Electronic Nose (Sensor-Based Odour Analysis System) for Odour Identification and Measurement | ARDI | Agri-Food Research & Development Initiative | Manitoba Agriculture, Food and Rural Initiatives

Increasing hog production and our capabilities for value-added processing continues to be a priority for Manitoba agriculture. This increase is also needed to fully utilize the new processing facilities which have been constructed and which promise to provide a major source of ongoing revenue and employment. However, this increasing production has met with concern over its environmental impact – with one of the major concerns being the odour produced by large-scale hog operations and the effect on nearby residents.

Measuring odour is a challenge to researchers and regulatory agencies. The electronic nose (e-nose) potentially allows the measurement of both quality and quantity of odour. In the simplest terms, the electronic nose is a sensor array which responds to volatiles in much the same way as the human nose (hence the name). Volatile compounds from a source (the sample) pass over the sensors and cause a change in the electrical response of each individual sensor (just as do the receptors in the human nose). These changes form a pattern that is recorded in a data file and a label assigned to the pattern (equivalent to the area of the human brain where odours are processed). This pattern is used to compare to other patterns to decide if the odour is the same or different. The machine responses can be compared to human responses to the same samples and machine criteria set to enable the electronic nose to replace the human for a large range of uses.

Electronic nose analyses have moved into general usage in quality control of food products in other countries, but its use in Canada has been limited to date. One reason is that very few e-noses exist in Canada and consequently there has been little local demonstration of the uses of the electronic nose technology. The strength of the use of the e-nose technology is the ability of these systems to reproduce decisions based on sensory assessments of the products. Unlike other systems such as GC/MS analysis which break down the odour samples to individual components, the electronic nose systems evaluate the overall change in the pattern of volatiles from a sample and use this pattern to relate to other known patterns to evaluate the detectable differences among samples.

An Alpha MOS Fox 3000 electronic nose was purchased by the University of Manitoba in early 2001 and commissioned for use in July, 2001. This grant provided funding for opportunities to demonstrate the applicability of this technology to Manitoba livestock and food industries. There were two primary objectives for the current work:

OBJECTIVE 1 – Installation, Commissioning and Establishing Operating Procedures

An Alpha MOS Fox 3000 electronic nose was received on May 8, 2001 by the Department of Biosystems Engineering at the Canadian Food Inspection Agency’s Sensory Science Laboratory at the Freshwater Institute. The unit consists of five basic components:

The actual installation and preliminary training session occurred in July 2001. A representative from Alpha MOS France spent three days in the laboratory working through both the actual set-up and standardization of the equipment and presenting sessions on the operation of the equipment, and the AlphaSoft operating system. A “start-up kit” was purchased from London Scientific Ltd. for use in this process. The kit includes the small pieces of equipment needed and consumables necessary for installation and preliminary testing: 1000 10-mL vials, magnetic caps with silicon septa, hand crimper for sealing caps on vials, injection port septa, and one chemical kit for calibration and diagnostic testing.

A series of studies were conducted for a range of samples and testing conditions to demonstrate the feasibilities of the e-nose technologies in odour measurement and product quality identification. These studies are listed as follows:

Each feasibility study involved two stages: (1) the optimization of sampling conditions, and (2) the evaluation of known different samples to establish the analytical performance of the e-nose for the product. In optimizing sampling conditions, the following two variables were established: (i) the incubation time and temperature which were used to facilitate the release of odours into the headspace in the sample vial; and (ii) the amount of sample which was withdrawn and injected into the sensor array. There were several other variables which could be changed as well, but for practical purposes, they were generally held constant unless there was some specific reason to change them. The variables were selected based on the nature of the product being analyzed and the expected intensity of the odours that would be produced. This information had to be generated for every different product to be tested. Once these conditions have been standardized, data collection can continue in a given study.

Each of these projects used the basic multivariate statistical analysis methods which are part of the AlphaSoft Version 8.0 software. The proprietary software controls the collection of the responses from the 12 sensors in the sensor array and formats them for data analysis using the multivariate methods. The data analysis subroutines and their uses are:

Principal Component Analysis (PCA) – is used as the first analysis to examine the data for outliers and to extract information on the ability of the sensor array to differentiate among the samples. The discrimination index, which is also calculated here, gives an indication of the separation of the data between groups. A positive value indicates distinct groups and a negative value indicates overlapping groups. A value between 80 and 100 (the maximum) indicates the most useful separation of the samples.

Soft Independent Modelling of Class Analogy (SIMCA) – is used to differentiate between a standard sample (referred to as “good”) and any other form of the product (referred to as “bad”) and is useful as a quality control tool. The desired end product is designated “good” and any other version of the product is “bad” (e.g., missed ingredient in formulation) to ensure that only the appropriate product is used and any problem lots are identified for further investigation. The effectiveness of this method is expressed as the percentage of recognition of defective samples. The model is considered to be valid if the percentage is greater than 90%.

Discriminant Factor Analysis (DFA) – is used to construct a model to allow a new sample to be identified as one of several groups. These groups can be different quality levels, intensities of a particular compound, or different product origins or identities. This method requires the use of sensor optimization for statistical discrimination and a minimum number of samples per group.

Partial Least Squares (PLS) – is available in AlphaSoft in two formats: Concentration Quantification (log transformed data – with “0” and negative values not allowed) and Sensory Score Correlation (linear data – with “0” and negative values allowed). This method is used to predict quantitative values based on a calibration curve calculated using correlation methodology between quantitative data (concentrations, panel scores, etc.) with the sensor responses.

Statistical Quality Control (SQC) – is available in AlphaSoft in two formats: Gold Standard (reference samples define two limits, upper and lower, of the sample range) and Odour Intensity with Threshold (define 1 upper limit of acceptability for the sample). These methods are intended for direct application to industrial situations – to predict when samples from production lots are in or out of process control.

The evaluation of samples and the conditions for evaluation would also usually involve the use of the “sensor optimization” subroutine which selects the sub-set of sensors that are most important in evaluating the particular product. This is then used to specify conditions for on-going quality control analyses. Examples of the use of some of these methods are included in the individual studies reported. PCA analysis is included for all of them, as it is the first analysis done and establishes the ability of the system to differentiate among the samples presented to the sensor array.

AlphaSoft also allows for the raw data to be extracted from the Multivariate Analysis subroutines and placed into standard spreadsheet programs. It can then be used for analysis through other analysis packages, such as Excel, SAS, SPSS, or artificial neural network programs.

OBJECTIVE 1 – Installation, Commissioning and Establishing Operating Procedures

The metal oxide sensors are sensitive to air environment. Much effort was devoted to searching for the supply of high purity gas mixtures and testing the suitability of these gas mixtures. It was found that a special gas mixture of 20% O₂ ± 1% with the remainder being N₂would be needed. Praxair of Winnipeg was identified as the gas supply. It was found that a mixture of 19.8% O₂ proved to provide a stable air environment for the sensor array.

The diagnostic tests were conducted for ongoing monitoring of the status of the sensor array relative to the initial state of the system. Diagnostic testing must be done weekly using standard conditions specified by the manufacturer. At the end of each session, the computer analysis compares the test to the baseline transferability to the computer model has been built through the ongoing diagnostic testing. Sampling runs were done with a variety of products in order to evaluate the effects of different incubation times/temperatures in the auto-sampler and sample volumes injected into the sensor array. Data which have been collected can be used for comparisons with other systems and within this system.

Procedures were established for the development of test conditions required for particular samples. The process involved the evaluation of incubation time and temperature conditions for optimum generation of headspace volatiles. This was followed by testing of selected sample sets to demonstrate the unit's ability to distinguish among known samples.

Data analysis procedures were established based on multivariate methodologies within the machine software. A procedure was also established for transferring data to other programs (Excel, SAS, SPSS) for further analyses.

Sampling methodologies were developed for two specific applications – “saline wash” for evaluation of fish and fish products, and cloth swatches for evaluation of environmental malodours.

The e-nose was configured by the manufacturer to take headspace samples from the auto-sampler. The auto-sampler mode takes samples contained in 10-mL vials and allows analysis of liquid and solid samples. To measure gaseous odour samples collected in Tedlar bags, the unit must have the airflow pathway changed internally for the sample delivery to the sensor array. Sampling mode switching involved reconfiguration of air flow paths and re-calibration of air flow rate. A procedure was established for switching the sampling modes.

Communication with the manufacturer was part of the ongoing developmental process for using the e-nose technology. An extremely important progress has been that the pathways for communication and support from Alpha MOS were established with their technical support group. This has proved very valuable for on-going monitoring of the system, for problem solving, and for evaluation of potential sampling conditions. Part of this is the capability of sending computer files to the manufacturing location in Toulousse, France through e-mail and the capability of receiving new file information from them for installation into the custom software for the system.

During the commissioning process, common commercial soft drink products were tested as a set of uniform samples which required no extensive preparation. Using highly standardized known commercial samples for initial testing is common in establishing data handling methodologies for electronic nose systems. The tested samples included fresh Coca Cola (COC), 6 months old Coca Cola (CO6), 6 months old Diet Coke (COD), Pepsi Cola (PEP), Barq’s Root Beer (BAR), and Ginger Ale (GIN). The first step in the data analysis process was the examination of the data map through principal component analysis (PCA). The discrimination index (DI) indicates how well the e-nose was able to discriminate among the samples. The values of the discrimination index ranges from 0 to 100, with a value >80 considered to be good discrimination.

The resulting PCA map shows that the e-nose was able to discriminate among all of the samples tested to a high degree (DI = 94%, Figure 1). One interesting note is that the machine discriminated between the recently purchased and the 6-month old Classic Coca Cola and the 6-month old diet Coke.

The same data set was analyzed using the “gold standard (SQC)’ sub-routine which is part of the AlphaSoft to demonstrate another use of the data – as a quality control tool. The samples identified in the “blue band” are the product standard and other samples are compared to this standard (Figure 2). In a quality control situation, the samples outside the standard area would be considered to be defects. The example includes the cola beverage standard in the “in process control” area and the two samples outside the area are the same beverage stored 6 months and its “diet” formulation (Figure 2).

Odourous gas (air) from livestock operations contains many chemical compounds (odourants) resulting from microbial degradation of manure and other by-products. The odour intensity (or offensiveness) depends on the concentration of these compounds as well as the combination of these compounds. There exists no single analytical instrument for the direct measurement of odour sensation. The most widely accepted method is olfactometry evaluation by using the dynamic-dilution olfactometer. To use a dynamic-dilution olfactometer, odourous air samples have to be brought into the laboratory from the measuring sites by using impermeable bags, which are usually costly. Cloth swatch adsorption has been considered and tested by several researchers as an inexpensive alternative for odour measurement. When a cloth swatch is exposed to odourous gas (air), odourants are adsorbed on the fabric, or dust particles that carry odourants are deposited on the fabric. The odourant laden swatches are then assessed by a human sensory panel to determine the odour intensity of air that the swatches had been exposed to. However, odour contained on a swatch is affected not only by the concentration and types of odourants in the air, but also by other variables. This particular study involved evaluating the uptake of swine manure odour onto cloth swatches. A swine odour simulant compounded from standard chemicals was used to provide a constant odour source (Table 1). This allowed for evaluation of the samples with the electronic nose and for comparison to sensory panel evaluations of the same samples.

Component	Concentration, ppm
acetic acid	1500
propanoic acid (propionic acid)	375
2-methyl propanoic acid (isobutryic acid)	600
butanoic acid (n-butyric acid)	250
3-methyl butanoic acid (isovaleric acid)	300
Pentanoic acid (N-valeric acid)	100
Phenol	25
4-methyl phenol (p-cresol)	75
indole	7.5
3-methyl indole (skatole)	5
ammonium hydroxide (NH4OH) 5%stock solution	adjust to pH 8.2

Thirteen types of fabrics were tested as odour adsorbing swatch material. Swatches were exposed to odour through a sampling set up shown in Figure 3. Exposed swatches were then placed into vials for e-nose measurement. Preliminary tests were conducted to establish test conditions as follows:

PCA analysis was performed to compare 13 different fabrics for their ability in adsorbing swine odour. The data map shows that the Fox 3000 e-nose was able to distinguish between control (non-exposed) and exposed swatches – by the general grouping of the treatments (Figure 4). The responses were dependent on the fibre type. Figure 5 compares non-exposed and exposed swatches for cellulosic fibres only (cotton, linen and rayon). The difference in the exposed and unexposed groups was clear as was the effectiveness of the e-nose in being able to detect swine odour.

This work was done as part of a larger study by Dr. D. Abramson of the Cereal Research Centre. E-noses analysis was performed on Durum wheat samples which were stored over 28 weeks. Durum wheat at 19% moisture (high moisture) was held for 20 weeks and samples were taken at 4 week intervals. There was a clear difference between the day 0 samples and all of the others (Figure 6). There was a clear change in the samples by week 4. This definite change at 4 weeks corresponded to the rise in level of mycotoxin in the grain. The Fox 3000 was able to differentiate the quality of the wheat as it changed – and this change correlated highly with the results of GC-MS analyses which were conducted as part of the larger study.

The Durum wheat samples containing 15% moisture (low moisture level) showed no development of mycotoxins in storage. This corresponded to the finding of no difference among all of the samples – including the 0-week control and the 4-week samples and the others (Figure 7).

A small feasibility study was done jointly with R. McQueen and L. Lamari, University of Manitoba to evaluate the possibility of using the e-nose for the testing of potato tubers for the presence of spoilage conditions during storage. The spoilage conditions tested were soft rot, late blight (at slight and heavy stages of spoilage) and deep rot conditions. Cloth swatches (standardized cotton flannel) were exposed under controlled conditions to different potato spoilage sources. The samples were: unexposed control (COT), healthy tuber (POTCLR), late blight (low level) (POTLBL), late blight (advanced level) (POTLBH), soft rot (POTROT) and deep rot (POTDPR). The principal component analysis (PCA) maps of the samples show that the e-nose was able to detect differences, with the distances between the samples on the map indicating the relative degree of difference of the odours analyzed (Figure 8). There were two distinct groupings – unexposed and exposed to healthy tubers, and exposed to diseased tubers. There was a clear difference between the samples derived from the diseased tubers and the healthy ones, while there was no difference between the healthy samples and the unexposed cloth swatches. There was a difference between the control, healthy tuber and early level of late blight (shown by the high discrimination index), but in this sample set they were in close proximity on the PCA map. The high discrimination index indicates that this method of sampling allowed clear differentiation among all of the samples.

A series of studies was conducted (through funding from the Canadian Food Inspection Agency and Environment Canada) using various species of freshwater and marine species to determine the effectiveness of the Fox 3000 e-nose in discriminating quality levels, presence of contaminants and identity of fish species. Two graphs are included to demonstrate the results of these studies. Figure 9 shows the data from Whitefish fillets (fresh) incubated at 60°C compared to scores obtained from an expert product assessor in a laboratory testing situation. The correlation was obtained using the Sensory Score Correlation subroutine to correlate the sensory score for each spoilage increment to the sensor responses for that sample (3 reps for each of 0,2,4,6,8 and 10 days on ice). The correlation between the sensor responses and the expert analyst was found be very high with r = 0.986. Figure 10 shows a PCA map for examining the data for the ability of the Fox 3000 e-nose to differentiate among fish species. Fish flesh samples from 5 species of fish, 3 freshwater and 2 marine species tested using 60°C incubation for 2 minutes. The e-nose was able to distinguish among the different species included in this sample set.

Proposed regulatory changes under Food and Drug Act and Regulations (Health Canada) will allow the irradiation of four new commodities – ground beef (fresh and frozen), poultry (fresh, frozen and ground), shrimp and prawns (prepackaged fresh, frozen, prepared and dried) and mangoes. An important aspect of allowing a process is the ability to measure its presence. As irradiation processing, even at low levels, can cause flavor changes, an electronic nose system is a logical analytical tool to detect the process. In a small feasibility study in collaboration with Dr. R. Holley, University of Manitoba, the Fox 3000 e-nose was used to evaluate frozen ground beef, in both aerobic and anaerobic packaging for the presence of irradiation processing to the level allowed in the Health Canada regulations. The resulting PCA maps are shown in Figure 11 and Figure 12. PCA maps generated from data show that the e-nose was able to differentiate between the untreated control and the treated ground beef – both fresh and frozen when processed with electron beam irradiation.

A feasibility study was conducted with the aid of samples supplied by Bee-Maid Honey, Winnipeg, MB. First, a preliminary test was conducted to establish sampling conditions followed by an actual test run of the different source samples. It was found that an incubation temperature of 30°C was appropriate. The sources of samples tested were: liquid honey (LH), which was Bee-Maid honey purchased from a commercial outlet (No. 1 White, pasteurized), processed honey (not pasteurized) (PH), clover (80%, raw) (CLO), canola (95%, raw) (CAN), sunflower (20%, raw – would not be processed as greater than 5% of another source) (SUN), buckwheat (20%, raw – would not be processed as greater than 5% of another source) (BUC), and a blend of 23% mixture of sunflower, clover, and 77% canola (MIX). The PCA map shows that all samples were able to be differentiated by the Fox 3000 e-nose (Figure 13).

The Alpha MOS Fox 3000 electronic nose is operational and its use for the successful collection of reliable data over a wide range of product has been demonstrated. Specifically, the n-nose was capable of identifying artificial swine odour adsorbed on cloth swatches. Spoilage of grain and potatoes could also be detected by the Alpha MOS Fox 3000 e-nose. The e-nose was able to distinguish among the five different fish species included in this study, between irradiated and non-irradiated ground beef, and among seven sources of honey samples. The e-nose is now available to the Manitoba agriculture and agri-food industry for the development of quality control applications.

This project was made possible due to funding from the Governments of Manitoba and Canada through the Canada-Manitoba Agri-Food Research and Development Initiative (ARDI).