ABSTRACT
Drift is a dynamical process, caused by physical changes in the sensors and the chemical background, which gives an unstable signal over the time. The drift could be both reversible, e.g. condensation of vapor on the sensors, and irreversible, aging. The samples and the operator, via contamination of the instrument, could introduce drift. One kind of drift caused by the measurement history is memory effects. This means that measurement at time t is highly influenced of measurements at time t-n. This leads to that the same gas mixture will not give one well defined pattern. The drift causes the pattern recognition models to be very short-lived. If no drift correction of the sensor signals is made, the models will have a continuous need for recalibration. Since the training phase of the pattem recognition model should contain all future coming variance a lot of samples are necessary. In real applications, processes, the samples may be very expensive which makes it impossible to recalibrate the pattem recognition model often. Basically, there are
two ways to obtain long-term reproducible and reliable gas-sensor array based instruments and both of them are necessary. The first, and most obvious, alternative is to improve the sensors to give a more stable long-term performance. The second approach is to refine the calibration models so the influence of drift will be reduced. Different approaches to correct gas sensor data that suffer from drift have been tried. One way to monitor the instrument performance over time is to measure calibration standards, reference gas. Attempts have been made using a reference gas as a reference value and then correcting all subsequent readings accordingly.[7, 81 Component correction [9] is new interesting method using PCA and PLS algorithms to find the drift direction from measurements of a reference gas. This direction is then removed from the samples. Fundamental studies of the mathematical properties of the drift effects have been studied by Davide et a1.[10] It was seen that the behavior was different in different frequency ranges. This group has also developed successful pattern recognition models based on system identification theories.[ll, 121 These models do not require the use of reference gas. Adaptive resonance theory has also been used as recognition of measurements subjected to drift [13].
