Oil condition prediction in diesel generators using acoustic signals

Abstract

This research proposes a novel method for predicting the oil condition in CAT diesel generators using acoustic signals, aiming to overcome the limitations of traditional Scheduled Oil Sampling (S·O·SSM) testing methods. Conventional S·O·SSM programs often face challenges such as delayed results, sampling errors, and high operational costs. To address these, the study develops a non-invasive acoustic monitoring technique that correlates generator sound patterns with oil health indicators. Audio samples were collected using a Blue Yeti microphone under controlled conditions, and corresponding physical oil samples were subjected to standard S·O·SSM laboratory analysis. Various time-domain and frequency-domain features, including MFCCs, Zero Crossing Rate, and Spectral Centroid, were extracted using the Librosa library. A neural network model was developed to classify oil condition states into "Action Required," "No Action Required," or "Monitor," based on these acoustic features. Testing the system on a dataset of 75 samples showed encouraging results, with the second iteration achieving improved classification accuracy. Findings confirm that generator acoustic emissions change with oil degradation, validating the hypothesis that sound can predict oil condition. Conclusions highlight that this method enables timely, cost-effective oil monitoring without operational disruption. Recommendations for future work include optimizing the neural network through learning rate adjustments, particularly expanding the sample dataset, exploring ensemble methods, and validating performance across different generator types to enhance system robustness and industrial applicability.