MotorSense:
hear without engine distortion
Each motor produces a unique frequency signature when operating. Bearing wear, misfire, valve play, or rotor rod breakage will alter this signature before the fault escalates. MotorSense captures this change with frequency analysis and machine learning and warns the user before the engine stops.
How does it work?
From sound to spectrum,
from spectrum to stimulation
The mobile app captures 30 seconds of recording. The server converts the audio to single-channel WAV, scores the recording quality (duration, level, clipping), extracts the power spectrum and energy ratios in frequency bands using the Welch method, compares it with the reference profile of the selected engine. The result is not a definitive diagnosis; These are candidate findings that can be explained by their probability and recording quality.
Fault is a frequency
The firing frequency of a 4-stroke engine is revolutions/60 × cylinder/2: 26.7 Hz in a 4-cylinder engine at idle (800 rpm). A stalling cylinder produces a rhythmic component below this frequency; High-frequency impacts on the valve clearance leave distinctive side bands on the worn bearing.
Warning, not diagnosis
Results are presented with probability and recording quality. At high risk, the user is told not to force the engine and to consult a specialist; The low quality recording will be requested again.
The record is processed and deleted
Audio files are processed in the temporary folder and deleted when the request is completed. For model development, it is only included in the review pool if the user explicitly consents and the record meets the quality/anomaly conditions.
Internal combustion and electric engines
Two engine types,
single approach
Today's MVP focuses on the internal combustion engine sound. The same signal processing backbone is being extended to electric motors with current and vibration sensors: early warning from the pump in the factory to the electric vehicle motor.
| engine | signal | Required signature | Example fault |
|---|---|---|---|
| internal combustion | Audio (phone microphone) | Ignition frequency and subharmonics, high frequency pulses | Misfire, valve play, knocking, belt noise |
| internal combustion | Vibration (sensor attached to the body) | Circuit dependent (order) analysis, crank angle environment | Bearing wear, imbalance |
| Electrical (asynchronous) | Motor current (clamp current sensor) | (1 ± 2s) f sidebands around grid frequency | Broken rotor bar, misalignment |
| Electricity (all) | Vibration (MEMS accelerometer) | Bearing characteristic frequencies (BPFO, BPFI, BSF) | Bearing outer/inner ring damage |
s: slip, f: grid frequency. Electric motor support is on the roadmap; In industrial use, continuous monitoring is aimed with a sensor end device (current + vibration).
Switching to machine learning
From rule-based to learning model
Today's classifier is rule-based and stands behind a fixed API contract; In this way, the application does not change when the model changes. The next step is to train a model using a combination of the mel-spectrogram and engine and rpm information, calibrate for each engine, add uncertainty and out-of-distribution (OOD) rejection, and set up a versioned model register.
The new tag given by the administrator does not go live directly: after enough service-verified records from the same tag accumulate and pass data quality control and model evaluation, it enters the new model version.
For whom?
- Vehicle owner: lifetime access to its engine; Get a record, see the risk, know before going to the service.
- Professional: Monthly packages of 10 or 20 engines; for services and fleets.
- Administrator: tags permissioned records, communicates with those granting communication permissions, and manages model versions.