With the continuous advancement of target stealth and noise reduction technologies, the radiated noise of targets has become increasingly weak, often resulting in intermittent trajectories on the azimuth profile. These trajectory gaps challenge traditional tracking methods in maintaining continuous tracking of the target effectively. Furthermore, in complex multi-target scenarios, mis-tracking of targets is likely to occur due to the lack of a robust association algorithm. To address these issues, this paper proposes a method combining traditional tracking techniques with deep neural networks, which facilitates rapid association and continuous tracking when the target signal reappears. Additionally, the proposed method allows for unified management of tracked targets, which is beneficial for subsequent analysis on target motion characteristics. Specifically, the method first uses conventional sonar tracking techniques to autonomously track sonar-detected targets, while extracting the power spectrum of the tracked targets and surrounding noise as input for training a network model. Each target is learned individually, and the learning result for each target forms a “box”, which refers to a set of neural network parameters. A “neural network archive” is then established in the background. When a new signal appears, automatic association with known targets is performed through this archive. The method proposed in this paper does not require the construction of training samples in advance, is convenient for target management and association, and facilitates the subsequent judgment of target motion characteristics.