| Abstract [eng] |
This study examines the contactless emission prediction of internal combustion engines using artificial neural network models and statistical methods. Given the increasing requirements for emission reduction and transport electrification, there is a growing need for efficient emission monitoring and forecasting technologies, particularly in the industrial, energy, and heavy transport sectors, where internal combustion engines will remain essential. During the study, an experimental test stand was designed and built, utilizing a diesel generator to analyze the number of particulate emissions. The stand was equipped with specialized measuring instruments, including accelerometers for vibration measurement, a sound level meter, an infrared camera, and a particle analyzer for exhaust emissions. The collected experimental data was processed and used for the study. For analysis, a transformer-based artificial neural network model was selected. To compare its efficiency, a recurrent LSTM model and a statistical ANCOVA model were also tested. The study aimed to evaluate the effectiveness of transformer models in predicting particle emissions in a contactless manner, utilizing engine vibrations, sound levels, and thermal imaging data. Experimental results revealed that transformer models achieved the highest prediction accuracy, outperforming other methods. Based on this information, a modular model system was developed that can independently interpret all types of data collected during the experiment and integrate predictions into a unified forecast. The combined transformer model achieved 96.59% accuracy, successfully predicting varying concentrations of particulate emissions, surpassing LSTM and ANCOVA models, which attained 62.34% and 61.82% accuracy, respectively. Infrared images contributed the least to the overall prediction, demonstrating the lowest accuracy at 86.3% among separate data types. |
