Statistical simulator for the engine knock

Although the automotive industry is rapidly shifting to electric cars, Hybrid Electric Vehicles (HEVs) – with both an electric motor and an internal combustion engine – will still dominate for the next half century. Increasing their efficiency remains crucial to reducing emissions, and a key issue is optimising the dynamics of so-called “engine knock” caused by the spontaneous ignition of unburned fuel during the engine cycle. While some knock is good for combustion efficiency, frequent knock causes both increased emissions and serious cylinder damage.

Methods to minimise these issues have tried to exert engine control through statistical feedback, using experiments to quantify the probability distribution of knock events and how it varies with engine control parameters, and then using feed-back to move the engine towards optimal states. But collecting data through experiments incurs significant expenses, and researchers have therefore turned instead to knock simulators. These work to predict the knock behaviour under different parameter settings either by A) simulating the actually engine physics, or B) building a stochastic map from the engine inputs to the probability distribution of the knock intensity.

In a new paper, LML External Fellow Jiancang Zhuang and colleagues explore a limitation of both of these approaches, stemming from their assumption that the data is single Gaussian, which means these methods can only represent a highly restricted family of statistical outcomes. As an alternative, the authors propose a different statistical simulator based on the so-called Mixture Density Network (MDN) – a kind of neural network which can be trained to link input and output data. After training, this network applied effectively approximates the function from input signal to the knock intensity distribution. Zhuang and colleagues go on to evaluate the proposed method and conform that it outperforms standard models in generating synthetic data with a closer probability distribution to the real experimental data and in obtaining a closer relationship to the actual engine performance data. As the method has validated accuracy and low computation burden, it should be useful for practical industrial applications.

The paper is available at

0 replies

Leave a Reply

Want to join the discussion?
Feel free to contribute!

Leave a Reply

Your email address will not be published. Required fields are marked *