Unsupervised acoustic anomaly detection for industrial valves. Trained on normal
operating sounds only (no anomalous examples seen during training); anomalies are
detected as embeddings that fall far from the learned distribution of normal acoustic
signatures.
The model is a multiscale Transformer encoder trained on a pretext task: given a
log-mel spectrogram patch, predict which physical valve unit (machine ID) it came from.
This forces the encoder to learn fine-grained acoustic signatures characteristic of each
unit's normal operation. An ArcFace additive angular margin loss (s=32, m=0.10)
tightens the per-ID clusters on the embedding hypersphere during training.
After training, the classification head is discarded. What remains is a 128-dimensional
embedding function. At inference, the anomaly score is the Mahalanobis distance from
a clip's embedding to a single LedoitWolf-estimated Gaussian fitted over all training
embeddings pooled across machine IDs — this is what makes the deployed model usable
without knowing which physical unit produced a given recording.
Architecture:
Input: log-mel spectrogram (64 frames x 128 mel bins)
-> multiscale patch tokenisation: patch sizes [8, 4] frames, each scale gets its own
linear projection to 128-d + sinusoidal positional encoding, then concatenated
-> 2 x Transformer encoder block (4 heads, d_model=128)
-> mean pooling -> 128-d embedding
-> (training only) ArcFace head, s=32, m=0.10, over 4 machine-ID classes
Trained independently for 3 random seeds (0, 1, 2); the reference implementation
z-score-ensembles all three at inference for a more robust score (see Usage below).
Intended use
Primary use case: screening industrial valve acoustic recordings for anomalous
operating conditions (e.g. predictive maintenance), as a proof-of-concept / benchmark
model on the DCASE 2020 Task 2 dataset.
Not validated for: real-world deployment audio outside the DCASE 2020 / MIMII
valve distribution (different valve types, microphone placement, background noise
profiles, or SNR than the benchmark's synthetic 0 dB mixing). Treat scores from
out-of-distribution audio as unreliable without re-calibration.
Not included in this repo: fan and pump models. This repository currently ships
the valve model only.
How to use
This model uses custom Keras layers (ArcFaceLayer, multiscale patch tokenizer) that
are registered via @keras.saving.register_keras_serializable in the companion
iaad Python package — loading the .keras file with bare keras.models.load_model()
without that package installed will fail to resolve the custom layers. Install the
package from GitHub (not yet published to PyPI):
After training, a calibration pass fits a LedoitWolf covariance estimator over all
training-set embeddings (pooled across machine IDs) and records percentile thresholds
of the resulting Mahalanobis distances — this is what ships in valve_calib_s{seed}.npz
and lets inference flag a clip as anomalous (e.g. above the 95th percentile of normal)
without any refitting.
Evaluation results
Per-machine-ID mean AUC-ROC, averaged over 3 random seeds (chance = 0.500):
Machine
EMB pooled Mahalanobis
valve
0.947 ± 0.006
Pooled scoring is the deployment-safe metric (no machine ID needed at inference). See
the GitHub repo's evaluation docs
for the full comparison against earlier detector families (dense autoencoder,
Transformer-VAE) and the oracle (per-ID, ID-known) scoring variant.
Files in this repository
File
Description
version.json
{"emb_version": "v5", "run_tag": "v5.4"} — model family marker read by AnomalyDetector._load()
valve_emb_s{0,1,2}.keras
Trained EMBClassifier weights, one per seed
valve_norm_s{0,1,2}.npz
Per-mel-bin normalisation stats (mean, std) fit on that seed's training split
valve_calib_s{0,1,2}.npz
Calibration: pooled LedoitWolf location/precision + percentile thresholds
Limitations and biases
Trained and evaluated exclusively on the DCASE 2020 Task 2 synthetic 0 dB SNR mixing;
performance on cleaner or noisier real-world recordings is unverified.
Unsupervised training means the model has never seen a true anomaly label; its notion
of "anomalous" is entirely defined by the DCASE benchmark's specific injected fault
types, which may not cover all real failure modes.
Anomaly threshold (95th percentile by default) is a design choice trading recall for
precision — recalibrate on your own normal-operation data before deploying to a
different valve installation.
