A supervised machine learning technique for combustion diagnosis using a vibration sensor signal

Abstract

Machine learning (ML) techniques are increasingly spreading in the automotive area. The advantage of using data-based algorithms is evidenced in the study of complex nonlinear phenomena, such as engine combustion. An important aspect when using these techniques is selecting the appropriate features to feed the algorithms. Considering that an accurate estimation of the combustion parameters is necessary to maintain high efficiencies, the following study evaluates the potential of using the engine block vibration signal to estimate parameters, such as the indicated mean effective pressure (IMEP) or the combustion phasing and duration. The block vibration data has proven to contain information regarding engine combustion. Yet, adequate data processing is required to retrieve the features that best contribute to the ML model prediction. To this end, the methodology employs Singular Value Decomposition (SVD) to extract these features from the knock signal’s spectrogram. Then, a correlation with the combustion parameters is made through an artificial neural network (ANN). Results yielded an improvement in the estimation accuracy of the combustion phasing parameters, compared with an ANN model using conventional engine control inputs (spark advance, fuel mass and engine speed). The mean absolute error decreased from 8% for the CA90 up to 25% for the CA50. Furthermore, this approach achieved better generalisation capabilities when unlearned conditions were added to a test set, showing an 80% decrease in the MAE with the proposed method. Measurements were performed on a spark ignition engine at different operating conditions.