knowledge transfer


Odorous quality control in food packaging

Submitted by G. Tselentis, 28.06.2001, Service Center


One of the problems in food packaging quality control is to detect when the packaging material can chemically interact with other substances (like ink for labelling) and emit bad odours that could degrade food quality. In this case even if there is no evidence of health hazard the negative impact in the credibility and market value of the product is of high importance.

Odorous quality control in industry is typically performed by human operators. In some cases reference substances and human olfactory panels are used for better assessment. Olfactory devices (e-noses) emerged in the market and used for automating olfactory control, showing encouraging results. In this problem an e-nose is used for distinguishing malodorous packages used in a production site of pasta. The performance vs. typical classification techniques and the ability of the system to operate in acceptable limits while sensors might degrade from poisoning (i.e. sensor surface get saturated from chemical substances and the signal degrades) should be investigated.


Intelligent Techniques are investigated as they can provide soft thresholds for reject-accept similar to human operators. During the project a European consortium aimed to develop a prototype that can classify odours based on chemical sensor data. Data produced as sensors’ conductivity varies according to the volatile substance that contacts their surface. Partners experimented with Conducting Polymer (CP) sensors and concluded the last validation phase with Metal Oxide (MO) sensors. The aim was to investigate the power of uncertainty modelling techniques like fuzzy logic, neural networks and machine learning on chemical sensors data, as it is difficult to model the electrochemical interactions that take place on the surface of the sensor. Three parallel classification modules were developed using fuzzy sets, linguistic description and neural networks. Each module considered and treated data in a different way in order to provide greater system robustness. Classification results can be either merged or considered separately.

Status and results

The prototype is able to distinguish malodorous packages with a success ratio of ~80%. The prototype can be connected directly to a sampler and perform detection in short time (within minutes). The problem of sensor poisoning is the main factor for lowering the performance of the classifier.

Adaptivity and portability

The training of the classifiers with new data from the production line is in some cases time consuming and can take several hours when using machines with typical computing power. Sensor poisoning is confronted with a calibration procedure using a reference odorous substance. The classifiers use this reference to readjust the classes and thus the problem of adapting the quality control system is based on this calibration procedure. The system can be transferred in a new site with minimal adjustments but the classifiers should be retrained.

More information

For more search INTESA project (ESPRIT 25254) at

Tselentis G., Marcelloni F., Martin T., Sensi L., "Odour Classification based on Computational Intelligence Techniques" in "Advances in Computational Intelligence and Learning, Methods and Applications", Eds: Zimmermann, H-J., Tselentis G., van Someren M., Dounias G.,International Series in Intelligent Technologies, Kluwer Academic Publishers, (expected publication end 2001).

technical reports & case studies

executive summaries

scientific papers




search eunite with