This git repo contains the dataset, code and weights of the deep learning architecture, BeamsNet, which was introduced in the paper BeamsNet: A data-driven Approach Enhancing Doppler Velocity Log Measurements for Autonomous Underwater Vehicle Navigation.
Autonomous underwater vehicles (AUV) perform various applications such as seafloor mapping and underwater structure health monitoring. Commonly, an inertial navigation system aided by a Doppler velocity log (DVL) is used to provide the vehicle’s navigation solution. In such fusion, the DVL provides the velocity vector of the AUV, which determines the navigation solution’s accu- racy and helps estimate the navigation states. In our paper we proposed BeamsNet, an end-to-end deep learning framework to regress the estimated DVL velocity vector that improves the accuracy of the velocity vector estimate, and could replace the model-based approach. Two versions of BeamsNet, differing in their input to the network, are suggested. The first uses the current DVL beam measurements and inertial sensors data, while the other utilizes only DVL data, taking the current and past DVL measurements for the regression process. Both simulation and sea experiments were made to validate the proposed learning ap- proach relative to the model-based approach. Sea experiments were made with the Snapir AUV in the Mediterranean Sea, collecting approximately four hours of DVL and inertial sensor data. Our results show that the proposed approach achieved an improvement of more than 60% in estimating the DVL velocity vector
The dataset was collected using the "Snapir" AUV in Mediterranean Sea. The Snapir is an A18-D, ECA GROUP mid-size AUV for deep water applications. Capable of rapidly and accurately mapping large areas of the sea floor, Snapir has a length of 5.5[m], a diameter of 0.5[m], 24 hours’ endurance, and a depth rating of 3000[m]. Snapir carries several sensors as its payload, including an interferometric authorized synthetic aperture sonar (SAS) and Teledyne RD Instruments, Navigator DVL.
Additional information regarding the dataset is located in the dataset folder.
To cope with the different input sizes, BeamsNetV1 architecture contains three heads. The first is for the 100 samples of the three-axes accelerometer, and the second is for the 100 samples of the three-axes gyroscope, operating simultaneously. The last head takes the DVL beam measurements. The raw accelerometer and gyroscopes measurements pass through a one dimensional convolutional (1DCNN) layer consisting of six filters of size 2 × 1 that extract features from the data. Next, the features extracted from the accelerometers and gyroscopes are flattened, combined, and then passed through a dropout layer with p = 0.2. After a sequence of fully connected layers, the current DVL measurement is combined and goes through the last fully connected layer that produces the 3×1 vector, which is the estimated DVL velocity vector. The architecture and the activation functions after each layer are presented in Figure above.
The network’s input is n past samples of the DVL beam measurements. Same as for the BeamsNetV1 architecture, the input goes through a one-dimensional convolutional layer with the same specifications. The output from the convolutional layer is flattened and passes through two fully connected layers. After that, the current DVL measurement is combined with the last fully connected layer output and goes into the last fully connected layer that generates the output
A code for networks is provided and can be seen in the code folder with additional information.
- Download this git repo and run the codes in the code folder.
- The code is written in python and uses the PyTorch framework.
If you found the experimental DATA useful for your research or used the BeamsNet architectures and code, please cite our paper:
@article{COHEN2022105216,
title = {BeamsNet: A data-driven approach enhancing Doppler velocity log measurements for autonomous underwater vehicle navigation},
journal = {Engineering Applications of Artificial Intelligence},
volume = {114},
pages = {105216},
year = {2022},
issn = {0952-1976},
doi = {https://doi.org/10.1016/j.engappai.2022.105216},
url = {https://www.sciencedirect.com/science/article/pii/S0952197622003013},
author = {Nadav Cohen and Itzik Klein},
}