Machine learning–based analysis of in-cylinder flow fields to predict combustion engine performance

Abstract

Cycle-to-cycle variations in an optically accessible four-stroke direct injection spark-ignition gasoline engine are investigated using high-speed scanning particle image velocimetry and in-cylinder pressure measurements. Particle image velocimetry allows to measure in-cylinder flow fields at high spatial and temporal resolution. Binary classifiers are used to predict combustion cycles of high indicated mean effective pressure based on in-cylinder flow features and engineered tumble features obtained during the intake and the compression stroke. Basic in-cylinder flow features of the mid-cylinder plane are sufficient to predict combustion cycles of high indicated mean effective pressure as early as 180 degree crank angle before the top dead center at 0 degree crank angle. Engineered characteristic tumble features derived from the flow field are not superior to the basic flow features. The results are independent of the tested machine learning method (multilayer perceptron and boosted decision trees) and robust to hyper-parameter selection.

A correlation-based model order reduction approach for a diesel engine NOxand brake mean effective pressure dynamic model using machine learning

“…Therefore, the HOM and LOM nonlinear controloriented models (NCOMs) for NO x and BMEP are obtained as equations ( 31) and (30) expected when looking at Figure 11…”

Deep learning procedure for knock, performance and emission prediction at steady-state condition of a gasoline engine

“…As Adam algorithm introduces moments of the gradient, it can overcome the local minima problem. The gradients of the stochastic objective function at timestep t (g t ) are obtained as equation (5), then exponential moving averages of the gradient (m t ) and the squared gradient (v t ) are updated by equations (6) and (7). To solve the bias initialization problem that the moment biases toward 0 for the initial timestep, bias-correction is adopted as equation (8).…”

“…Hanuschkin et al conducted feature extraction, including in-cylinder flow and tumble features, to classify high indicated mean effective pressure (IMEP) cycles using a machine learning approach. 7…”

“…Hanuschkin et al conducted feature extraction, including in-cylinder flow and tumble features, to classify high indicated mean effective pressure (IMEP) cycles using a machine learning approach. 7 Recently, deep learning has been recognized as a main stream of AI, because deep learning has shown much higher accuracy than other machine learning approaches. In the automotive research field, several studies adopting deep learning algorithms have been conducted.…”

“…Most previous studies, [6][7][8][9]13,15,16,19,21,[24][25][26] studied machine learning, and deep learning specifically, with MATLAB. Although MATLAB is easy to use, it has limitations of expandability and standardization of deep learning algorithms.…”