Abstract: Aiming at the problem that the bending of the sound ray leads to the deviation of traditional filtering result in the non-equal sound speed channel,a tracking model based on recurrent neural network is proposed.In the absence of sound speed profile,the model learns the mapping relationship between input observations and output state values through data-driven iterative training to achieve the precise information of the change of the target position and transient characteristics.The results of the Monte Carlo simulation experiment show that the horizontal distance tracking error is decreased by 4.06%and 1.57%,and the depth estimation error is decreased by 0.87%and 0.85%,compared with the traditional single-model filtering algorithm and interactive multi-model algorithm in the complex maneuvering scene.The proposed model has higher tracking accuracy,and is able to carry out transfer learning under the mismatched sound speed channel,so as to improve the generalization of the model to the mismatched environment.