Improvement of Knock Onset Determination Based on Supervised Deep Learning Using Data Filtering Seoul National University

Format:

Book

Conference/Event

Author/Creator:

Park, Jihwan, author.

Contributor:

Cho, Seokwon

Kim, Minjae

Lee, Sangyul

Min, Kyoungdoug

Shin, Seunghyup

Song, Chiheon

Conference Name:

SAE WCX Digital Summit (2021-04-13 : Live Online, Pennsylvania, United States)

Language:

English

Physical Description:

1 online resource cm

Place of Publication:

Warrendale, PA SAE International 2021

Summary:

Regulations regarding vehicles' CO2 emissions are continuing to become stricter due to global warming. The CO2 regulations urge automobile manufacturers to develop gasoline engines with improved efficiency; however, the main obstacle to the improvement is the knock phenomenon in spark-ignition engines. If knock is predicted, the efficiency potential can be maximized in an engine by applying modest spark timing. Several research regarding knock prediction modeling have been conducted, and typically Livengood-Wu integral model is used to predict the knock occurrence. For the prediction, knock onset should be determined on a given pressure signal of given knock cycles for establishing the 0D ignition delay model. Several methodologies for knock onset determination have been developed because checking all the knock onset position by hand is impossible considering the breadth of data sets. Deep learning technique have been presented as a solution to the establishment of the knock onset methodology in the previous study. The model showed high accuracy and feasibility on knock onset determination; however, there was a limitation on robustness; the unstable robustness of knock onset determination could adversely affect the 0D ignition delay model. In this study, an adversarial attack was applied to the pervious deep learning model for verifying the robustness, and it was shown that the model had unstable robustness. This model was improved on the aspect of the robustness with data filtering. The model showed stable robustness even with the adversarial attack. The variance of the determination decreased by 64.1% from 0.136 to 0.082, and the standard deviation of the determination decreased by 40.0% from 0.019 to 0.007. Additionally, it was verified that the robustness of the base knock onset model could affect the robustness of the 0D ignition delay model; the variance and the standard deviation decreased by 18.2% and 9.6% respectively, when the improved deep learning knock onset model was used as base knock onset determination rather than using the previous model

Notes:

Vendor supplied data

Publisher Number:

2021-01-0383

Access Restriction:

Restricted for use by site license