Features

🧠 Implemented Machine Learning Approaches

MTSA currently integrates the following anomaly detection models:

Hitachi
A robust baseline model specifically designed for industrial anomaly detection tasks.
RANSynCoders
A state-of-the-art model enhanced with Mel-Frequency Cepstral Coefficients (MFCC) for improved acoustic feature representation.
GANF
A model that combines graph structures, recurrent neural networks (RNNs), and normalizing flows to perform anomaly inference, incorporating MFCCs to enhance performance on acoustic signals.
Isolation Forest
A tree-based ensemble method that isolates anomalies, adapted here with MFCC features.
OSVM (One-Class SVM)
A kernel-based approach for novelty detection, also enhanced with MFCC features.

MTSA includes a set of handcrafted statistical and signal-based features designed to enhance anomaly detection in acoustic signals:

M (Magnitude Mean) – Mean value of MFCC magnitudes over time.
S (Magnitude Std) – Standard deviation of MFCC magnitudes.
C (Correlation) – Correlation between MFCC coefficients, capturing inter-frequency relationships.

Root Mean Square – Measures the signal’s overall energy.
Square Root of Amplitude – Related to the average amplitude level.
Kurtosis – Captures the “tailedness” of the signal distribution.
Skewness – Measures asymmetry in the signal distribution.
Peak-to-Peak – Difference between the signal’s max and min values.
Crest Factor – Ratio of peak amplitude to RMS value.
Impulse Value – Ratio of peak amplitude to mean absolute value.
Margin Factor – Ratio of peak amplitude to root amplitude.
Shape Factor – Ratio of RMS to mean absolute value.
Kurtosis Factor – Modified kurtosis for enhanced anomaly sensitivity.
Frequency Center – Weighted average frequency indicating dominant energy region.
RMS Frequency – Root mean square of the frequency components.
Frequency Variance – Variability in the signal’s spectral distribution.