Phone-Based Audio Fault Detection

ISEF Category: Robotics and Intelligent Machines

Subcategory: Machine Learning  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: Full Year

The Hook

A fan can sound fine for days before it fails. That hidden warning is the whole challenge. Your phone microphone can catch tiny sound changes before your ears do. If you train a model to notice those changes, you can turn ordinary noise into an early fault alarm.

What Is It?

This project asks a simple question with a smart twist, can a computer learn the sound of a machine before humans hear anything wrong? You would record normal sounds from a desk fan, printer, or fridge compressor, then compare them with recordings from the same machine as it starts to wear out. The model looks for patterns in the audio, like a doctor listening for a tiny change in heartbeat rhythm.

The machine learning part is called self-supervised learning. That means the model learns structure from the recordings themselves, not from hand-labeled examples of every failure type. Contrastive learning pushes similar sound clips closer together and different clips farther apart. In plain language, the model learns what “normal” sounds like by comparing many clips, then flags clips that do not fit that pattern.

Why This Is a Good Topic

This is a strong science fair topic because you can measure a real signal, test a clear change, and compare your model against human hearing. The real-world link is clear too, early fault detection saves money, reduces waste, and can prevent breakdowns in home devices and small machines. You can learn data collection, feature extraction, model training, and evaluation without needing a professional lab.

Research Questions

• How does the type of machine, such as a fan, printer, or fridge compressor, affect anomaly detection accuracy?
• What is the effect of recording distance on the model's ability to detect early bearing wear?
• Does contrastive learning detect failure earlier than a baseline model trained on hand-crafted audio features?
• To what extent does background noise reduce early fault detection performance?
• Which audio representation, raw waveform, spectrogram, or mel-spectrogram, gives the best anomaly score separation?
• How does intentional wear progression change the model's confidence before a human notices a difference?

Basic Materials

• Smartphone with a microphone app that can export audio files.
• Desktop fan, printer, or another small machine with a rotating part.
• Second smartphone or laptop for backup recording and note-taking.
• Tripod, phone stand, or stack of books to keep the microphone position fixed.
• Tape measure to keep recording distance consistent.
• Notebook or spreadsheet for recording machine condition, date, and test notes.
• Basic safety gear, such as safety glasses if you plan to open or inspect hardware.

Advanced Materials

• Access to a laptop with enough storage for audio datasets.
• External USB microphone for comparison with phone recordings.
• Arduino or smart plug for synchronized machine state logging.
• Vibration sensor or accelerometer for optional multimodal comparison.
• Access to a small test motor or fan that can be worn intentionally under safe supervision.
• Labeling tool for sorting clips by machine state and recording session.
• Python environment for feature extraction, model training, and evaluation.

Software & Tools

• Python: Prepares audio files, extracts features, and trains anomaly detection models.
• Audacity: Cleans up recordings, trims clips, and checks for clipping or background noise.
• ImageJ: Not used for audio, so skip it unless you also analyze visual damage photos.
• Librosa: Converts audio into spectrograms and audio features for machine learning.
• Google Colab: Runs Python notebooks in a browser when your own computer is slow.

Experiment Steps

1. Define the fault you want to detect and the exact machine state you will compare.
2. Plan a recording setup that keeps microphone position, device settings, and background noise as constant as possible.
3. Build a clean dataset with enough normal clips and enough clips from each wear stage.
4. Choose one representation of sound to test first, then decide how you will convert clips into model inputs.
5. Set up a baseline method, then compare it with your self-supervised approach using the same test split.
6. Plan an evaluation metric that rewards early warning, not just final failure detection.

Common Pitfalls

• Changing microphone distance between sessions, which makes the model learn room geometry instead of bearing noise.
• Using clips that are too short or too long, which can hide the fault signature or mix in unrelated sound.
• Letting background noise from people, HVAC systems, or traffic dominate the recordings.
• Training and testing on the same machine session, which inflates accuracy and hides weak generalization.
• Calling every unusual sound a failure, which turns normal load changes into false alarms.

What Makes This Competitive

A strong version of this project does more than say, “the model worked.” It shows when the model starts to fail, how early it can warn, and whether it still works on a different machine of the same type. Better projects compare several audio representations, use a real baseline, and report false alarms, missed faults, and lead time before failure. A top-tier entry also explains why the signal changes, not just that the model found a pattern.

Project Variations

• Use a cheap box fan instead of a desk fan, then test whether the same model transfers across motor types.
• Compare phone microphone audio with a vibration sensor to see whether sound or motion gives earlier warning.
• Analyze one machine under different loads, such as idle, medium use, and heavy use, to separate normal variation from true wear.

Learn More

• PubMed: Search review articles on condition monitoring, bearing fault detection, and acoustic anomaly detection.
• NASA NTRS: Search technical reports on machinery health monitoring and signal-based fault detection.
• NOAA: Use background-noise resources and acoustic monitoring ideas to think about sound contamination in field recordings.
• MIT OpenCourseWare: Find free machine learning and signal processing course materials for feature extraction and model evaluation.
• USGS Earthquake Hazards Program: Explore how researchers detect small signals in noisy data, then adapt that thinking to audio anomalies.
• OpenReview: Search for recent peer-reviewed machine learning papers on self-supervised anomaly detection and contrastive learning.