AI Layer Features

The three analyses on this page sit one level above the individual descriptors. Instead of producing a single physical quantity, they combine many features into a decision-ready result: a scalar deviation from a known-good baseline, a learned embedding vector, or a map of the modulation structure hidden in a stationary-looking signal.

AI Anomaly Score

The anomaly score answers a question that has no textbook formula: does this machine sound different from the machines I know are good? It does so without any labelled fault data — only recordings of normal behaviour are required.

The 14-dimensional feature vector

Each analysis frame is reduced to fourteen features, eight from the waveform and six from the spectrum of the same frame:

The definitions are exactly those of the corresponding stand-alone analyses, so a high score can always be traced back to an interpretable quantity. Any feature that cannot be evaluated on a frame (for example a ratio on a silent frame) contributes 0 rather than invalidating the whole vector.

Score definition

The baseline stores a mean and a standard deviation per dimension. A frame is scored by its normalized distance from that baseline — a diagonal Mahalanobis distance averaged over the dimensions:

The normalization by the number of dimensions makes the score easy to read: a frame that sits exactly on the baseline scores 0, and a frame that deviates by exactly one standard deviation in every dimension scores exactly 1. A threshold of 3 therefore means “on average three sigma away from normal”, independent of how many features are used.

Each standard deviation is regularized to a small positive floor, so a feature that happens to be constant across the baseline recordings cannot blow the score up to infinity.

Baseline modes

Three ways of obtaining the baseline are supported, tried in this order:

Whichever mode is active, setting BaselineExportPath writes the baseline statistics to a JSON file, so a baseline derived from a good unit can be captured once and reused later through BaselineStatsPath.

Outputs and alarm

The result is a score curve over time. Its single value is the 95th percentile of the score, which is far more stable than the maximum and works well as a pass / fail criterion. The annotation additionally reports the maximum score and the moment at which it occurred. An alarm is raised when any frame exceeds AnomalyThreshold, and it states how many frames exceeded it and where the worst one was.

Parameters

AI Embedding (ONNX)

An embedding is a fixed-length vector produced by a pretrained neural network, in which acoustically similar recordings end up close together. It turns “does this sound like that” into a distance computation, and it lets a classifier be trained on a handful of examples instead of thousands.

Front end and model input

The analysis reuses the standardized Log-Mel front end with the training-domain scaling ln(mel + 0.01) fixed, computes the mel matrix for a channel and feeds it to the model given by OnnxModelPath. The first output of the model is flattened into the embedding vector.

The front-end parameters must match the convention the model was trained with. The defaults correspond to the widely used 25 ms / 10 ms, 64-band, 125…7500 Hz layout.

Input shape adaptation

Models declare their input differently — some expect frames first, some bands first, some carry a batch dimension, some leave dimensions dynamic. The analysis reads the shape declared by the model and fills it in: a dimension whose declared size equals the number of mel bands receives the bands, other fixed dimensions are kept as declared, and dynamic dimensions are filled from the back with the band count, then the frame count, then 1. If the model declares no shape at all, the tensor is built as [1, frames, mels].

Outputs and model handling

The embedding is displayed as a curve over the dimension index; its single value is the vector norm |v|, and the annotation states the dimensionality. Two vectors can then be compared by cosine similarity to find the nearest known sound.

No model file is shipped with the software. If OnnxModelPath is empty or the file does not exist, the analysis reports this and finishes without an error — the engine is ready and only waits for a model. Because the result is entirely determined by the model you supply, this entry carries no verification stars in the Insert catalogue.

Parameters

Cyclic Spectral Coherence

Many machine faults are invisible in an ordinary spectrum because their signature is not a line but a modulation: a high-frequency resonance whose amplitude is switched on and off at a low repetition rate. Cyclic spectral analysis separates the two frequencies — the carrier that is being excited and the cyclic frequency at which it is modulated — and shows them on two independent axes.

Cyclic modulation coherence

The signal is transformed by a strongly overlapping short-time Fourier transform. For each carrier line f, the sequence of frame powers et = |Xt(f)|² is the envelope of that line over time. Removing its mean and transforming it along the frame axis gives the modulation content, normalized by the mean envelope so the result is a relative modulation depth rather than an absolute level:

The factor 2 merges the positive and negative cyclic frequencies, so the value can be read directly as a modulation ratio. A stationary, unmodulated signal gives values close to 0; an amplitude modulated carrier produces a clear ridge at the modulation rate and weaker ridges at its harmonics.

Resolution and range

Because the modulation is observed at the frame rate of the short-time transform, the accessible cyclic frequency range and its resolution follow directly from the window and hop settings:

where K is the number of frames used, truncated to the largest power of two available so that the cyclic frequency grid is exact and no zero padding leakage occurs. This range comfortably covers the modulation rates of bearing and gear faults, which lie between a few hertz and a few hundred hertz. CscAlphaMaxHz can limit the displayed range further, which speeds up the analysis when only low modulation rates are of interest.

Reading the map

The result is a colormap with the cyclic frequency α on one axis and the carrier frequency f on the other:

• a vertical ridge at one α, spanning a wide carrier range, is the classical signature of a repetitive impact exciting a broadband resonance — read α and compare it with BPFO, BPFI, BSF or FTF from the bearing analysis;
• the carrier range covered by that ridge tells you which resonance is excited, and thus which band to demodulate;
• a flat map means no significant amplitude modulation is present.

The analysis also reports the cyclic frequency at which the coherence, averaged over all carrier lines, is highest — the dominant modulation rate of the recording. The first two cyclic bins are skipped in this search because they are dominated by slow level drift rather than by a genuine modulation.

Parameters

Applications

• End-of-line testing without fault samples — train the anomaly baseline on good units and reject on the 95th percentile of the score.
• Search by sound — compare embedding vectors by cosine similarity to find the recordings in an archive that sound most like the one at hand.
• Few-shot classification — train a small classifier on embeddings instead of on raw audio, using a handful of labelled examples per class.
• Modulation diagnosis — cyclic spectral coherence exposes repetitive impacts that remain hidden in a conventional spectrum, and points at both the repetition rate and the excited resonance.
• Feature stacking — the embedding vector and the anomaly score can themselves be used as inputs to a higher-level model.

© 2026 Deepsona. All Rights Reserved.