TinyML Sound and Vibration Anomaly Detection for Backup Generators
Shops, clinics, mosques and offices rely on diesel or petrol backup generators during power cuts, and they fail at the worst moment. Build a small sensor node that listens to a generator's sound and vibration, learns its normal behaviour on the device itself with a tiny anomaly detection model, and sends an alert when something sounds wrong, without a cloud connection or a data scientist.
Intermediate
A small team, or one strong student willing to learn something new.1 to 2 semesters · 2 students
7 to ship
3 optional extrasSuggested stack
What you should ship
- A sensor node with microphone and accelerometer, mounted safely and protected from heat and dust
- Feature extraction on the device (spectrograms or band energies) and a small autoencoder or Gaussian model that learns 'normal' from the first hours of running
- On-device anomaly scoring with a threshold set from collected normal data
- Alerts by SMS, WhatsApp-style notification via a gateway, or a local buzzer and LED when there is no network
- A training and evaluation pipeline on a public machine sound dataset, plus your own recordings
- Evaluation with controlled faults you can safely create (loose mount, blocked air intake, low oil warning state, unbalanced load) on a generator you have permission to use
- A measurement of memory, inference time and power on the microcontroller
If you have time left
- Run-hour logging and fuel consumption estimates from vibration
- LoRa link for sites with many generators
- A small dashboard for a maintenance contractor
The problem: the generator fails during the power cut
Where load-shedding is routine, a backup generator is essential: for a pharmacy fridge, a clinic, a small factory or a home. Most are maintained when someone notices a strange noise, which is often too late. Industrial plants use vibration monitoring, but those systems are built for large machines and large budgets.
Experienced mechanics can hear a problem. A tiny model on a cheap microcontroller can learn to do a rough version of the same thing, for one machine at a time.
What already exists and the gap
- Industrial condition monitoring products from large automation companies monitor vibration on motors and pumps.
- Edge Impulse, now part of Qualcomm, and TensorFlow Lite for Microcontrollers make it possible to run small models on microcontrollers.
- Public datasets such as MIMII (malfunctioning industrial machine sounds) and the DCASE challenge's anomalous sound detection task support research.
The gap is a low-cost, self-learning node for small generators, which learns each machine's own normal sound on the device, works without internet, and alerts in a way a shop owner understands.
Inspiration
The DCASE anomalous sound detection task trains models only on normal sounds, because fault recordings are rare. That is exactly the situation with a single generator in a single shop.
What you build
- The sensor node and its mounting.
- On-device learning of normal behaviour and anomaly scoring.
- Alerts that work with and without a network.
The technically interesting part
- Learning on the device. Training fully on a microcontroller is hard; a practical approach is a fixed feature extractor with simple statistics or a small model fine-tuned on device. Compare options.
- Changing conditions. Load, temperature and fuel level change the sound. Separate normal variation from faults.
- Tight limits. Fit features and model into a few hundred kilobytes of RAM and measure it.
How you will evaluate it
- Detection rate and false alarms on public data and on your controlled faults.
- False alarms per week during normal running.
- Inference time, memory and power on the board.
Path to a product
Generator dealers and maintenance contractors could offer the node as part of a service plan. The same approach applies to water pumps, cold room compressors and other small machines.
Scope warning
Generators are hot, loud and dangerous, and exhaust fumes can kill. Work only with permission, with the owner present, outdoors or in a well-ventilated space, and never create faults that risk damage or fire. Do not touch the electrical output.
Learn what you need on Sefism
- Embedded and IoT roadmap: microcontrollers, sensors and connectivity.
- Embedded and IoT cheat sheet: quick reference for sensors, buses and power.
- Machine learning roadmap: anomaly detection and evaluation basics.
- ESP-IDF programming guide: building firmware for the ESP32.
- Arduino documentation: if you choose an Arduino board.
- How to become an embedded systems engineer: where this project can take you.
Get help with this idea
The risk is a model that alarms all day or never at all. Tauseef can help you choose the hardware, design safe fault tests and set thresholds sensibly. If you are new here, create a free account first.
Ideas and guidance, not finished projects
These are project ideas and scoping guidance, published free for students to use as a starting point. I do not build, write, or sell final-year projects, and I do not complete coursework for anyone. Take an idea, make it yours, and build it.
Not sure this one fits you?
I help students pick an idea that matches their skills and their deadline, then scope it down until it is finishable. Guidance only, never done for you.
Stuck on something specific?
Writing only gets you so far. If you want an answer to your situation rather than the general case, book a session and we will work through it together. Sessions are free for approved Sefism members, and a few slots open each week.
Follow along
New writing, resources and project ideas land here first.