevant features Knowledge Base Pattern Recognition Training & Testing Identification Sensor Technologies Metal Oxide Semiconductor sensors (MOS) Chemical Field Effect Transistors (ChemFET) Conducting Polymers (CP) Fiber Optical Sensors (FOS) Quartz Crystal Microbalances (QCM) Surface Acoustic Wave devices (SAW) Mass Spectrometry Gas Chromatography Pattern Recognition technologies Statistical pattern recognition (SPR) Bayes classifiers Discriminant analysis (DA) Maximum likelihood estimate Principal component analysis (PCA) Non-parametric techniques Artificial neural networks (ANN) Fuzzy logic (FL) Rule-based / expert systems Commercially Available Systems Electronic Nose System Manufacturer WMA Airsense Analysentechnik GmbH Fox x000 AlphaMOS AromaScan OsmeTech Inc. BH114 Bloodhound Sensors Inc. Cyranose 320 Cyrano Sciences Ltd. Enose 5000 Marconi Ltd. Znose 4200 Electronic Sensor Tech. QMB6 – HS40XL HKR Sensorsystem GmbH MOSES II Lennartz Electronik GmbH NST 3210 Nordic Sensor Technologies OligoSense OligoSense SAM Daimler RST Rostock SMart Nose 300 SMart Nose VOCmeter MoTech Sensorik, GmbH FreshSense Element Ltd. 4440B HP – Agilent Technologies VaporLab MicroSensor Inc. Libra Nose Technobiochip Sensor Technology # of Sensors Pattern Recognition Algorithms MOS 10 ANN, DC, PCA QCM, SAW, CP, MOS 6-24 ANN, DFA, PCA CP 48 ANN, FL CP 14 ANN, CA, DA, PCA ? U.K. CP 32 PCA 5,000 USA QCM, MOS, CP, SAW 8-28 ANN, DA, PCA ? U.K. SAW, GC 6-15 SPR 19,50025,000 USA QCM 6 ANN, PCA ? Germany QCM, MOS 16 ANN, PCA ? Germany MOS, FET, QCM 22 ANN, PCA 40,00060,000 Sweden CP ND/PR ND/PR ? Belgium QCM, SAW, MOS 6-10 ANN, PCA 50,000 Germany MS N/A DA, PCA ? Switzerland QCM, MOS 8 ANN, PCA ? Germany MOS ND/PR ND/PR ? Iceland MS N/A Various Chemometrics 79,900 USA SAW 2 ND/PR 5,000 USA QCM 8 ND/PR 5,000 Italy Price ($) 20,00043,000 20,000100,000 20,00075,000 Country of Origin Germany France U.K. Quartz Crystal Microbalances & Gas Sensing Bare piezoelectric crystal Central part of the crystal coated with first gold, and then polymer material Electrode on back Electrode on front Crystal holder W F 2.3 10 F A 6 2 Coating Selection Considerations For desired levels of selectivity and sensitivity • Thickness, softness / stiffness, reversibility, operation temperature Advantages Disadvantages Thickness sensitivity resistance, phase lag, attenuation Softness response time, reversibility Attenuation Stiffness Attenuation Reversibility Temperature Softness and hence response time sorption and hence sensitivity. • Viscoelastic properties: thermal expansion, swelling due to sorption, film resonance • Solubility parameters: coating – analyte interactions VOCs and Coatings Used Acetonitrile Acetone H O H H C C H H C H H Ethanol H H N C C H H CH3C N CH3CH2 OH Tricholoroethane Dicholoroethane O APZ Poly(isobutylene) PIB H C C OH H H Apiezon (grease, not a polymer) CH3CCH3 Methanol H H Cl H C OH H C H CH3CCl3 Methylethylketone O H C C C H H C H Octane H H H H H H H H Hexane H C C H Cl ClCH CCl2 Toluene CH3 C C C C C C H H H H H H H C6H14 Xylene CH3 CH3 C C C C C C C C H H H H H H H H H C8H18 Sol-gel SG Poly(siloxane) OV275 Poly (diphenoxylphosphorazene) PDPP H H H H H H Cl Cl H C2H5COCH3 H ClCH2CH2Cl Tricholoroethylene H C C H DEGA H H H Cl CH3 OH H H Cl Cl H C Cl Poly (diethyleneglycoladipate) C6H5CH3 C6H4 (CH3)2 • 12 individual VOCs at 7 different concentrations (84 patters). • 24 Binary Mixtures of VOCs at 16 different concentrations (384 patterns) Block Diagram of the Experimental Setup PERSONAL COMPUTER MASS FLOW CONTROLLER CARRIER GAS (NITROGEN) VOC BUBBLERS NETWORK ANALYZER MULTIPLEXER 3-WAY VALVE Bi - directional information flow Unidirectional information flow Gas flow SENSOR CELL 6 QCMs Experimental Setup Mass Flow Controller Network Analyzer PC Switching Box Nitrogen VOC VOC in bubbler Sensor Cell EXPERIMENTAL SETUP Mass Flow Controller Mass Flow Meter Network Analyzer Post-It notes Sensor Cell Gas Bubbler Switching Box How Does Odor Signal look Like? Problems With Identification Of Mixtures • Existence of dominant VOCs APZ: Apiezon, PIB: Polyisobutelene, DEGA:Poly(diethyleneglycoladipate), SG: Solgel, OV:Poly(siloxane), PDPP: Poly (diphenoxylphosphorazene) • Approach: Identify dominant VOC first, and identify secondary VOC based on the identification of the dominant VOC. Pattern Separability Issues Sensor 2 Measurements Sensor 2 Measurements (a) Sensor 1 Measurements Measurements from class 1 class2 (b) class 3 Sensor 1 Measurements class 4 (a) Well separated patterns and (b) densely packed / overlapping patterns Pattern (In)separability in Mixture VOC Problem ETHANOL TOLUENE XYLENE Sensor 3 OCTANE TCE Identification of VOCs Raw Sensor Readings (6-D) Preprocessing Increasing Pattern Separability Filtering, Normalization, De-trending, etc. Fuzzy nose (FNOSE), Feature range stretching, or Nonlinear cluster transformation Neural Network Training Multilayer perceptron LEARN++ (for incremental learning) Neural Network Validation Classification VOC Identification . . . . . Nonlinear Cluster Transformation Outlier Removal x x2 2 æ ö M1 çç å ( m i - m1)÷÷ - m1 è i ¹1 ø m 5-m 1 m5 m4 m3 x1 m2 m 4-m 1 x1 m 3-m 1 m 2-m 1 m1 Translation Cluster 1 S i -M i C x2 å (m j - mi ) C j 1 x2 Generalized regression neural networks Nonlinear Cluster Transformation x1 Similar to RBF networks Do not require iterative training Successful in multidimensional function approximation x1 PRINCIPLE COMPONENT ANALYSIS A Comparison XYLENE ETHANOL OCTANE TOLUENE TCE TOLUENE OCTANE XYLENE TCE ETHANOL . . . . . Artificial Neural Networks Signals From sensors (six) Output signal based on a weighted average of input signals Toluene Xylene The Multilayer Perceptron Neural Network . . . . . d input nodes x1 H hidden layer nodes x1 wJi J xd d æ ö ç f netJ å wJi xi ÷ ç ÷ i 1 è ø xd Wkj netj y j netk zk .. zc … x(d-1) Wji z1 .. …….... x2 ……... c output nodes æd ö ç y j f net j f å w ji xi ÷ ç ÷ è i 1 ø ( ) netj æH ö ç zk f (netk ) f å wkj y j ÷ ç ÷ j 1 è ø netk i=1,2,…d j=1,2,…,H k=1,2,…c Results Single VOC Identification 7 patterns obtained for each VOC, corresponding to seven different concentration values between 70 ppm and 700 ppm. Thirty (30) of the total 12*7=84 patterns were used to train the neural network. Remaining patterns were used to validate the performance of the network All 54 validation patterns were identified correctly ! Results Binary Mixture of VOCs TCE Mixtures TCE & TL 1 TL Mixtures 5 XL Mixtures OC Mixtures ET Mixtures TL & ACN 0 XL & ACN 0 OC & ACN 0 ET & ACN 0 TCE & MEK 1 TL & MEK 0 XL & ET 0 OC & MEK 0 ET & MEK 0 TCE & TCA 0 TL & HX 5 XL & HX 0 OC & TL1 3 ET & HX 0 TCE & HX 0 TL & ET 0 XL & MEK 0 OC & ET 0 ET & TCA 0 TCE & ET 1 TL & TCA 0 XL & TCA 0 OC & TCA 0 Dominant VOC Performance: 96% TCE Mixtures TCE & TL 1 TL Mixtures 0 XL Mixtures OC Mixtures ET Mixtures TL & ACN 0 XL & ACN 0 OC & CAN 0 ET & ACN 0 TCE & MEK 0 TL & MEK 2 XL & ET 2 OC & MEK 0 ET & MEK 1 TCE & TCA 0 TL & HX 3 XL & HX 3 OC & TL2 0 ET & HX 0 TCE & HX 0 TL & ET 0 XL & MEK 1 OC & ET 3 ET & TCA 0 TCE & ET 0 TL & TCA 0 XL & TCA 1 OC & TCA 0 Secondary VOC Performance: 96% 196 (50%) patterns used for training and remaining 196 used for testing. Conclusions QCM technology along with neural network identification can be used as an efficient tool for electronic nose applications Challenges: Identification of components in mixtures Identification of gases at very low concentrations (ppb levels ?) Adverse environmental conditions (temperature, humidity, etc.) New sensor technologies for improved sensitivity and selectivity Incremental learning of additional odorant (Algorithm: Learn++) Questions