AI Feature Library — Overview

The AI Feature Library is a set of 26 analyses whose purpose is not to produce a picture for a report, but to produce numbers a machine can learn from. Every analysis turns a recording into a compact, reproducible descriptor — a coefficient matrix, a condition indicator curve, a fault frequency set or an embedding vector — that can be exported and fed directly into a classifier, an anomaly detector or a production test limit.

The library is organized in four groups: feature extraction (MFCC, Log-Mel, spectral contrast), time-domain condition indicators, machine diagnostics (spectral kurtosis, Kurtogram, bearing frequencies, HNR, onset) and the AI layer (anomaly score, ONNX embeddings, cyclic spectral coherence).

The 26 Analyses at a Glance

Most analyses exist in two forms: one along the time axis and one along the RPM axis. Both are listed as separate entries in the Insert catalogue.

Feature extraction

Insert entry	Output	vs. RPM variant	Described in
MFCC	Colormap, coefficient × time (plus Δ and ΔΔ maps)	Yes	MFCC & Log-Mel
Log-Mel Spectrogram	Colormap, mel band × time	Yes	MFCC & Log-Mel
Spectral Contrast vs. Time	Colormap, sub-band × time, in dB	Yes	MFCC & Log-Mel

Time-domain condition indicators

Insert entry	Output	vs. RPM variant	Described in
Shape Factor	Curve, dimensionless	Yes	Time-Domain Indicators
Impulse Factor	Curve, dimensionless	Yes	Time-Domain Indicators
Margin Factor	Curve, dimensionless	Yes	Time-Domain Indicators
Energy Entropy	Curve, 0…1	Yes	Time-Domain Indicators

Machine diagnostics

Insert entry	Output	vs. RPM variant	Described in
Spectral Kurtosis (Antoni)	Curve, kurtosis vs. frequency	—	Bearing Diagnostics
Kurtogram	Colormap, level × frequency + optimal band	—	Bearing Diagnostics
Bearing Fault Features	Envelope spectrum + BPFO/BPFI/BSF/FTF harmonics + alarm	—	Bearing Diagnostics
Harmonic-to-Noise Ratio	Curve, dB	Yes	Bearing Diagnostics
Onset Detection	Novelty curve + onset count and times	—	Bearing Diagnostics

Rotating machinery

Insert entry	Output	vs. RPM variant	Described in
Fundamental Frequency (YIN)	Curve, F0 in Hz	—	Rotating Machinery
RPM from Sound	Curve, estimated rpm	—	Rotating Machinery
Gear TSA Indicators	TSA waveform over 360° + FM0/FM4/NA4*/ER/SBI	—	Rotating Machinery

AI layer

Insert entry	Output	vs. RPM variant	Described in
AI Anomaly Score	Score curve + threshold alarm	—	AI Layer Features
AI Embedding (ONNX)	Embedding vector (128…2048 dimensions)	—	AI Layer Features
Cyclic Spectral Coherence	Colormap, cyclic frequency × carrier frequency	—	AI Layer Features

Counting the vs.-RPM entries separately, the library contributes 26 entries to the Insert catalogue: 14 in feature extraction and time-domain indicators, 6 in machine diagnostics, 3 in rotating machinery and 3 in the AI layer.

vs. Time and vs. RPM

The vs. Time form frames the signal into blocks along the time axis and reports one value (or one column) per block. The vs. RPM form instead places one block at every RPM step crossing, so the feature can be read directly against machine speed during a run-up or coast-down.

All vs.-RPM analyses obtain speed the same way: the pulse-decoded speed is used when a pulse channel is present; otherwise a tacho / RPM channel is resolved from the loaded channels. The step width is set with RpmStep (default 50 rpm) and the direction with SlopeMode (AutoDetect, Runup, Rundown, All).

Verification Stars in the Insert Catalogue

Each entry in the Insert catalogue carries a star rating. The stars describe the verification level that the analysis has reached — they let you decide how much weight to put on an absolute number versus a relative trend.

Rating	Meaning	Typical use
★★★★★	Closed-form definition, verified against analytically known expected values on synthetic signals (square wave, sine, impulse train).	Absolute values may be used directly, e.g. as production test limits.
★★★★	Mathematical core verified on synthetic signals with known behaviour; cross-checks against published reference implementations are still in progress.	Excellent for trends, comparisons and machine learning; pin the analysis parameters before defining absolute limits.
(blank)	Not rated — typically because the result depends on a user-supplied model file.	Result quality is defined by the model you provide.

In this library the four time-domain indicators (shape / impulse / margin factor, energy entropy) and their vs.-RPM variants carry five stars; the remaining analyses carry four stars, except AI Embedding (ONNX), which is unrated because it runs a model that you supply.

Typical Uses

Feeding an AI model

Insert Log-Mel Spectrogram (or MFCC) on the channel of interest.
Keep the default parameters if the target model expects the common convention (25 ms frames, 10 ms hop, 64 mel bands, 125…7500 Hz); switch LogMode to ln for the training-domain scaling.
Export the resulting matrix and use it as model input, or run AI Embedding (ONNX) to obtain the embedding vector directly inside the software.

Machine diagnostics on a rolling bearing

Kurtogram → read the optimal demodulation band (centre frequency, bandwidth, level).
Bearing Fault Features → enter the bearing geometry and shaft speed; read which of BPFO / BPFI / BSF / FTF dominates the envelope spectrum.
Onset Detection and the impulse / margin factors → confirm that the signature is impulsive rather than a stationary tone.

Building a health index without labelled data

Record a set of known-good machines.
AI Anomaly Score with BaselineExportPath → store the baseline statistics.
On new recordings, load that baseline with BaselineStatsPath; the score curve and the threshold alarm then express deviation from the known-good population.

Where to Go Next

All analyses in this library work on ordinary time-domain channels — microphone, accelerometer or any other measured quantity. Nothing has to be prepared beforehand except a speed source for the vs.-RPM variants and for the gear and bearing analyses.