Concept of an electronic nose
The concept of an electronic nose originated in the seventies. Up to then, analytical chemistry had been pre-occupied with developing highly specific sensors and methods, aimed at identifying unique
The new availability of personal computing made it possible to apply pattern recognition techniques to complex measurement data.An important consequence of the concept is that a substance, or mixtures of substances, can only be recognized after a calibration phase: in order to match a pattern, it must be known beforehand (‘seen’ before) substances.
The proposal was to have a general, broadly responsive sensor system generating complex multi-dimensional measurement data and use pattern recognition techniques to match measured ‘patterns’ to previously ‘seen’ patterns.This is analogous to how we smell, hence the name ‘electronic nose’. This is illustrated in the figure.
Electronic nose sensor types
The basic concept of an electronic nose, or machine olfaction, is a measurement unit that generates complex multi-dimensional data for each measurement combined with a pattern recognition technique that interprets the complex data and relates it to a target value or class.
In academic literature systems based on (for example) a mass-spectrometer in combination with pattern recognition are sometimes presented as an ‘electronic nose’ application or artificial olfaction. However, in this section only relatively low-cost sensor technologies are discussed which are in principle suitable for bench-top or portable devices are discussed.
The requirement that a multi-dimensional measurement signal is generated excludes single detection elements used for example PID meters. This is often overcome by using an array of broadly sensitive elements with different sensitivities to important chemical compounds. As an electronic nose device is frequently exposed to volatile chemicals arrays of potentiometric sensors are not useable.
The latter type are aimed at leak detection because they have a limited amount of reactive chemicals (the working principle is similar to a chemical battery) which is depleted when exposed to the target substances.
The technologies which are feasible for application are QMB/SAW, conducting polymers and metal-oxide sensors.
A QMB is a quartz crystal with a chemically active surface, usually a polymer. When gas molecules adsorb to the surface, the mass changes and the resonant frequency of the crystal shifts. These minute shifts need to be measured with high frequency electronics (complex, expensive).
Small temperature variations result in similar frequency shifts thus dictating strict environmental temperature control. A variation of a QMB is a SAW (surface acoustic wave) sensor which also works on the principle of frequency shifts.
Conducting polymers are polymers which are either intrisically conducting or non-conducting types which have been ‘loaded’ with graphite. In the former type the conductivity may alter when exposed to volatiles. In the latter case the graphite provides an electrical resistance path which can be measured very easily. When gas molecules associate with the polymer, it will swell thus breaking contact points between graphite particles and thus changes the resistance.
In this case also, temperature changes will result in expansion/contraction and thus in resistance changes, therefore it is also advisable to apply strict environmental temperature control for this sensor type.
Although the possible variations in polymers is enormous (and thus the variations in arrays also), they are chemically not very stable. Strong oxidizers such as chlorine and ozone can fairly easily disrupt a polymer.
Metal Oxide sensors
The basic choice of sensor is a so-called micro hotplate Metal-oxide sensor (MOS).
Certain metal-oxides behave as semiconductors at higher temperatures. Sensors based on this are designed as having a heater element and an sensor element (sintered metal-oxide with or without catalyst). Both elements are separated by a very thin isolating membrane.
Redox-reactions occurring at the sensor surface result in changes in resistance which can be measured. These redox-reactions depend on the nature of the metal-oxide/catalyst, the reacting gas(es), and the temperature. A minimum of 0.1% of ambient oxygen is required for normal operation.
Depending on sensor type and temperature, a very broad range of substance will give a redox reaction. Notable exceptions are N2, CO2 (will not oxidize further) and noble gases such as Helium and Argon.
Since the 1980’s there has been a commercial market for electronic nose devices, albeit a niche market. The current market leader is the French company Alpha Mos which targets the very expensive, very high-performance laboratory equipment market.
The relatively cheaper and more portable devices of Scensive Ltd (Bloodhound) and Smiths Detection (Cyranose models) frequently reported in academic literature are both based on conducting polymer arrays. All of the current devices are intended to strictly be used as laboratory instruments analogous to HPLC/GC’s and spectrometers. All of these devices need to be individually calibrated for a particular application.
It is not possible to develop an application using one device and then apply that model to another device of the same type without using actual calibration measurements with the same type of samples.
The approach of the eNose company
The philosophy of the eNose company is to not compete with the existing specialist laboratory market, but to address mass-markets.
This is the currently non-existing segment of low-cost, low power, mass-producible and mass-employable electronic noses. The target is to have a tiny, battery operated electronic nose with consistent unit-to-unit properties thus enabling ‘calibrate on one unit, apply to all units’ methods. The latter is required to make mass-production and mass-employability possible.
For example, a feasible goal is to have a device small enough to incorporate in a mobile phone and distribute new ‘smelltones’ (downloadable calibration models) which can be independently developed without requiring any form of chemical calibration of already sold mobile phones.
eNose core module
The workhorse and basic core of our eNose application devices is essentially a stand-alone module in its smallest form. This module measures 13.5×41.5 mm. The module contains a sensor with the supporting temperature control and measuring electronics, a microcontroller and ram and flash memory. The module also has a silicon serial number chip with a unique identifier that allows tracking and management of each manufactured module.
The microcontroller performs the temperature control and calibration of the sensor heater, measures the thermal cycles according to the settings, and buffers measurements until they can be unloaded to a central data store. Numerous settings are stored persistently in the onboard ferro-static ram (FRAM). The module communicates over a serial bus system with a host.
As of yet there are 5 chemically different sensor types that can be mounted on a module, giving five subtypes with the same electronics and embedded software.
As explained in the previous section, the redox reactions are temperature dependent. This allows the generation of much more specific patterns by a technique called temperature cycling. This is depicted in the left figure. In this example a sinusoidal temperature cycle is depicted, with the resulting sensor responses as function of the temperature variation in time. The time axis is the same for the upper and lower parts of the figure.Different substances exhibit strong responses at differing temperatures for the same chemical sensor type.
By taking the response of a complete cycle and plotting the response as function of the temperature the patterns are obtained. This is illustrated in the right-hand figure.
The elegance of this approach is that one acquires the relative temperature dependent redox reactions on a single sensor. The relative information is not device dependent as it is with arrays of sensors used at a static temperature. The crux of the pattern recognition problem is to get the actual temperatures right in all units/sensors. The variation in heater element production (even using MEMS fabrication) can result in temperature variations in the order of ±50 °C between two sensors of the same type.
The eNose company has developed a patented unique automatic temperature calibration technique. With this technique the standard deviation in inter-unit temperature variation is slightly less than 1°C. This is a requirement to allow calibration models developed on one unit to be used on another ‘as is’ without alteration or chemical tuning.
Data analysis is a crucial factor in any electronic nose device. The complex nature of the measurements combined with timing effects due to sampling mechanisms (such as exposure/recovery dynamics) creates multi-way data. When the thermal cycling technique of 2.3 is applied, each sensor will generate a 3 dimensional matrix of (exposure) time versus thermal cycle versus response value. When a combination of chemically different sensors is used to measure a sample, the chemistry of the sensor forms a fourth dimension in the multi-way data. This is illustrated in the left part of the figure below for a system with three sensors A,B and C.
Classical multi-variate pattern recognition techniques require that a single measured sample is represented as a 1-dimensional vector giving a 2-dimensional matrix for the data set as a whole. One dimension of the matrix represents the individual measurements, the other the elements of the vector belonging to each measurement. This is illustrated in the right part of the figure in which each row represents a single sample.
A naïve unfolding of the 3-way or 4-way data in order to derive a 1-dimensional vector is not feasible as the number of elements of the resulting vector would exceed the number of measurements/samples by orders of magnitude leading to unresolvable over-fitting. In many cases researchers by necessity resort to using only a very small subsection or even a single point of the measurement data and thus waste most of the available information. Thus a form of pre-processing is needed which generates 1-dimensional vectors for each sample while retaining the essential information needed for the application.
The eNose company has developed a highly sophisticated hybrid TUCKER3/PARAFAC algorithm that is tailored to this particular type of multi-way data. With this algorithm the full information contained in the measured data is preserved when generating 1-dimensional vectors for each measurement while at the same time elegantly removing redundancy, noise and scaling information.
The patterns obtained by thermal cycling are not only dependent on the temperatures applied, but also on the dynamics of the temperature. This is illustrated in the figures below. Four substances were measured at several concentrations with the same sensor/hardware setup.
In the left figure a slow ramping temperature cycle was used, in the figure on the right a series of much faster pulse cycles. As is illustrated, the relative sensitivity for methylamine is increased dramatically (more than 1000 fold) by the change in temperature dynamics.
The implication is that the sensitivity of a sensor can be enhanced by selection of the correct dynamics in software without any change to the hardware or sensor.