Technical communication on engine fault detection with small datasets

Predictive maintenance is becoming an increasingly important tool in the marine industry, where unexpected engine failures are costly, particularly due to the resulting downtime. Acoustic sensors have the potential to detect developing damage before critical failures occur based on anomalies in the sound signature. However, in order to determine the baseline engine condition using learning algorithms, a comprehensive set of data in both normal and fault conditions is required, which is difficult to obtain.In this paper, we propose an approach that uses transfer learning and synthetic data for fault detection in marine engines on small datasets. The synthetic data generation methods investigated in this paper include WaveGAN, WaveNet and SpecGAN. These deep-learning-based networks can be used to find patterns in audio files and use them to generate new data that is indistinguishable from the original data to the human ear. Using the synthetic data, several unsupervised transfer learning techniques were investigated, including k-means and autoencoder.The algorithms were evaluated using real data from a river cruise ship engine and test measurements on a small electric motor under laboratory conditions. The combination and parameterisation of the methods should be thoroughly investigated for a final implementation of the method in an industrial setting. Overall, the approach proposed in this paper gives promising results for anomaly detection with small datasets, but needs to be further improved to achieve the best possible results.