Engineered a 5,000 sample wind noise dataset and employed a CNN to estimate wind speed from their spectrograms. Gathered data using MEMS microphones and a cup anemometer as ground truth using ESP32. Recorded a mean-absolute error (MAE) of 1.12 m/s for wind speed
This project was developed under the CEN 598: Embedded Machine Learning course at Arizona State University during Fall 2023. The goal was to estimate wind speed using audio data from two closely spaced microphones on an ESP32 embedded system, leveraging machine learning to analyze wind noise correlation.
REPORT: Final Project Report
PRESENTATION: Project Presentation
- Meteorological Data Sparsity: Accurate weather predictions require denser data networks, which current Automatic Weather Stations (AWS) struggle to provide at scale due to cost and infrastructure.
- Low-Cost Wind Estimation: This project explores the feasibility of deploying microphone-based wind speed sensors in scalable, cost-effective solutions.
- Future Applications: Enabling weather-sensitive devices like smartphones or cars to include wind estimation as a feature.
-
Hardware Setup:
- ESP32 Microcontroller with I2S peripherals for audio data processing.
- INMP441 MEMs Microphones capturing wind noise across two channels.
- Hall-Effect Anemometer for ground-truth wind speed labels.
- DHT11 Sensor for temperature and humidity data collection.
-
Data Collection:
- Gathered over 5000 samples across 10 fan speeds, varying microphone directions for broader dataset coverage.
- Audio sampled at 1 kHz, 16-bit format.
-
Machine Learning Model:
- Spectrograms generated from audio samples to visualize frequency components below 51 Hz.
- Trained a Convolutional Neural Network (CNN) to predict wind speed.
- Samples were gathered using a controlled fan and multiple directions to explore classification potential.
- ESP32 managed stereo microphone inputs, streamed to a PC via serial communication.
-
Preprocessing:
- Data was normalized using min-max scaling.
- Spectrograms generated from both microphone channels.
-
Model Architecture:
- CNN with dropout layers to address overfitting.
- Input: Paired spectrograms from the two microphones.
- Output: Estimated wind speed in meters per second.
-
Challenges:
- Difficulties in maintaining high audio sampling rates due to embedded system limitations.
- Overfitting observed in training, requiring further feature engineering.
- The model showed signs of learning from frequency differences between spectrograms at various wind speeds.
- However, results were inconclusive without further data and experimentation.
- Improve data collection efficiency and increase dataset size.
- Experiment with feature extraction techniques to enhance model learning.
- Explore classification tasks using directional microphone data.