From audio to a decision in four steps
Analyze
The system receives the audio and computes a deterministic DSP report for each ten-second recording: waveform, FFT, Welch spectrum, spectrogram, band power, envelope and autocorrelation diagrams with their measurements.
Supply references
Each known condition contributes one report, from a recording chosen before testing. The unknown report carries no label.
Compare
A general multimodal model studies the unknown report beside every reference and cites evidence from both. One configuration adds Welch spectral distances.
Decide
The model returns the best-supported known condition, or different so that a person reviews the case.
Every recording is published with its diagrams, so you can listen to an unknown sample and each reference side by side.
Good references make good decisions
A reference should resemble its condition as it will be recorded: microphone position, speed and load. References are fixed before testing, by rules that use no test outcome.
When recordings with the same label behave very differently, the label covers several acoustic families or operating regimes, and each gets its own reference. Such differences can come from speed, load or mounting as well as from the fault itself.
On Jin, front-microphone references carried over to right-microphone recordings of the same four conditions. On Ottawa, references from operating profile 1 now need companions from the other profiles.
Recorded results
Each row is one configuration with its own counts. Publisher labels are the ground truth. In excluded-class tests the matching reference is withheld, so the right answer is different.
| Collection | Configuration | Correct class | Wrong class | False rejection | Correct rejection, class excluded | Wrong acceptance, class excluded |
|---|---|---|---|---|---|---|
| Jin: reports + distances | Reports plus Welch spectral distances | 11/12 | 0/12 | 1/12 | 2/4 | 2/4 |
| Ottawa: reports only | Reports only | 4/16 | 12/16 | 0/16 | Not tested | Not tested |
| Jin: calibrated replay | Calibrated distance rule (exploratory replay) | 9/12 | 0/12 | 3/12 | 4/4 | 0/4 |
The calibrated replay is an exploratory variant tuned on eight extra labeled windows. Nearest Welch distance alone also labels all 12 reserved Jin windows, and 2 of 16 Ottawa recordings; the next comparisons measure what the model adds beyond these numbers. Summary / all 68 recorded attempts
Six decisions worth inspecting
What we are improving
- References for every operating regime. By Welch distance, 14 of 16 Ottawa profile-2 recordings lie closer to another class's profile-1 reference; references from each profile and load address this directly.
- Separating close classes. The Jin magnet-fracture and tight-bearing references lie 3.04 dB apart, which explains the one missed magnet-fracture window and one wrong acceptance.
- Explanations that follow the numbers. In one excluded-class test the model accepted a reference 6.35 dB away on the strength of similar levels and a shared 16 kHz ridge.
- Balancing recognition and review. Calibrated thresholds set aside all four excluded-class windows together with the three magnet-fracture windows.
The current evidence comes from two public motor datasets, few acquisitions and recordings the project has already studied. Fresh recordings and blind listening comparisons are the next validation steps. Published scores from other models use their own sensors, splits and label budgets and serve as context.
Other directions explored
We also tested a wider set of 26 diagrams, the removal of selected diagram types and an adaptive class memory built from aggregated references. On 72 Ottawa acquisitions, fixed memory gave 16 correct, 11 wrong and 45 reviewed decisions; adaptive memory gave 18 correct, 41 wrong and 13 reviewed. The archive keeps every measurement and guides the current focus on references.