(Invited) Using Machine Learning to Decode Output of a Mixed-Potential Sensor Array for Automotive Exhaust Monitoring

Abstract

Introduction In order to meet the latest requirements for more fuel-efficient cars with low emissions, it is essential to equip vehicles with advanced sensor systems whose readouts can be used to determine concentrations of key automotive exhaust pollutants such as NOx, NH3 and C3H8 in real time [1]. Arrays of mixed-potential electrochemical sensors (MPES) provide a competitive sensor platform for this task. MPES are electrochemical devices that develop a non-Nernstian potential due to differences in the redox kinetics of various gas species at each electrode/electrolyte gas interface [2,3]. To mitigate signal drift and poor reproducibility, we have developed a patented MPES design utilizing dense electrodes with an overcoat of porous electrolytes such as yttria-stabilized zirconia (YSZ). This design has resulted in highly stable, reproducible, and durable devices which were tested for up to 1000h of operation [4]. In this study, we have employed an array of four mixed-potential sensors (Cr450, Au475, H545, and Cr470) to monitor gas mixtures that mimic automotive exhaust in a laboratory setting. However, even with the latest sensor designs we observe significant cross-specificity and non-linearity in sensor responses to each target gas. Thus, reliable inference of the concentration (or even the presence) of each constituent gas in a complex mixture requires application of advanced machine-learning techniques that will be the focus of this presentation. The ability to decipher the content of gas or liquid mixtures both quickly and reliably has potential applications in many areas of science and technology, including monitoring of various technological processes and continuous observation of air quality in the interests of ecological studies and national security. Our computational methodology can be readily adapted to various sensor array platforms in which gas or liquid mixtures elicit complex response patterns. Method Our previous work has focused on developing computational models which were based on the fundamental electrochemistry of MPES and utilized detailed quantitative descriptions of gas-sensor interactions [5]. We were able to treat both two- and three-gas mixtures (Fig. 1) and take into account the non-linearities of sensor responses. Our model was used to estimate relative concentrations of C3H8, NH3 and NO2 with respect to NO with the maximum error of 14.0% and the average error of 1.8%. Furthermore, we predicted the absolute concentration of each gas in two- and three-gas mixtures with the average error of 3.2%. In this presentation, we will discuss a novel approach to analyzing sensor array output based on using support vector machines (SVM) to classify gas mixtures into types, thus inferring the number and the identity of gases present in a given test mixture. SVM classification is followed by Gaussian kernel regression, a Bayesian machine-learning technique which allows us to predict the concentrations of each constituent gas. The primary advantages of this two-stage algorithm are its versatility: the approach is independent of the sensor platform employed to provide the measurements but requires retraining. Furthermore, given a novel set of measurements, the algorithm is able to estimate both the expected concentration of each component in the mixture and the corresponding uncertainty of the predictions. Unlike maximum-likelihood methods, Gaussian kernel regression is less prone to overfitting and readily lends itself to model comparisons, allowing us to choose the optimal model complexity given the available data. Discussion and Conclusions There is a growing interest in computational approaches designed to infer gas concentrations in complex mixtures such as diesel exhaust solely on the basis of real-time sensor array readouts. In this presentation, we will discuss SVM/Gaussian kernel regression analysis of two-, three- and four-gas mixtures (Fig. 1) using voltage readouts of an array of four mixed-potential sensors as input. Our machine learning approach is scalable, robust, and easily transferrable to other sensor platforms that are currently employed in numerous scientific and industrial settings. After training the model parameters on gas-mixture data generated in the lab under controlled conditions, the computational model can be employed to predict gas concentrations from novel measurements in real time, as required by real-world applications. Figure 1. Plot of the voltage response (in volts) of three sensors in the array: Cr450, Au475 and H545 (Cr470 was excluded due to its similarity with Cr450). Color coding indicates mixture type.