A MODIFIED MULTILAYER PERCEPTRON MODEL FOR GAS-MIXTURE ANALYSIS

An investigation was carried out on the ability of artificial neural n etworks (ANNs) to quantify the concentrations of individual gases and gas mixtures in air from patterns generated by an array of chemically modified sintered SnO2 sensors. The aim of this study was to design a neural paradigm that could compute the concentrations of four gases (H -2, CH4, CO and CO2) in simple gas mixtures. The experimental data wer e gathered by a gas test station with an array of three commercial Tag uchi sensors (822, 813 and 815) and three catalytically modified senso rs (812 with 1 mug of Pd, Au, Rh, respectively). The change in conduct ance of each of the six sensors was measured up to concentrations of 1 5 000 (H-2), 10 000 (CH4), 500 (CO) and 15 000 (CO2) PPM. Analysis of the raw data showed that the individual sensor responses were highly n on-linear over the chosen concentration ranges and that the CO2 data f ell in the noise. So the detection Of CO2, on its own or in gas mixtur es, was problematic with sintered SnO2 sensors. Initially, three prepr ocessing algorithms were applied to the input data and fed into fully connected multilayer perceptron models with the backpropagation paradi gm. The network error was minimised by changing the number and size of the hidden layers and the learning rate and momentum, yet its overall performance was still poor. Consequently, the model was modified by u sing three non-linear target functions (log, sigmoid and tanh). These models only gave slightly improved results. Finally, we adopted a part ially connected network with the six input elements connected to all 9 elements in a single hidden layer. This corresponded to 3 for each ga s (excluding the CO, data), but each group of three elements in the hi dden layer was only connected up to one output. This helped to compens ate for the relatively small signal for CO compared with H-2 and CH4, the idea being to separate the learning characteristics for each gas a nd thus obviate poor data for one gas affecting another with better da ta. The best results were obtained using log input and tanh output pro cessing functions. In this case, the maximum prediction error was 10% for H-2, CH, and CO gases. It was also possible to quantify H-2:CH, ga s mixtures to a similar accuracy with no interference effect observed from humidity changes. The CO concentration could also be detected in H-2:CH4:CO gas mixtures but to a much lower degree of accuracy.