APPLICATION OF NEURAL-NETWORK SYSTEMS TO THE DYNAMIC-RESPONSE OF POLYMER-BASED SENSOR ARRAYS

The conventional calibration method for sensor arrays uses steady-stat e signals that depend on the gas concentration. This method can be tim e consuming if many concentrations and compositions of a multicomponen t mixture are required. Good experimental design may reduce the necess ary effort so that the number of calibration experiments is minimized. Dynamic measurements may significantly reduce the time of each calibr ation experiment. In the present approach a random walk through the do main of the gas concentrations is chosen with each step of the walk ad justed for a short time only. The sensor array consists of six polymer (polysiloxanes with functional groups)-coated bulk acoustic wave (BAW ) devices. The concentration domain is defined by a binary mixture of n-octane and toluene (150 to 800 ppm). Neural networks evaluate both q ualitative and quantitative information from; the sensor response. In particular, the extensions of feed-forward nets towards recurrent or t ime-delay structures can be used to solve problems related to dynamic evaluations (e.g., no steady-state signal, parameter drift). These net work architectures with different numbers of bidden neurons are applie d to evaluate the data from the BAW device array. The networks are tra ined with back-propagation-like training algorithms and are validated with arbitrary gas mixtures.