Environmentally adaptive acoustic transmission loss prediction in turbulent and nonturbulent atmospheres

Gordon Wichern, Mahmood R. Azimi-Sadjadi, Michael MungioleNeural Networks 2007 · DOI

Abstract

Predicting acoustic transmission loss (TL) in the atmosphere is a computationally demanding task, particularly when accounting for complex environmental factors like turbulence. Traditional physics-based models, while accurate, are often too slow for real-time applications. This paper introduces an "environmentally adaptive" system that utilizes a bank of specialized back-propagation neural networks (BPNNs) integrated with fuzzy logic. By partitioning the environmental state-space and employing a nonlinear fusion mechanism, the proposed system achieves high accuracy—comparable to gold-standard physics models—while offering a computational speedup of up to 50,000 times in complex turbulent scenarios.

Introduction

Accurate prediction of Transmission Loss (TL)—the reduction in sound intensity as it propagates from a source to a receiver—is critical for the effectiveness of Unattended Ground Sensors (UGS) in defense and environmental monitoring. The atmosphere is a highly dynamic medium where wind, temperature gradients, and turbulence significantly alter sound propagation, creating complex interference patterns and "shadow zones."

The standard for high-fidelity simulation is the Parabolic Equation (PE) model. However, the PE model is computationally expensive. As shown in the flowchart below, the process involves multiple iterative steps, including grid determination and marching the solution forward through range steps. In turbulent atmospheres, where results must be averaged over numerous random realizations to ensure statistical accuracy, the computational burden becomes prohibitive for real-time environmental awareness.

Flowchart of the traditional Parabolic Equation (PE) model for calculating transmission loss.

Methodology

To overcome the limitations of traditional physics models, this research proposes an "environmentally adaptive" architecture. The core strategy is a "divide and conquer" approach that specializes neural networks for specific atmospheric conditions.

Expert Predictors and Fuzzy Logic

The atmospheric state-space is partitioned into six distinct regimes based on physical parameters such as friction velocity (wind) and sound speed scale (temperature). For each regime, a dedicated BPNN "expert" is trained. Because atmospheric transitions are continuous rather than discrete, the system employs a Fuzzy Confidence Measure using Gaussian membership functions. This allows the system to determine the relative "trustworthiness" of each expert for any given set of input conditions.

System Architecture and Fusion

The outputs of these specialized experts are combined through a nonlinear Fusion Center, which is itself a neural network. This fusion layer blends the weighted predictions into a single, optimized TL estimate. To improve training efficiency, the researchers utilized a nondimensional PE model, reducing the input dimensionality from over 15 parameters to 10 key variables, including frequency, range, and ground resistivity.

Architecture of the proposed environmentally adaptive prediction system.

Results and Discussion

The system was evaluated using massive synthetic datasets generated by the PE model, comprising over 40,000 nonturbulent and 25,000 turbulent data pairs.

Accuracy and Reliability

The adaptive system achieved Root Mean Square (RMS) errors of 1.84 dB for nonturbulent conditions and 1.36 dB for turbulent conditions. Notably, the system performed better in turbulent environments. The authors attribute this to the "smoothing" effect of turbulence, which randomizes wave phases and fills in the sharp, erratic interference nulls that typically challenge neural network learning. Furthermore, the adaptive approach reduced the frequency of unreliable estimates (errors > 2 dB) by 24% to 27% compared to a single-network baseline.

Computational Efficiency

The most significant achievement is the dramatic reduction in computation time. While the speedup is approximately 10x for low-frequency, short-range signals, it scales exponentially with complexity. For high-frequency signals at long ranges in turbulent atmospheres, the neural network system is approximately 50,000 times faster than the traditional PE model.

Conclusion

The environmentally adaptive neural network system provides a robust solution for real-time acoustic transmission loss prediction. By combining specialized experts with fuzzy logic fusion, the model maintains the accuracy of high-fidelity physics simulations while enabling near-instantaneous results. Future work will focus on validating the model against real-world field data and exploring physics-informed neural network architectures to further enhance generalization across diverse atmospheric conditions.

Summary updated

BibTeX
@article{wichern2007environmentally,
  author = {Gordon Wichern and Mahmood R. Azimi-Sadjadi and Michael Mungiole},
  title = {{Environmentally adaptive acoustic transmission loss prediction in turbulent and nonturbulent atmospheres}},
  journal = {Neural Networks},
  year = {2007},
  doi = {10.1016/j.neunet.2007.04.025},
}
Figure