Acoustic Anomaly Detection (TinyML)

Edge AI · MI-EAI-01

A microcontroller board with a MEMS microphone, mounted on one machine and trained on its normal sound, flagging abnormal sound on the device itself — no cloud, no streaming, under ₹10k.

Request a Quote All solutions

• ₹5,500pilot BOM · hardware only
• 3 – 5 daysto a live pilot
• Starterintegration level
• 3validated providers

Why the best fault detector in the plant is retiring

The problem

Compressors, gearboxes, fans and pumps sound different before they fail, but only to the one technician who has walked past them for fifteen years. Streaming audio to the cloud is expensive and usually forbidden; so the plant waits for the failure.

• Experienced ears retire; the knowledge leaves with them
• Audio streaming to the cloud is costly and often not allowed
• Vibration sensors are not always practical on every asset
• No early warning on the assets that are only 'listened to'

The solution

A small microcontroller board with an on-board MEMS microphone is mounted near the machine. In the pilot it records a few days of normal operation; a TinyML anomaly model is trained on that sound and deployed back to the board. The board then scores each second of audio locally and raises an alert (LED, relay or BLE/LoRa message) when the sound leaves the normal envelope. No audio leaves the device.

• Trained on the machine's own normal sound
• Inference on the microcontroller, no cloud, no audio streaming
• Alert by LED, relay contact or BLE/LoRa event
• Under ₹10k per monitored asset

Validated service providers

Validated by TwoElectrons and ranked by validation score: verified credentials, completed jobs on the platform and client ratings.

1. Sample Provider S — Bengaluru · On-prem LLM & RAG · ★ 4.9 · 12 jobs
2. Sample Provider T — Hyderabad · TinyML & edge inference · ★ 4.7 · 9 jobs
3. Sample Provider U — Pune · Secure AI infrastructure · ★ 4.6 · 7 jobs

Provider listings are samples pending live registrations.

Pilot BOM and cost

Reference hardware to run the pilot. Unit costs are indicative Indian market prices for the component class.

POC BOM cost only — hardware for a pilot. Installation, integration and provider services are quoted separately. Request for Quote

How on-device acoustic anomaly detection works

1. Mount. Fix the board in an enclosure near the machine; power by USB or battery.
2. Record. Capture a few days of normal operation across shifts.
3. Train. Build an anomaly model in Edge Impulse; deploy to the board.
4. Listen. Seed a fault or wait for a real one; review the alert log.

Pilot architecture

MCU board + MEMS mic (ESP32-S3 / nRF52840 class) → On-device TinyML model (anomaly score per second) → Alert output (LED · relay · BLE/LoRa) → Maintenance view (events only)

Only anomaly events leave the board; the audio itself is never stored or transmitted.

Business case: cost per asset, early warning, data egress

Ranges are typical figures reported for this class of solution; your pilot establishes the numbers for your plant.

• < ₹10k per monitored asset (the cheapest early warning available)
• 0 bytes of audio leave the device (privacy and bandwidth solved)
• Days to a trained model (from mounting to first alert)
• 3 – 5 days to a live pilot (one machine)

Who this is for: Maintenance and reliability engineers, Plants with tight data-egress rules, Compressor houses and HVAC plants, Gearboxes, fans, blowers and pumps, Utilities and remote pumping stations, Anyone who wants to try edge AI cheaply.

Illustrative case study

Illustrative scenario · not a client reference

A compressor house with four reciprocating compressors in a textile mill in Coimbatore (illustrative)

Set-up. One board per compressor in a vented enclosure, powered from the panel, three days of normal-sound recording, model trained and deployed, alerts over BLE to a gateway in the compressor house.

What happened. Compressor 3's anomaly score began rising nine days before a valve plate failure that the maintenance team then found on inspection and replaced during a planned stop. The mill added boards to its main blowers next.

• Pilot budget: ₹5,500 hardware
• Warning before failure: 9 days
• Audio transmitted: none

FAQ

Does it need to be trained on faults?

No — it learns what normal sounds like and flags deviation, so it works without any fault examples. Specific fault classes can be added later if you have recordings.

What about background noise from other machines?

The model learns the normal mix at the mounting position; the pilot checks that a neighbouring machine starting or stopping does not trigger false alerts and adjusts thresholds.

Can it run on battery?

Yes, for months at low duty cycle; mains or USB power is simplest for a pilot.

What does the ₹5,500 cover?

The board, enclosure, mount and power on a free-tier training workflow. Model training effort and provider services are quoted after a Request for Quote.