| KNOCK ONSET DETERMINATION WITH 1D CNN USING RANDOM
SEARCH HYPERPARAMETER OPTIMIZATION AND DATA
AUGMENTATION IN SI ENGINE |
| Jihwan Park 1, Seunghyup Shin 2, Sechul Oh 3, Sangyul Lee 4, Woojae Shin 1, Kyoungdoug Min 1 |
1Department of Mechanical Engineering, Seoul National University
2Department of Artificial Intelligence, Sejong University
3Department of Mobility Power Research, Eco-Friendly Energy Conversion Research Division, Korea Institute of Machinery and Materials
4Department of Mechanical and Electronics Engineering, Hansung University |
|
|
|
|
|
|
|
| ABSTRACT |
|
For avoiding knock occurrence in SI engines, spark timing is retarded whenever the knock has occurred which
leads to a loss of thermal efficiency. Therefore, the knock occurrence needs to be properly controlled. For doing that, knock
should preemptively be predicted and controlled. Prerequisite data for knock prediction modelling is a knock onset position,
which can be figured out by finding the starting point of the oscillation on pressure data. A deep learning knock onset
determination model was developed in a previous study, and showed the highest accuracy among the comparable methods,
the model showed weak robustness on knock cycles obtained in different engine experiments. Meanwhile, the 1D CNN
model has been widely used in signal processing fields with its advantage of having a feature extraction layer, and the model
is introduced in this study for determining the knock onset. Dataset from four different engine types were used for verifying
the model accuracy and robustness. The dataset was augmented by calculation windows for producing various data with
limited data sources. Hyperparameters of the model were optimized with random search. The accuracy standard deviation
following engine types in terms of RMSE was improved by 77.4 % from 0.827 CA to 0.187 CA. |
| Key Words:
Spark ignition engine, Knock, Knock onset determination, Deep learning, 1D convolution neural network,
Data augmentation |
|
|
|
Corresponding Author.
Kyoungdoug Min , Email. 
Preview
Full text via DOI
Download Citation 
Print 
