Classification of a Ferry Network Activity Based on ES Sensors and Convolutional Neural Networks

In this article, a convolutional neural network (CNN) is proposed to classify the activity of a marine vessel network based on real-time data from passive electronic support (ES) sensor. The data are measured by a passive ES sensor based on the daily activity of the network for 24 h. The data are stored as formats of pulse descriptor words (PDWs). The pulse parameters of pulse amplitude (PA), pulsewidth (PW), pulse frequency (PF), and angle of arrival (AoA) are extracted from each PDW and combined into arrays with multiple denominations. The hybrid strategy based on oversampling and undersampling approaches is proposed to handle the imbalanced problem in the PDW datasets. The CNN model is trained and tested in two different cases. The first case involves the measurement of the accuracy and the training time of the model using different array combinations of PA, PW, PF, and AoA, whereas the approach of hybrid data sampling is applied for the second case. An exhaustive analysis is conducted to investigate the number of activity classes that can be achieved from the PDWs data while maintaining the high accuracy of the classification model. Experimental results show that the proposed system can achieve higher accuracy by utilizing more pulse parameters. In addition, the proposed system performs better than the existing system in terms of accuracy in different classes. More importantly, the results also show that the proposed system can get a high number of classes of 24 that cover the entire activities of the ferry network with the highest overall accuracy of 0.9877.