Hybrid Model-Based Hardware Acceleration for Diesel Engine NOx Emission Prediction

Nitrogen oxides (NOx) from diesel engines pose significant environmental and public health challenges. To comply with stringent emission standards, this study proposes a hybrid CNN-LSTM model for real-time prediction of engine-out NOx emissions. Using real-vehicle operating data, the CNN extracts spatial features from engine parameters while the LSTM captures temporal dependencies in emission sequences. The model achieves a mean absolute error (MAE) of 28.05 ppm, reducing errors by 15.71% and 19.97% compared to standalone CNN and LSTM models, respectively. In the context of in-vehicle deployment, INT16 quantization limits MAE degradation to 8.3% while enabling FPGA acceleration. The customized hardware accelerator leverages parallel computing and on-chip memory to optimize convolution and LSTM operations via time-division multiplexing. Implemented on a Kintex-7 FPGA at 100 MHz, it achieves a latency of 0.12ms with a power consumption of 0.71 W. This solution offers a high-precision, low-latency deployment option for real-time monitoring of diesel engine emissions.