Data-Driven Prediction of Key Combustion Parameters Based on an Intelligent Diesel Fuel Injector for Large Engine Applications

Abstract

<div class="section abstract"><div class="htmlview paragraph">Digital technologies are capable of making a significant contribution to improving large internal combustion engine technology. In particular, methods from the field of artificial intelligence are opening up new avenues. So-called “intelligent” engine components rely on advanced instrumentation and data analytics to create value-added data, which in turn can serve as the basis for applications such as condition monitoring, predictive maintenance and controls. For related components and systems, these data may also allow for novel condition monitoring approaches. This paper describes the use of value-added data from an intelligent diesel fuel injection valve that give detailed information about the injection process for real-time prediction of key combustion parameters such as indicated mean effective pressure, maximum cylinder pressure and combustion phasing. These parameters are usually involved in combustion controls and power unit condition monitoring and normally acquired using in-cylinder pressure indication systems, which are costly and prone to wear. On the one hand, a data-driven model for key combustion parameters based on an intelligent fuel injection valve could replace an indication system. On the other hand, such a model may enable backup functionality and mutual condition monitoring of the fuel injection valve and the indication system. The data required for model building were acquired from a medium-speed four-stroke single-cylinder research engine with a displacement of approximately 15.7 dm³. Different machine learning methods are compared to obtain an accurate yet reliable model for each of the desired combustion parameters. In addition to the value-added injection data, readily available parameters on production engines serve as model inputs (e.g., engine speed, charge air and exhaust gas pressures). Based on the results, the quality of the model predictions is evaluated, and it is assessed whether the approach might be useful for series engine applications.</div></div>