Deep-learning based in-cylinder pressure modeling and resolution of ion current signals

摘要

Engine calibration becomes much more costly due to strict exhaust emission regulations and individual consumer needs. Measuring the in-cylinder pressure could be a promising alternative for various sensors, thus reducing the calibration cost because the measurement provides diverse, low-delay, and precise information inside the cylinders. However, a piezoelectric pressure sensor for this measurement is too expensive, which prevents its use in production vehicles. The less-expensive, more reliable, and responsive ion current measurement provides signals that highly correlate with the in-cylinder combustion process and pressure. Many proposed methods correlate the ion current and pressure through chemical-kinetic models or manually tuned machine-learning models. A few methods can automatically estimate the pressure change or predict the peak pressure with the ion current, which provides valuable information for advance control and engine monitoring. In this paper, an autoencoder deep-learning model is developed that is unprecedentedly well fit for these two tasks. It automatically encodes ion current signals into a semantic representation with a convolutional neural network. With this, the model can either predict the peak pressure or estimate the pressure change through the gated-recurrent-unit decoder of the model. The evaluation of the model predictions and estimations is performed on an actual engine-based dataset, where the model demonstrates state-of-the-art performance on both tasks, and the mean relative error is 7.84% and 19.68%, respectively. Additionally, an orthogonal analysis method is applied to study the resolution of the ion current signals, making it possible to categories these signals by converting them into semantic representations.