Unmanned underwater vehicle radiated noise recognition based on instantaneous frequency estimation and feature fusion network

Given the frequency modulation characteristics of line spectrum components in unmanned underwater vehicle (UUV) radiated noise during maneuvering, as well as the problem that the number of signal components is unknown a priori, a UUV radiated noise recognition method that adaptively extracts and fuses instantaneous frequency (IF) features is proposed. To overcome the limitation of traditional ridge extraction methods that require a preset number of components, the constant false alarm rate detector is combined with the Viterbi algorithm. Through an iterative masking and automatic termination mechanism, IF estimation is achieved without prior knowledge of the component number. To fully exploit the discriminative information in the estimated IF features, the Gramian angular field is employed to convert the IF into two-dimensional texture features, which are then fed into a dual-feature fusion network together with time-frequency features. A feature fusion attention mechanism is further introduced, where the IF features are used to guide the time-frequency branch to focus on key channels, enabling collaborative extraction of deep information across different feature modalities. To verify the discriminative property of IF features on UUV radiated noise and the effectiveness of the proposed method, the measured UUV radiated noise data was classified into categories according to the navigation status for classification experiments. The experimental results show that the proposed method achieves a classification accuracy of 75%, a balanced accuracy of 70%, and a macro F1 score of 68%, which is an improvement compared to the comparison methods.