Lightweight DenseNet for Submarine Cable Anchor Damage Event Classification

Rapid identification and localization of submarine cable anchor damage events is crucial and challenging for its safe and stable operation. In this paper, a submarine cable anchor damage recognition system using lightweight DenseNet network is developed by pruning the four fully connected layers in the original code into two layer and adjusting the parameters of the convolutional layer to make it lightweight. Experiments show that the accuracy reaches from 93.06% to 97.03%, the model size is reduced from the original 44.8m to 4.48m, and the total training time of the model is reduced from 4698s to 3296s. The lightweight DenseNet network proposed in this paper reduces the cost of time and reduces the size of the model without affecting the recognition accuracy.