An FPGA-Based On-Chip Neural Network for TDLAS Tomography in Dynamic Flames

Abstract

A back propagation (BP) neural network is introduced and implemented on a field programmable gate array (FPGA) chip into tunable diode laser absorption spectroscopy (TDLAS) tomography for fast response and a high signal-to-noise ratio (SNR) in temperature and water concentration imaging. The network implemented on the FPGA extracts the peak values of normalized second harmonics of absorption spectra. A recursive demodulator based on a Kalman filter generates outputs of the training database. Compared with the typical quadrature demodulator, the sampling time of the BP neural network drops to one-quarter of the typical one, with a 5 dB increase in the SNR at different noise levels. To verify and evaluate the proposed method, the FPGA-based 32-bit fixed-point BP neural network was compared with the floating-point BP neural network and the quadrature demodulator in numerical simulations. In the real experiments, Bunsen burner flames with acoustic excitation were imaged, and noisy distortions of the normalized second harmonics and average temperature were effectively reduced by using the proposed method. Temporal fluctuations of up to 750 Hz can be clearly identified from variations of reconstructed distributions of temperature and water vapor molar concentration, in which cases the typical quadrature demodulator failed to work. These results showed that the proposed method works well in the FPGA chip and reveals more details in dynamic flame monitoring.