Available online at www.sciencedirect.com Sensors and Actuators B 128 (2008) 594–602 Predicting gas concentrations of ternary gas mixtures for a predefined 3D sample space Bekir Mumyakmaz a,∗ , Ahmet ¨ Ozmen a , M. Ali Ebeo ˘ glu a,b , Cihat Tas ¸altın b a Department of Electrical and Electronics Engineering, Dumlupınar University, K¨ utahya 43100, Turkey b TUBITAK, Marmara Research Center, Materials and Chemical Tech. Res. Inst., P.O. Box 21, 42470 Gebze-Kocaeli, Turkey Received 6 April 2007; received in revised form 11 July 2007; accepted 12 July 2007 Available online 22 July 2007 Abstract This paper presents a QCM sensor array and a data processing system to find gas concentration ratios of ternary gas mixtures. QCM sensors are very commonly used for gas sensing and detecting systems since they have linear responses to variable gas concentrations, especially for single gas samples. However, analyzing gas mixtures with a QCM sensor array creates a huge amount of data. Processing this data to obtain meaningful information becomes a complex and non-parametric problem. In this work, data processing is divided into pre- and post-processing sections to increase performance. The pre-processing section filters redundant data out, and extracts meaningful data. Then, a quadratic polynomial curve fitting is applied to the data. The post-processing section includes two feed-forward multi-layer artificial neural networks (ANNs): one for classification of species, and the other for quantification of the concentration ratios. Three industrial chemicals used in the experiments are acetone, chloroform and methanol. Variable volumes of these chemicals and their single, binary and ternary mixtures are applied to the sensor array, and the dynamic data is collected from the sensor responses. The ANNs are trained with 309 preprocessed data set (90% of whole data) using Levenberg–Marquardt training algorithm. Finally, the system is tested with 34 set of real data (10% of whole data). The average success rate of finding the concentration amounts in the testing phase is 93.87%, and identifying the species (classification) is 100%. © 2007 Elsevier B.V. All rights reserved. Keywords: QCM; Gas sensor array; Concentration prediction; Artificial neural network 1. Introduction Gas analysis is commonly used for health care, food freshness control, forensic analysis and industrial applications [1–4]. Gas chromatography is a conventional way of analyzing gas mix- tures to find concentration ratios, which is also very expensive and daunting task. Increasing demands for gas analysis force scientists to find cost effective solutions to this problem [5]. Recent technological achievements on electronic gas sensors are tempting, but they usually generate a huge amount of sensor data. Hence, classification of species and/or prediction of con- centration ratios in a gas mixture become a challenging problem [6,7]. Current research trends in the area focus on smart sensor systems implemented with artificial neural networks (ANNs) to analyze gas mixtures [8–10]. ∗ Corresponding author. Tel.: +90 274 2652062x4252; fax: +90 274 2652066. E-mail address: mumyakmaz@dpu.edu.tr (B. Mumyakmaz). This work aims to analyze a ternary gas mixture using a smart sensor system. The analysis study is carried out with two aspects: (1) qualitative work to find either a gas specie existing or not existing in a mixture, and (2) quantitative work to predict con- centration ratios of existing gases in a mixture. Although the research can be implemented in many different gas types, the framework of experiments is limited to three industrial gases: acetone, chloroform and methanol. Their single, binary and ternary mixtures are used during the experiments from approx- imately 0 ppm to 12,000 ppm with a 2000 ppm increment in 7 steps. The target set of all experiments is created in a 3D cubic output space. Quartz crystal microbalances (QCM) type electronic gas sen- sors are used in the experiments. QCM basics can be described as follows: chemically coated quartz crystal frequency drifts unusually from its center depending on the gas type and con- centration in the ambiance. Since single sensor response is not sufficient to resolve ternary gas mixture properties, an array of 18 chemically different coated QCM sensors are used in a 0925-4005/$ – see front matter © 2007 Elsevier B.V. All rights reserved. doi:10.1016/j.snb.2007.07.062