Diesel Engine Acoustic Emission Airflow Clogging Diagnostics With Machine Learning

Abstract

A diesel engine electrical generator set (“gen-set”) was instrumented with in-cylinder indicating sensors as well as acoustic emission microphones near the engine. Air filter clogging was emulated by progressive restriction of the engine's inlet air flow path during which comprehensive engine and acoustic data were collected. Fast Fourier transforms (FFTs) were analyzed on the acoustic data. Dominant FFT peaks were then applied to supervised machine learning neural network analysis with matlab-based tools. The progressive detection of the air path clogging was audibly determined with correlation coefficients greater than 95% on test data sets for various FFT minimum intensity thresholds. Further, unsupervised machine learning self-organizing maps (SOMs) were produced during normal-baseline operation of the engine. The degrading air flow engine sound data were then applied to the normal-baseline operation SOM. The quantization error (QE) of the degraded engine data showed clear statistical differentiation from the normal operation data map. This unsupervised SOM-based approach does not know the engine degradation behavior in advance, yet shows clear promise as a method to monitor and detect changing engine operation. Companion in-cylinder combustion data additionally shows the degrading nature of the engine's combustion with progressive airflow restriction (richer and lower density combustion).