🩺 AI Sound Analyst & Health Manager for Industrial Assets

Team

Table of Contents

  1. Introduction
  2. Solution Architecture
  3. Software Designs
  4. Testing
  5. Conclusion
  6. Links

📖 Introduction

PERA-SAM (Predictive Equipment Reliability & Acoustics - Sound Analysis Manager) is a centralized acoustic management system designed to listen to the “heartbeat” of machines.

Traditional maintenance is reactive—fixing things only after they break. PERA-SAM shifts this to a predictive model. By processing acoustic signatures using FFT (Fast Fourier Transform) and MFCC, the system detects subtle frequency shifts caused by friction, imbalances, or wear before catastrophic failure occurs.

Currently prototyped for laptop cooling fans, server fans, engine fans, this system is designed to scale up to heavy industrial machinery and vehicle engines.

Solution Architecture

graph LR
    A["mimii_baseline<br/>(Research Lab)"] -->|"dataset + pickle"| B["Model/server<br/>(ML Backend API)"]
    B -->|"REST API :8000"| C["pera-sam<br/>(Frontend Web App)"]
    C -->|"Supabase Auth"| D["Supabase Cloud<br/>(User DB + Auth)"]

| Folder | Role | Tech Stack | |——–|——|————| | mimii_baseline/ | Original Hitachi research code + raw dataset storage | Python, Keras, librosa | | model/server/ | Production ML API — trains models, serves predictions | Python, FastAPI, TensorFlow, uvicorn | | pera-sam/ | Web dashboard — user login, upload audio, view results | React, Vite, TypeScript, TailwindCSS, Supabase |

Step-by-Step: What happens when run the system

sequenceDiagram
    participant User
    participant Frontend as pera-sam<br/>(React :5173)
    participant Backend as Model/server<br/>(FastAPI :8000)
    participant Dataset as mimii_baseline/<br/>dataset/
    participant Supabase as Supabase Cloud

    Note over Backend: SERVER STARTUP
    Backend->>Dataset: 1. Scan dataset folder for machine IDs
    Backend->>Backend: 2. Check if .h5 models exist in assets/
    alt Models missing
        Backend->>Dataset: 3. Read normal/*.wav files
        Backend->>Backend: 4. Extract mel-spectrograms → pickle cache
        Backend->>Backend: 5. Train autoencoder (50 epochs)
        Backend->>Backend: 6. Calibrate threshold (normal vs abnormal)
        Backend->>Backend: 7. Save model.h5 + metrics.yaml → assets/
    end
    Backend->>Backend: 8. Load all .h5 models into memory

    Note over User: USER INTERACTION
    User->>Frontend: 9. Open browser → Landing Page
    Frontend->>Supabase: 10. Login / Register (auth)
    Supabase-->>Frontend: JWT token
    User->>Frontend: 11. Go to Analysis page → upload .wav
    Frontend->>Backend: 12. POST /analyze (audio file)
    Backend->>Backend: 13. Preprocess audio → mel-spectrogram
    Backend->>Backend: 14. Auto-detect machine ID (lowest MSE)
    Backend->>Backend: 15. Compare score vs calibrated threshold
    Backend-->>Frontend: 16. JSON result (Normal/Warning/Anomaly)
    Frontend->>User: 17. Display health score, recommendation

Software Designs

Testing

Conclusion