SOUND-BASED ABNORMAL COMBUSTION CLASSIFICATION MODEL
FOR HIGH COMPRESSION RATIO, SPARK-IGNITION ENGINES USING
MEL-FREQUENCY CEPSTRUM COEFFICIENTS AND ENSEMBLE
LEARNING ALGORITHMS |
Seongsu Kim , Junghwan Kim |
School of Energy Systems Engineering, Chung-Ang University |
|
|
|
|
ABSTRACT |
A supervised machine learning model was developed to determine knocking in a spark ignition engine. The
engine sound was delivered to an operator via an extended metal tube installed on the cylinder block. The sound was recorded
using a smartphone at a sample frequency of 48,000 Hz. Thirty-nine features were extracted from the mel-frequency cepstrum
and spectral analysis. Neighborhood component analysis was performed to select eight features with the highest contributions.
The gentle adaptive boosting scheme, available in MATLAB, achieved the best results among the nine ensemble algorithms
used in this study, regardless of whether it was trained using all 39 features or the eight selected features. The best model
exhibited 99.98 % accuracy in classifying knock sounds and 99.85 % in classifying normal sounds. A second round of validation
was performed to investigate the robustness of the proposed model. The dataset used in this round was acquired from a slightly
advanced spark timing case, in which knock intensity varied from mild to severe. The model achieved 100 % accuracy in
detecting both knock and normal sounds. Each signal segment contained an individual cycle sound to evaluate the feasibility
of a model for detecting individual knock cycles during real-time engine sound monitoring. |
Key Words:
Engine knock, Supervised learning, Mel-frequency cepstrum coefficients, Ensemble learning |
|
|
|