The novel EOS835 electronic nose and data analysis for evaluating coffee ripening

https://doi.org/10.1016/j.snb.2005.01.019Get rights and content

Abstract

We present an analysis of roasted coffee ripening performed by the novel Electronic Olfactory System EOS835, manufactured by the Italian company Sacmi Imola s.c.a.r.l., which is based on thin film semiconductor metal oxide gas sensors. We focused our analysis on: (1) exploratory data analysis for systematically investigating the outcomes of different sampling conditions and therefore selecting advantageous settings; (2) feature selection for improving classification performance and ranking the contribution of the different sensors and feature types. Exploratory analysis, via the successive generation of PCA plots, showed that the main factors influencing discrimination between diverse ripening times are headspace generation time (HGT, i.e. time elapsed between vial filling and measurement performing) and sample preparation. A relatively long HGT (18 h) allows to follow the ripening progression of the coffee blend over time and to correctly classify the best coffee ripening (as determined by an expert taster). In forming the feature vector, we added a feature calculated in the phase space to the standard features. Feature selection showed that, the phase space feature consistently lead to improved classification and that, of the three sensor types constituting the array, the two indium–tin oxide sensors perform better for our application.

Introduction

Smart systems for odour recognition appeared on the market ten years ago (MOSES I, 1993) due to a fruitful collaboration between industry and scientific world. After that, many commercial Electronic Noses (EN) have been put on the market; this type of technology is now widely employed in different fields: automotive, environmental monitoring, medical diagnostic, food processing. A recent and detailed review can be found e.g. in Ref. [1].

Routine analysis in food quality control is one of the most promising applications of EN [2], [3]. In fact, in standard routine controls a full chemical analysis of the product is often not necessary. Instead other factors should be considered: ease of use, low cost, high speed (few minutes), good correlation with sensory panel assessments, reliability and reproducibility. Some of the above requests are satisfied by EN technology, while great efforts have still to be made in order to improve long-term reproducibility and system stability towards changing environmental conditions.

In particular, the coffee production chain has been widely investigated by ENs. The EN has been used to distinguish different types of coffee beans [4], to identify various brands and mixtures [5], to classify commercial coffee blends [6] and samples with different roasting levels [7]. In such cases, a very good sensitivity to the coffee aroma and satisfying time stability has been observed for semiconductor metal oxide (SMO) sensors. We also experienced this result. In previous works the Pico EN developed at the Sensor Lab, equipped with thin film SMO sensors [8], has been used to classify different monocultivar coffees and commercial coffee blends [9], [10]. The results were encouraging: the classification rate was above 90% and the human hedonic index was also correctly predicted by the EN. This good correlation between human and EN judgement is an important finding, because the coffee quality assessment is usually performed by an expert taster or by a trained panel. However, an expert taster is not very objective, because it is due to a single judgement, and the panel test has also many drawbacks: recruitment and training, people availability, time consume, high costs.

We used the EOS835 for the determination of the ripening level of a roasted coffee blend inside the production chain of an Italian company. The EOS835 electronic nose has been manufactured by the Italian company Sacmi s.c.a.r.l. The EOS835 is an advanced, engineered version of the Pico EN previously (1998–2002) developed at the Sensor Lab.

The paper concentrates on: (1) exploratory data analysis for the optimisation of the headspace sampling protocol and (2) the selection and ranking of the best sensors and features for the ripening levels discrimination. The applicative long-term goal is to design a robust system that monitors the coffee blend during the ripening. The system should be able to predict the right aroma intensity, established for example by a human panel, in order to stop the seasoning of coffee and to put the product onto the packaging line.

In order to reach the best system performance, the optimisation of the sampling conditions is of paramount importance [11]. The relevant sampling parameters include: headspace technique (e.g. static, dynamic, purge and trap, solid phase microextraction), amount of sample and vial volume, thermostatting temperature of the sample and equilibration time for headspace generation [12], [13]. Other important environmental variables (such as temperature, relative humidity, pressure, flow) should be also monitored and controlled.

In order to evaluate the influence of the various sampling parameters on the EN responses, it is useful to visually explore the data. In fact, the initial examination of the data is one of the most important parts of the data analysis cycle [14]. This aspect of data investigation should ideally be performed iteratively together with data collection. Termed initial or exploratory data analysis it constitutes the first phase of the analysis and comprises three parts: checking the quality of the data; calculating summary statistics and finally producing plots of the data in order to get a feel of their structure. Histograms, scatter plots and principal component analysis (PCA) are the default methods for graphical data representation.

In this paper we make systematic use of PCA for detecting clusters in the data, indicating important variables and detecting outliers. To this end, we produce a sequence of PCA plots either by using diverse data labellings (where the labels correspond to the variables of interest) or by taking data subsets which correspond to well defined target values. A similar approach has been used for the exploratory analysis of a previous EN coffee dataset in Ref. [10].

Diverse features have been extracted from the sensor response curve in the literature. In the classical time domain, feature extraction can be performed either considering the steady state part of the signal or the transient response [15], [16], [17]. Recently, feature extraction from the phase space (i.e. the space formed by the time response and its first derivative) has been introduced [18]. This feature could be interesting because it takes into account both static and dynamic information at the same time. In this paper, for every sensor, we extracted the classical relative response R/R0, the relative integral and the phase space integral.

If a large number of different features is extracted from the sensor signal, a selection of relevant variables is necessary in order to improve classification and to evaluate the sensor and feature importance [19]. Feature selection requires two fundamental elements: a selection criterion (e.g. the classification error evaluation) and a search algorithm. The research procedure can be exhaustive, i.e. through all the possible subsets, or sub-optimal [20]. We here performed exhaustive feature selection with k-NN to rank features and sensors according to their classification rate.

Section snippets

Experimental method

Commercial coffees are usually blends, i.e. mixtures of different (monovarietal) coffees. In our case, the blend is made by 12 different types of monocultivar Arabica coffee. After the roasting and before the packaging step, the coffee grains are stored in 250 kg batches within suitable silos to undergo a process named blend ripening or seasoning.

During this time (usually not more than 1 week) the coffee blend modifies its aroma. In order to monitor the coffee quality, an expert coffee taster

Results and discussion

The expert taster assessment is that the coffee quality increases during the ripening process and that the best quality corresponds to the sample with a ripening time of 96 h. The two indexes “positive odors” and “olfactory intensity” grow with increasing ripening time and reach the maximum value after 96 h indeed.

The aim of this study is therefore to discriminate between the ripening levels and in particular to single out the 96 h ripening level which is the best. In the following, we will

Conclusions

The results presented show that the EN, after optimization of the sampling parameters and suitable data processing, can be used to monitor the coffee blend during the seasoning process for evaluating the optimal ripening time.

Not surprisingly, we observed that the measurement outcomes are strongly influenced by the sampling parameters. We showed that simple PCA plots are sufficient to interpret the changes in the measurements that reflect the different experimental conditions and therefore to

Acknowledgements

This work has been financially supported by the FIRB project no. RBNE019TMF. We would like to thank the anonymous reviewer for the thorough and stimulating review, which we think enhanced the quality of the paper.

Matteo Falasconi got his degree in Physics (with honours) from University of Pavia in 2000 defending a thesis on non-linear optical properties of porous silicon. In 2001, he obtained a PhD scholarship in Material Engineering from University of Brescia. He is involved in functional characterization of thin-film metal oxide gas sensors and in the development of an innovative electronic nose for food quality control.

Cited by (0)

Matteo Falasconi got his degree in Physics (with honours) from University of Pavia in 2000 defending a thesis on non-linear optical properties of porous silicon. In 2001, he obtained a PhD scholarship in Material Engineering from University of Brescia. He is involved in functional characterization of thin-film metal oxide gas sensors and in the development of an innovative electronic nose for food quality control.

Matteo Pardo got a degree in Physics (summa cum laude) in 1996 with a thesis in theoretical surface physics at the University of Milano. In March 2000 he obtained the PhD in Computer Engineering with a dissertation on Multivariate Data Analysis for Gas Sensor Arrays. Since 2002 he is a researcher of the National Institute for Matter Physics (INFM), now part of the Italian National Research Council (CNR). His research interest is data analysis and in particular the applications of machine learning and pattern recognition techniques for the analysis of chemical sensor arrays and, recently, DNA microchips data. He was an invited lecturer at three international conferences and co-director of the Short Course on Fundamentals of signal and data processing for the 2nd EU Network of Excellence on Artificial Olfactory Sensing. He is the winner of the 2003 Gopel award.

Giorgio Sberveglieri received his degree in Physics from the University Parma, where starting in 1971 his research activities on the preparation of semiconducting thin film solar cells. In 1994, he has been appointed full Professor in Physics, formerly at the Faculty of Engineering of University of Ferrara and then in 1996 at the Faculty of Engineering of University of Brescia. Since 1988, he is the Director of the SENSOR Lab. mainly devoted to the preparation and characterisation of thin film chemical sensors. During his 25 years of scientific activities, Giorgio Sberveglieri published more than 200 papers on international reviews; he presented more than 90 Oral Communications to international congresses and numerous oral communications to national congresses.

Ibanez Riccò received his MSc in Physics from Parma University in 2000 with a thesis on preparation of TiO2:Fe gas sensors. In 2001 he joined SACMI Imola scarl where he is currently in charge of preparation of thin film gas sensors for the electronic nose.

Andrea Bresciani received his MSc in Industrial Chemistry from Bologna University in 1984. He started his activity on plastic materials at the “Giulio Natta” Research Centre of Himont spa, a worldwide leader in polypropylene manufacturing. In 1988 he joined SACMI Imola scarl, a worldwide leader in machinery for ceramics, plastics and food and beverage, where he is currently Technology Manager for advanced R&D.

View full text