Machine sound can reveal changes in friction, looseness or operating rhythm, but a microphone hears the whole room: people, tools, echoes and neighboring equipment. An ESP32 acoustic monitor should capture the smallest useful window, retain machine context and report evidence a technician can review. It should not quietly become an always-on room recorder or claim a fault from one loud event.
Digital MEMS microphones may expose I2S or PDM interfaces, but support depends on the ESP32 target, peripheral instance and ESP-IDF driver mode. Check the exact chip documentation and microphone timing/format; PDM-to-PCM conversion support is not uniform across targets. Verify channel selection, sample width, gain and clock before collecting a baseline.
For a fixed machine, the goal may be detecting a repeatable tonal or broadband change, not intelligible speech. Select sample rate and window length from the frequency range and event duration of interest, then test aliasing and clock stability. Use DMA-backed bounded buffers, track dropped samples, and tag every window with sample rate, sensor configuration, mounting position, machine ID, load/speed and a monotonic sequence.
Mount the microphone consistently and shield it from direct airflow or handling noise when appropriate. Measure the healthy machine across normal shifts, startup, shutdown and load changes. A baseline recorded in a quiet lab will not represent a noisy production floor.
Start with window RMS, peak level, crest factor, zero-crossing rate and energy in selected frequency bands. A spectrogram or short-time Fourier transform can show how spectral energy changes over time; do not confuse a visually bright bin with a validated fault signature. Normalize carefully and retain units or clearly label dimensionless features. Compare feature distributions by operating state before creating a score.
Edge processing can send features rather than raw PCM, reducing network use and exposure. If a raw snippet is needed for review, capture only after a defined trigger, keep it short, encrypt transfer, restrict access and apply a documented retention period. Avoid collecting ambient speech; evaluate applicable workplace privacy and consent requirements before deployment.
Thresholds should account for machine speed, load, distance, enclosure, microphone aging and room acoustics. A rule might require several consecutive windows outside a state-specific baseline band, with hysteresis before clearing. Evaluate false alarms per operating hour and missed events on held-out runs. Keep a human confirmation path and label alerts as “unusual acoustic pattern” until a qualified investigation establishes meaning.
| Feature | What it can show | Important confounder |
|---|---|---|
| Window RMS | Overall level trend | Microphone gain, distance and room noise |
| Band energy | Change in selected frequency region | Speed shifts and nearby tonal sources |
| Spectral centroid | Movement in spectral brightness | Microphone response and broadband noise |
| Clip/gap counters | Acquisition quality | Never classify corrupted audio as healthy |
Log capture success, sample gaps, feature version, threshold version, machine state, alert count and last-seen age. A dashboard should show feature trends and missing windows, not replay audio by default. Keep device identity and access controls separate from machine labels where practical. If the network is offline, bound local buffering and retain only the minimum needed for the agreed investigation window.
Begin with a repeatable normal baseline, then use safe, controlled lab variations and independently label each run. Compare microphone placement changes and environmental noise so the model learns the asset rather than the room. Do not intentionally damage production machinery or use an acoustic score as a safety interlock. Document the exact assets and conditions tested, and require maintenance expertise to interpret any suspected fault.
An ESP32 microphone can support useful edge observability when the sampling path, context, privacy and false-alarm budget are designed together. Capture short windows, validate signal quality, compare like operating states and send compact evidence. The output is a reviewable change signal, not a machine diagnosis.