Prediction of Acoustic Communication Performance for AUVs using Gaussian Process Classification

Yifei Gao, Harun Yetkin, McMahon James, Daniel J. Stilwell

TL;DR

This paper introduces a data-driven approach to help underwater robots (AUVs) predict whether their acoustic messages will actually get through. By using Gaussian Process Classification instead of traditional signal-strength regression, the system learns from both successful and failed communication attempts, resulting in more accurate, less conservative "probability maps" for underwater coordination.

What Problem Does This Paper Address?

Underwater acoustic communication is notoriously finicky. Factors like multi-path interference, ambient noise, and distance lead to high bit-error rates. Currently, most AUV teams play it safe by assuming they can only communicate within a very short, "conservative" range.

While researchers have previously used Gaussian Process (GP) Regression to predict Signal-to-Noise Ratio (SNR), this method has a major flaw: it only learns from success. If a message fails to arrive, there is no SNR to record, so the model ignores the failure. Furthermore, regression requires a human-defined threshold (e.g., "anything above 10dB is a success") which may not hold true in different environments. The authors aim to create a model that explicitly accounts for failures and predicts the probability of success directly.

Methods

The authors move beyond standard regression to explore several classification-based techniques:

  • Gaussian Process Classification (GPC): Instead of predicting a continuous SNR value, the model predicts a binary outcome (Success/Failure). This allows the model to utilize "negative" data—the instances where a message was expected but never arrived.
  • Handling Noisy Inputs (NI-NN): A unique challenge in underwater robotics is that if a message fails, the receiver doesn't know exactly where the transmitter was at that moment. The authors adapted a "Noise-Input Neural Network" framework (inspired by NIMGPC) to account for this location uncertainty in the training data.
  • Sparse Variational GP (SVGPC): To handle large datasets without melting the AUV's onboard computer, they used "inducing points"—a subset of data that represents the whole—to keep the computational cost manageable.

Datasets

The study used three real-world datasets collected from a pair of Virginia Tech 690 AUVs during trials in Claytor Lake, Virginia:

  • DataSet1: Poor communication environment (72 successes, 144 failures).
  • DataSet2: Mixed environment (120 successes, 135 failures).
  • DataSet3: Favorable environment (311 successes, 103 failures).

The AUVs used WHOI micromodems at 25KHz and tracked their positions using inertial navigation, DVL (Doppler Velocity Log), and GPS.

Key Insights and Ideas

  • Failures are Data: In traditional regression, a failed message is a "missing value." In classification, it’s a critical data point that defines the boundaries of the communication map.
  • The Threshold Trap: Regression models are highly sensitive to the chosen SNR threshold. The authors found that a threshold optimized for one mission (e.g., 9.8dB) performed poorly on others. Classification bypasses this by learning the success probability directly from the raw outcomes.
  • Uncertainty Matters: By accounting for the fact that transmitter locations are "fuzzy" during failed attempts, the NI-NN method achieved the best performance, showing that modeling the noise in the navigation data actually leads to a better communication model.

Key Results

The researchers compared their classification methods against standard GP Regression (GPR) and Sparse Variational GPR (SVGPR).

  • Higher Accuracy: In almost all scenarios, classification methods (GPC, SVGPC, NI-NN) outperformed regression in predicting the "ratio" of correct communication attempts.
  • Better Confidence: The NI-NN method consistently achieved the lowest Negative Log-Likelihood (NLL), meaning it was not only more accurate but also more "certain" in its correct predictions.
  • Visual Superiority: Qualitative heatmaps showed that while regression models often predicted high SNR in "danger zones" (areas with many failures), the classification models correctly identified these as low-probability regions.
Metric (DataSet 2) GPR (Regression) GPC (Classification) NI-NN (Noisy Input)
Success Ratio 0.5667 0.6049 0.7068
NLL - 0.7503 0.6205

Limitations and Future Work

  • 2D Assumption: The current study assumes AUVs are operating in the same horizontal plane, ignoring vertical displacement and depth-related acoustic effects (like thermoclines).
  • Centralized Processing: The models were trained offline. The authors note that the next step is implementing these algorithms decentralized and in real-time on the AUVs themselves.
  • Computational Load: While sparse methods help, training these models on low-power underwater hardware remains a challenge that requires further optimization.

Summary updated

Abstract

Cooperating autonomous underwater vehicles (AUVs) often rely on acoustic communication to coordinate their actions effectively. However, the reliability of underwater acoustic communication decreases as the communication range between vehicles increases. Consequently, teams of cooperating AUVs typically make conservative assumptions about the maximum range at which they can communicate reliably. To address this limitation, we propose a novel approach that involves learning a map representing the probability of successful communication based on the locations of the transmitting and receiving vehicles. This probabilistic communication map accounts for factors such as the range between vehicles, environmental noise, and multi-path effects at a given location. In pursuit of this goal, we investigate the application of Gaussian process binary classification to generate the desired communication map. We specialize existing results to this specific binary classification problem and explore methods to incorporate uncertainty in vehicle location into the mapping process. Furthermore, we compare the prediction performance of the probability communication map generated using binary classification with that of a signal-to-noise ratio (SNR) communication map generated using Gaussian process regression. Our approach is experimentally validated using communication and navigation data collected during trials with a pair of Virginia Tech 690 AUVs.
BibTeX
@misc{gao2024prediction,
  author = {Yifei Gao and Harun Yetkin and McMahon James and Daniel J. Stilwell},
  title = {{Prediction of Acoustic Communication Performance for AUVs using Gaussian Process Classification}},
  year = {2024},
  eprint = {2411.07933},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.07933},
}
Figure