In order to solve the problem of underwater precise positioning caused by underwater unstructured environment, this paper proposes an inertial odometer positioning method based on LSTM(long short term memory) for the positioning of underwater working robots. In the training phase, the network adds Gaussian white noise to the acceleration and angular velocity data of the IMU(inertial measurement unit) by simulating the noise model to achieve data enhancement, and then uses the ResNet18 to extract the motion characteristics of the robot. At the same time, the sampling time of the IMU is introduced in the input space of the network to enhance robustness. Then, the three-channel LSTM is used to map the extracted features to high-dimensional space, and feature fusion is performed. Finally, the full connection layer is used to predict the relative displacement and rotation of the robot. In the training process, the relative loss function and the absolute loss function are combined to ensure the short-term and long-term positioning accuracy of the network. Finally, multiple data sets and pool experiments are carried out to verify the effectiveness of the proposed method. The experimental results show that the method has good positioning performance and strong robustness in most scenarios.