or4k2l/enhanced-audio-anomaly-detection
Hybrid ensemble (AST + classical) for industrial anomaly detection. Pump: 0.874 AUC (+7% vs baseline). Machine-specific strategy: Hybrid for Pump/Slider (0.87+), Classical for Fan/Valve (0.81+). DCASE 2020.
What's novel
An experimental machine-learning system for industrial audio anomaly detection using pretrained transformer embeddings and classical signal-processing features. See Reproducibility for supported environments, benchmark scope, and how to record a real-data run. In Cooperation whith AI.
Code Analysis
14 files read · 4 roundsIntegrates HuggingFace AST embeddings and librosa classical features into a 1723-dim vector, then applies standard sklearn anomaly detectors (GMM, OCSVM, XGBoost) for DCASE 2020 industrial audio anomaly detection.
Strengths
Clean, well-documented code with good type hints, lazy dependency loading, and a well-structured embedding_anomaly module (Mahalanobis/kNN/IsolationForest ensemble). The audio_anom package shows thoughtful error handling with NaN recovery, zero-division protection, and fallback mechanisms.
Weaknesses
The '955-dim' and '256-dim' feature vectors are achieved by padding/truncation rather than genuine fixed-size extraction, making the feature content inconsistent. The synthetic anomaly generation (3σ Gaussian noise) is scientifically unsound, the CAE is not actually convolutional, and two parallel code trees create confusion about which implementation is canonical.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
100
Contributors
2
Files
109
Active weeks
4
Repository
Language
Python
Stars
0
Forks
0
License
NOASSERTION