PMGCC: a high-precision underwater lightweight delay estimation model based on causal partial convolution module

Abstract An essential evaluation metric for ocean observation devices like the Inverted Echo Sounder is the precision of time delay estimation (TDE). Only large mesoscale ocean dynamical processes, such as ocean circulation, Kuroshio, and mesoscale eddies, can be observed with the current equipment; sub-mesoscale and small-scale ocean dynamical processes, like internal waves, cannot be measured. This is attributed to the low accuracy of conventional delay estimation algorithms such as generalized cross-correlation (GCC) in the presence of heavy seafloor reverberation, sea surface reflections and scattering. A high-precision underwater lightweight delay estimation model based on the channel attention causal partial convolution (PConv) block is proposed in this paper to solve this issue. The model uses a convolutional neural network to learn and predict by feeding it the parameterized multiclass frequency-domain weighted GCC matrix. However, regular convolutional layers are computationally intensive and are not suitable for deployment on resource-constrained devices. Thus, a causal PConv layer is proposed, which uses channel attention weights to determine which input channels need to be convolved and suppresses irrelevant input channels, significantly reducing computational complexity. Through simulation experiments, the model overcomes the effects of the low signal-to-noise ratio and complex ocean environment, as well as obtains more accurate TDE results than the traditional algorithm and other models, with an root mean square error (RMSE) reduction of 40.35% than the GCC with phase transform in the −20 dB hydroacoustic environment, and an RMSE reduction of 5.23% than the MGCCFF method.