Improved pulsed photoacoustic detection by means of an adapted filter

We present a numerical and experimental study of two adapted filters devised to the quantitative analysis of weak photoacoustic signals. The first one is a simple convolution-type one and the other is based on neural networks of the multilayer perceptron type. The theoretical signal used as one of the inputs in both filters is derived from the solution of the transient response of the acoustic cell modeled with a simple transmission-line analogue. The filters were tested numerically by using the theoretical signal corrupted with white noise. After 500 iterations it was possible to define an average error for the returned value of each filter. Since the neural network outperformed the convolution-type, we assessed its performance by measuring SF 6 traces diluted in N 2 and excited by tuned TEA CO 2 laser. The results show the use of the neural network filter allows recovering a signal with poor signal-to-noise ratio without resorting to extensive averaging, thus reducing the acquisition time while improving the precision of the measurement.