Acoustic signature-based approach for machinery condition monitoring in autonomous ships

In conventional ships, machinery conditions are monitored through a human watch system. Sensor systems, integrated into the machinery, provide a live feed of the parameters to the engine control room. Operators in the control room continuously monitor the sensor feed, while watchkeepers visit the engine room regularly to observe the machinery operation and identify any abnormalities. The knowledge and experience of the watchkeepers and operators play a key role in this process. In fully autonomous ships, sensory inputs would be transferred to a shore-based control centre to be monitored by a remote operator. Even though a human element is still present in the remote operation, the physical presence of a watchkeeper in the engine room does not exist. Thus, the human experience that identifies deteriorating machinery conditions is absent.

This study developed an acoustic-based machine learning algorithm to mimic the ability of engine room watchkeepers to classify machinery conditions from auditory inspection. Firstly, a series of experiments was conducted to collect machinery acoustic signals. These acoustic signals were processed using time domain and frequency domain techniques to extract the acoustic features to be used in training several supervised machine learning algorithms. Based on the highest classification accuracies and the consistency of the results, the dominant feature and algorithm were selected. Then the experiments were extended to validate the findings and identify which kernel function can be generalised to classify acoustic signals of machinery.

The study found that the Support Vector Machines (SVM) algorithm, with a second-order polynomial kernel function, had the highest generalisation capability of the kernel functions in classifying machinery conditions using their acoustic data. In each experiment, the performance evaluation of kernel functions was conducted incorporating the percentage of training accuracy, ROC curve, confusion matrix, and running time of the algorithms. The consistency of the classification accuracy results was obtained with the Fast Fourier Transform (FFT) magnitudes from the frequency domain as the dominant input feature in the SVM algorithm. It was found that the accuracy of the second-order polynomial kernel function was independent of the microphone used to collect the acoustic data. The validation results have shown an accuracy of over 95% for unseen data when fed into the algorithm trained using the experimental data. The results indicate that the SVM algorithm with a second-order kernel function is an effective and efficient way to classify the machinery acoustic signals and contribute to mimicking the engine room watchkeepers' auditory inspection ability.

The findings of the study are beneficial to mitigate the lack of human observation in the engine room of autonomous ships. The proposed method of classifying machinery-condition using acoustic data can play a key role in shore-based control centres, keeping operators informed about the machinery condition onboard ship. An SVM acoustic-based engine room monitoring algorithm could also be used in existing conventional ships, thereby reducing the hours that watchkeepers need to spend in the engine rooms. Implementing the algorithm in conventional ships will provide future researchers with informed expertise and a wide range of data from actual engine rooms. Further research can then incorporate this enhanced understanding into the monitoring of machinery on fully autonomous ships.

In future studies, the classification of machinery conditions will be improved by implementing a similar approach using the data collected from other human senses (e.g. sight, smell, touch and taste). Consequently, there will be an enhanced awareness of the real-time condition of the ship's engine room machinery.