Diesel Engine Acoustic Diagnostics With Machine Learning During Various Degradation Modes

Abstract

Abstract A diesel engine electrical generator set (‘gen-set’) was instrumented with an in-cylinder pressure indicating system as well as an acoustic emission sensor near the engine. Air filter clogging, rocker arm gap and fuel cetane changes were applied during which engine combustion and acoustic data were collected. Fast Fourier Transforms (FFTs) were analyzed on the acoustic data. FFT data were then applied to categorical supervised machine learning neural network analysis with MATLAB based tools. The detection of the various degradation modes was audibly determined with correlation coefficients greater than 99% on test data. Next, an unsupervised machine learning Self Organizing Map (SOM) was produced during normal-baseline operation of the engine. Application of the degraded mode engine sound data from operation with the various faults were then applied to the normal-baseline SOM. The quantization error of the various 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 promise as a method to monitor and detect changing engine operation. Companion in-cylinder combustion data shows how changing combustion characteristics result in emitted sound differences.