Learning Gasoline Direct Injector Dynamics Using Artificial Neural Networks General Motors LLC

Format:

Conference/Event

Author/Creator:

Lucido, Lucido, author.

Contributor:

Shibata, Jonathan

Conference Name:

WCX World Congress Experience (2018-04-10 : Detroit, Michigan, United States)

Language:

English

Physical Description:

1 online resource

Place of Publication:

Warrendale, PA SAE International 2018

Summary:

In today's race for improved fuel economy and lower emissions from gasoline engines, precise metering of delivered fuel is essential. Gasoline Direct Injection fuel systems provide the means for improved combustion efficiency through mixture preparation and better atomization. These improvements can be achieved from both increasing fuel pressure and using multiple injection events, which significantly reduce the required energizing time per injection, and in a number of cases, force the injector to operate at less than full stroke. When the injector operates in this condition, the influence of variation in injector dynamics account for a large percentage of the delivered fuel and require compensation to ensure accurate fuel delivery. Injector dynamics such as opening delay and closing time are influenced by operating conditions such as fuel pressure, energizing time, and temperature. In addition, build variation in injector hardware including spring force, needle lift, and coil resistance have significant influence on the dynamics and continue to vary as the injector ages. Conventional techniques for learning injector dynamics such as injector closing time, require finding an inflection point in the differential voltage signal; however, injectors with weaker signals can compromise the effectiveness of these techniques. This paper's approach uses machine learning via artificial neural network fitting to analyze Gasoline Direct Injector differential voltage signals to learn injector dynamics over the life of the injector.The paper uses a standard two-layer feedforward neural network and explores different training algorithms for optimum performance and efficiency. The training sets encompass data under numerous operating conditions along with a wide range of injector hardware variations including parts toward the end of useful life. The predicted injector dynamics from the neural network are compared to the measured data in order to evaluate the performance of the algorithms. Furthermore, multiple injector designs are evaluated in order to assess the robustness of the techniques across platforms

Notes:

Vendor supplied data

Publisher Number:

2018-01-0863

Access Restriction:

Restricted for use by site license