EchoGuard

Unsupervised Audio Anomaly Detection using Deep Learning

EchoGuard is a research-driven project focused on machine audio anomaly detection using unsupervised learning and signal processing techniques. Rather than treating anomaly detection as a black-box modeling problem, the project begins with a systematic investigation of machine acoustics to understand how faults manifest in audio recordings.

Current focus:

• Dataset exploration
• Signal processing analysis
• Acoustic feature investigation
• Unsupervised anomaly detection

---

Project Status

🚧 In Progress

Current stage:

• Dataset exploration completed
• Channel selection completed
• RMS Energy analysis completed
• Spectrogram analysis completed
• Spectral Centroid analysis completed
• MFCC analysis in progress
• Model development pending

---

Dataset

This project uses the MIMII (Malfunctioning Industrial Machine Investigation and Inspection) dataset.

MIMII is a benchmark dataset designed for machine condition monitoring and audio-based anomaly detection.

Machine Types

The dataset contains recordings from four industrial machine categories:

• Pumps
• Fans
• Valves
• Slide Rails Each machine category contains multiple machine IDs representing different physical machines.

---

Current Scope

The current analysis focuses exclusively on:

Machine Type: Pump
Machine ID  : id_00
Channel     : 0

Limiting the scope to a single machine allows detailed investigation of acoustic behaviour expanding to other machine types.

---

Recording Setup

The MIMII dataset was recorded using:

• TAMAGO-03 circular microphone array
• 8 microphone channels
• Industrial operating environment
• Normal and abnormal operating conditions Each recording contains approximately 10 seconds of machine audio

---

Why Channel 0?

The original recordings contains 8 microphone channels.

An exploratory analysis was performed to identify the most representative channel.

Channel 0 was selected because:

• It exhibited the highest minimum cross-correlation with the remaining microphones.
• It showed slightly higher RMS energy, indicating improved signal quality.
• Its spectral centroid occupied a central frequency range without strong directional bias.
• It is the reference microphone used in the original MIMII publication.

---

Research Questions

This project is driven by the following questions:

Q1: Can classical signal processing features explain why some anomalies are difficult to distinguish from normal machine sounds?

Q2: Why do certain abnormal recordings overlap with normal recordings in RMS Energy?

Q3: Do subtle anomalies exhibit different spectral characteristics despite having similar energy distributions?

Q4: Can unsupervised deep learning models learn these subtle differences automatically?

---

Key Observations So Far

RMS Energy

• A subset of abnormal recordings overlaps significantly with the normal RMS distribution.
• Simple energy thresholding is therefore insufficient.

Spectrogram Analysis

Two anomaly categories emerged:

• Obvious anomalies
• Subtle anomalies Some faults produce clear broadband disturbances, while others remain visually similar to normal recordings.

Spectral Centroid

• Mean spectral centroid provides limited separation.
• Temporal variability measures (Std and IQR) contains more useful information.

MFCC Analysis

Current hypothesis:

Temporal stability of cepstral features may be more informative than their avergae values when detecting subtle machine faults.

---

Repository Structure

EchoGuard/
│
├── README.md
├── requirements.txt
├── .gitignore
│
├── data/
│   |── .gitkeep
|   └── README.md
│
├── notebooks/
│   ├── 01_dataset_exploration.ipynb
|   ├── 02_channel_selection.ipynb
│   ├── 03_waveform_analysis.ipynb
│   ├── 04_rms_analysis.ipynb
│   ├── 05_spectrogram_analysis.ipynb
│   ├── 06_spectral_centroid_analysis.ipynb
│   ├── 07_mfcc_analysis.ipynb
│   ├── 08_feature_summary.ipynb
│   └── 09_model.ipynb
│
├── src/
│   ├── __init__.py
│   ├── utils.py
│   ├── features.py
│   ├── visualization.py
│   └── dataset.py
│
├── figures/
│
├── results/
│
└── research_notes/
    ├── Research_Note_01.md
    ├── Research_Note_02.md

---

Planned Work

• Complete MFCC investigation
• Feature correlation analysis
• Log-Mel spectrogram exploration
• Autoencoder baseline
• Normalizing Flow-based anomaly detection
• Comparative evaluation
• real-world machine audio validation

---

References:

MIMII Dataset:

• Purohit et al., "MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection"

---