Deep-sea transponders serve as critical infrastructure, yet their operational lifetimes are significantly constrained by limited energy supplies. Consequently, acoustic wake-up mechanisms are essential to mitigate standby power consumption. These transponders are typically deployed on the seabed, deviating from the SOFAR channel axis, and remote transmission relies on incomplete channels characterized by multiple surface-bottom reflections. In such environments, severe and time-varying multipath effects degrade the detection performance of distant weak signals. To address these challenges, this paper proposes a robust detection framework that integrates a dual-frequency signal mechanism with lightweight knowledge distillation. First, dual-frequency signals are employed to circumvent the need for complex multipath channel estimation. Second, a TabPFN-based teacher network is constructed to extract robust feature representations. A structural-semantic collaborative distillation strategy is subsequently applied to align the student model with the teacher in both the embedding and logit spaces. This approach effectively transfers generalization capabilities to a lightweight micro-network comprising only 832 parameters, achieving high-precision diversity detection using a computationally efficient architecture. Sea experiment results demonstrate that the proposed method achieves a detection probability of 95.48% and a false alarm rate of 0.04% under incomplete channel conditions at ranges from 0 to 100 km.
🌊
