Real-time FPGA-based laser absorption spectroscopy using on-chip machine learning for 10 kHz intra-cycle emissions sensing towards adaptive reciprocating engines

Abstract

Fast emissions sensing is needed to enable rapid optimization on-the-fly for increasingly adaptive engine architectures to improve performance over a wide range of loads and to offer fuel flexibility towards a low-carbon energy future. In this work, a real-time laser absorption spectroscopy technique is developed for 10 kHz on-line measurements of engine exhaust gas temperature and carbon monoxide. Latency due to data reduction is significantly shortened through a machine learning approach to spectral analysis and minimization of post-processing complexity for implementation on a high-bandwidth field-programmable gate array (FPGA). The data reduction method is tested on cycle-resolved laser-absorption thermochemistry measurements in reciprocating piston engine exhaust. The sensor employs a quantum cascade laser to spectrally-resolve two fundamental rovibrational absorption lines of carbon monoxide near 4.9 μm at a rate of 10 kHz. The recorded signals are passed to an FPGA to infer CO concentration and gas temperature through either a pre-trained ridge regression or artificial neural network model. Models are constructed to limit complexity with the aim of minimizing resource utilization and latency on the FPGA. Experimental data are used to evaluate the prediction accuracy of the models, with the neural network achieving RMS errors in CO mole fraction and gas temperature of 0.0390% and 15.0 K, respectively, compared to spectral fitting of the absorption lineshapes. The data reduction latencies are measured through hardware-in-the-loop demonstration, achieving a 10 kHz throughput and 25 μs latency. Computational time of the end-to-end data reduction process on the FPGA is measured at 300 ns and 600 ns for the ridge regression and neural network, respectively. The data reduction methods presented in this work expand the utility of laser absorption spectroscopy for low-latency sensors that match the timescales of combustion and increase the potential for real-time sensing and control to minimize engine exhaust emissions and maximize performance.

Keywords

• Laser absorption spectroscopy
• Real-time
• FPGA
• Machine learning
• Neural network
• Low-latency sensor