A Deep Learning-Based Fault Detection Model for Optimization of Shipping Operations and Enhancement of Maritime Safety

Abstract

The ability to exploit data for obtaining useful and actionable information and for providing insights is an essential element for continuous process improvements. Recognizing the value of data as an asset, marine engineering puts data considerations at the core of system design. Used wisely, data can help the shipping sector to achieve operating cost savings and efficiency increase, higher safety, wellness of crew rates, and enhanced environmental protection and security of assets. The main goal of this study is to develop a methodology able to harmonize data collected from various sensors onboard and to implement a scalable and responsible artificial intelligence framework, to recognize patterns that indicate early signs of defective behavior in the operational state of the vessel. Specifically, the methodology examined in the present study is based on a 1D Convolutional Neural Network (CNN) being fed time series directly from the available dataset. For this endeavor, the dataset undergoes a preprocessing procedure. Aspiring to determine the effect of the parameters composing the networks and the values that ensure the best performance, a parametric inquiry is presented, determining the impact of the input period and the degree of degradation that our models identify adequately. The results provide an insightful picture of the applicability of 1D-CNN models in performing condition monitoring in ships, which is not thoroughly examined in the maritime sector for condition monitoring. The data modeling along with the development of the neural networks was undertaken with the Python programming language.

Keywords: Convolutional Neural Network (CNN); condition monitoring; deep learning; fault detection; maritime.

Figure 1

Demonstration of a trigger event from an actual signal collected from the D/G LO pressure parameter onboard a bulk carrier.

Figure 2

Overview of the 1D-CNN proposed in this study.

Figure 3

Conceptualization of a CNN.

Figure 4

Confusion matrix demonstrations. Left: Class 0 (vessel healthy), right: Class 3 (cylinder); TP: True Positive, TN: True Negative, FP: False Positive, FN: False Negative.

Figure 5

Operational flow of LAROS™ CMS providing synchronized and reliable data.

Figure 6

Sources of collected data and the respective features.

Figure 7

Outline of the study.

Figure 8

Visualization of filters applied to RPM and ship’s velocity variables.

Figure 9

Outlier removal relative to main engine rpm. (Left): Fuel Oil Consumption (FOC); (Right): exhaust gas temperature of cylinder 1.

Figure 10

Real (blue line) vs. smoothed (red line) signal obtained with a 15 min averaging window.

Figure 11

Signal of the average temperature among the main engine’s six cylinders for 8 h.

Figure 12

Demonstration of segmentation.

Figure 13

Time interval varying models.

Figure 14

ROC curves and AUC values comparison of corresponding time-varying models.

Figure 15

Distribution of accuracy scores of the numerous models during hyperparameter tuning across the three developed instances.

Figure 16

Models developed with diverse levels of degradation and input time horizons.

Figure 17

Demonstration of levels of artificially induced degradation.

Figure 18

ROC curves and AUC values comparison of corresponding degradation varying models.

Figure 19

Distribution of accuracy scores of the numerous models during hyperparameter tuning across the three instances developed.

Figure 20

Comparison between proposed methodology 1D-CNN (blue line) and two benchmark classifiers (Random Forest c—pink line, Support Vector—red line) through their yielded AUC values and the respective ROC curves, for the three instances investigated.