On-site wind speed recovery from smartphone audio: Time domain deep learning approach, laboratory validation and outdoor field test

Abstract

On-site wind speed information is paramount in various engineering domains. Conventional methods, reliant on expensive equipment and skilled operators, present challenges for practical uses. This study presents a novel time domain approach to recover wind speed from smartphone audio time series using deep learning. Leveraging the ubiquity of smartphones, the proposed approach offers a cost-effective and accessible alternative for wind speed measurement based on audio signals, allowing personnels to obtain on-site wind speed using smartphones. The model is built upon muti-layer architecture of one-dimensional convolutional neural network, and is trained and fine-tuned with high-fidelity wind tunnel laboratory data. The model’s performance is tested across different smartphone models and outdoor environments to assess its generalizability. The presented approach is generally effective, providing an overall accuracy with mean squared error of 0.2 to 0.9 (m/s)2. Although the model’s accuracy diminishes slightly in outdoor settings, it still shows favorable agreement with specialized wind gauges.