Proof of concept / acoustic condition recognition

ModelMetis recognizes machine condition from sound

ModelMetis records a machine, turns the audio into signal-processing diagrams and measurements, and lets a general multimodal model compare them with reference recordings of known conditions. It names the matching condition and cites the evidence, or sends the case to a person.

The proof of concept works on public motor datasets recorded at changing speeds, loads and microphone positions. A production test bench runs every unit at set-points chosen in advance, so its recordings are directly comparable from the start.

11/12reserved Jin motor recordings recognized, with references recorded in the same regime
40/72MAFAULDA cases recognized across speed changes with speed-normalized reports; 27/72 before normalization
8/18recognized at doubled speed after normalization; 2/18 before
1 recordingper known condition is the whole reference set: no model training

How the method works

Each decision follows five steps. Every step produces a file a person can open, so each answer can be traced back to the sound.

  1. Record at a known regime

    The machine runs at a set speed and load while a fixed microphone records ten seconds of sound.

  2. Normalize to the shaft

    When speeds differ, the audio is resampled to its own shaft rotation, so rotation-locked tones sit at the same order in every recording.

  3. Analyze

    A deterministic signal-processing report shows spectra, spectrogram, envelope, cycle average and their measurements, in hertz and in shaft orders.

  4. Compare

    A general multimodal model studies the unknown report beside one reference report per known condition and cites figures and numbers from both.

  5. Decide

    It names the best-supported condition, or answers different so that a person reviews the case.

The model receives evidence it can read: diagrams and measurements of known physical meaning. A new condition joins the system as one more reference recording, and a new machine type as one more reference set.

Built for controlled test benches

The method is strongest when the reference and the unit under test are recorded in the same conditions. Industrial benches provide exactly that: automated set-points, fixed microphones and a known operating sequence. Two deployments follow directly.

End-of-line quality assessment

  1. The motor arrives at an automated bench and is driven through regimes fixed in advance.
  2. Microphones in fixed positions record the noise at each regime.
  3. Each recording is compared with references of good units and of known defects recorded at the same regime.
  4. The bench returns pass, a named defect with its evidence, or review.

Vehicle diagnostics in service

  1. The vehicle arrives at the workshop and is placed in standard conditions, such as idle and set engine speeds.
  2. A fixed microphone setup records each condition.
  3. Recordings are compared with references of the same model in the same conditions.
  4. The technician receives the recognized condition with the diagrams behind it, or a case flagged for inspection.

A pilot needs one reference recording per condition and regime, the speed set-point or a tachometer signal, a fixed microphone position and a review step for different answers. Without a speed signal, the base frequency can be estimated from the audio: on MAFAULDA the estimate agreed with the tachometer within 3% for 797 of 798 accepted recordings.

Proof-of-concept results

The public datasets add difficulties that a controlled bench removes: references and test recordings come from different speeds, loads or microphone positions. Each row reports its own denominator, scored against publisher labels.

DataWhat changes between reference and testConfigurationResult
Jin motor recordings, 4 conditionsMicrophone position: front references, right-microphone testsReports plus spectral distances11/12 recognized; 2 of 4 cases without a reference correctly sent to review
MAFAULDA rig, 10 conditionsShaft speed: references at 1.0-2.0 times the test speedSpeed-normalized reports40/72 recognized; 27/72 with reports before normalization
MAFAULDA rig, same speedAnother acquisition and severity at the same speedSpeed-normalized reports13/18 recognized
Ottawa motors, 8 conditionsDrive frequency doubled from 15 to 30 Hz, plus loadReports only4/16 recognized; regime-matched references are the next step

MAFAULDA has 10 conditions, so guessing scores about 1 in 10. The model used is GPT-5.6 Sol with the same prompt and answer schema throughout.

Normalization keeps recognition stable when speed changes

Without a model, the nearest reference by spectrum shows the effect on 400 MAFAULDA test recordings. In hertz, recognition falls quickly once speeds differ; on the shaft-order axis it holds near half of all recordings even at doubled speed.

Nearest spectrum in hertzNearest spectrum in shaft orders, 0-100

At equal speed the hertz spectrum separates 246 of 387 recordings by itself: recordings taken in the same regime carry the condition clearly. A bench that fixes the regime keeps every test in that favorable row, and the model adds a cited explanation and the option to send a case to review.

Aggregate counts: speed-normalization record. Methods and every arm: speed study.

See and hear the evidence

Listen and compare48 recorded decisions: play each unknown recording beside every reference, switch among eight diagrams, read the explanation and the exact model input. Recorded resultsPer-configuration counts on Jin and Ottawa, six highlighted decisions and the full attempt record. MethodInputs, decisions, reference selection and speed normalization. Speed-normalization toolsAngular resampling with a known speed, and a base-frequency estimator for recordings without one. Results and scopeDenominators, distances and how each dataset was used. Other directions exploredDiagram selection, adaptive class memory and family separation, with their measurements.

Next step: a pilot on a real bench

The proof of concept used three public datasets with publisher labels and few machines per condition. A pilot records references at the bench's own set-points, runs a batch of units through the same sequence and compares the decisions with the existing quality outcome. Because each recording is reproducible at its regime, the pilot measures the method in the conditions it was built for.