Laser recognition of underwater wakes based on principle component analysis and machine learning

To achieve the recognition of underwater wakes, this study integrates principal component analysis (PCA) with various machine learning methods to classify different types of wakes using the characteristics of forward-propagated laser signals. In processing the forward-propagated optical signals, time-domain and frequency-domain features are first computed, followed by dimensionality reduction of the original features via PCA. The reduced features are then fed into machine learning classifiers to distinguish wakes generated by different underwater thrusters under tank conditions. Furthermore, lake test validates the practical feasibility of the proposed method. Results indicate that under tank conditions, the optimal machine learning method is support vector machine (SVM), achieving a accuracy of 89.93%. Whereas in the lake test, the optimal methods are k-nearest neighbors (KNN) and random forest (RF), with accuracies of 96.39% and 96.25% respectively. As the wake intensity varies and the number of thruster types increases, the inter-wake dissimilarity decreases, and the PCA feature boundaries become blurred. Additionally, the importance of time-domain versus frequency-domain features differs significantly across classification tasks. Both experimental and lake test results demonstrate that the proposed method can effectively discriminate among different underwater wakes.