Sound Levels Forecasting in an Acoustic Sensor Network Using a Deep Neural Network.

Area of Science:

• Environmental Science
• Acoustics
• Artificial Intelligence

Background:

• Wireless acoustic sensor networks are crucial for urban noise pollution monitoring.
• Existing noise prediction models often rely on long-term averages and auxiliary data.
• Short-term noise level forecasting is needed for effective urban planning and management.

Purpose of the Study:

• To propose and validate a Long Short-Term Memory (LSTM) deep neural network for short-term noise level prediction.
• To model the temporal behavior of sound pressure level and loudness level.
• To integrate predictive capabilities into wireless acoustic sensor network nodes.

Main Methods:

• Utilized a Long Short-Term Memory (LSTM) deep neural network.
• Trained models using one-minute equivalent sound levels from a two-month measurement campaign.
• Compared LSTM models against Auto-Regressive Integrated Moving Average (ARIMA) models.

Main Results:

• LSTM models demonstrated superior performance in predicting short-term sound levels compared to statistical models.
• Achieved a mean square error below 4.3 dB for sound pressure level predictions.
• Achieved a mean square error below 2 phons for loudness level predictions.

Conclusions:

• LSTM deep neural networks offer a robust solution for near-time noise level forecasting in urban environments.
• The proposed technique enhances the functionality of acoustic sensor networks, enabling early warning systems.
• Accurate short-term noise prediction supports proactive urban noise management strategies.