This repository contains a 4-layer IoT anomaly detection and predictive maintenance pipeline trained on real industrial sensor data collected via MQTT protocol.

The system predicts machine health degradation and detects anomalies from IoT sensor readings in real-time, designed for integration with Oracle AI Agents and MQTT brokers.

Models in This Repository

File	Model Type	Purpose
`rf_forecast_model.joblib`	Random Forest Regressor (300 trees)	Predicts future machine health score (0–100)
`scaler.joblib`	StandardScaler	Feature normalization (must be applied before prediction)

Pipeline Architecture

Layer 1: Isolation Forest        → Anomaly Detection & Scoring
Layer 2: Random Forest Regressor → Machine Health Score (0–100)  
Layer 3: Rule-based Engine       → Root Cause Explanation
Layer 4: RF Forecasting Model    → Future Health Prediction + Time-to-Critical

Input Features

The forecasting model (rf_forecast_model.joblib) requires these 8 features, in exact order:

#	Feature	Description	Unit
1	`temp_celsius`	Temperature reading	°C
2	`high_temp_duration`	Consecutive steps above 75°C	steps
3	`thermal_stress_code`	0=NORMAL, 1=HIGH, 2=CRITICAL	categorical
4	`degradation_rate`	Rate of health degradation (1–8x)	multiplier
5	`health_score`	Current machine health	0–100
6	`health_delta`	Change in health since last reading	points
7	`light_value`	Ambient light sensor reading	lux
8	`anomaly_score`	Isolation Forest anomaly score	float

How to Use

python

import joblib
import numpy as np
from huggingface_hub import hf_hub_download

# Download and load models directly from Hugging Face
model_path = hf_hub_download("HarshaMuralidharan04/iot-anomaly-detection", "rf_forecast_model.joblib")
scaler_path = hf_hub_download("HarshaMuralidharan04/iot-anomaly-detection", "scaler.joblib")

model = joblib.load(model_path)
scaler = joblib.load(scaler_path)

# Example: One machine reading
features = np.array([[
    72.5,   # temp_celsius
    3,      # high_temp_duration
    0,      # thermal_stress_code (NORMAL)
    1.0,    # degradation_rate
    85.2,   # health_score
    -0.5,   # health_delta
    210.0,  # light_value
    0.12    # anomaly_score
]])

# IMPORTANT: Always scale before predicting
features_scaled = scaler.transform(features)
predicted_future_health = model.predict(features_scaled)
print(f"Predicted health in 5 steps: {predicted_future_health[0]:.1f}/100")

Training Data

Source: Real MQTT sensor data from industrial IoT devices
Size: ~3,500+ timestamped sensor readings
Sensors: Temperature, Accelerometer (3-axis), Gyroscope (3-axis), Magnetometer (3-axis), Pressure, Humidity, Light
Split: 80% training / 20% test (time-aware split — no future data leakage)

Performance

Metric	Forecasting Model
MAE (Mean Absolute Error)	< 3.0 health points
R² Score	> 0.95
Training samples	~2,800
Test samples	~700

Intended Use

✅ Intended for:

Industrial equipment health monitoring
Predictive maintenance scheduling
Real-time anomaly alerting via MQTT / Oracle AI Agents

⛔ Not intended for:

Medical devices or safety-critical systems without additional validation
Consumer electronics (trained on industrial sensor ranges)

Authors

HarshaMuralidharan04 — IoT Anomaly Detection Project
Developed as part of an Oracle AI Agent + MQTT integration pipeline for industrial predictive maintenance.

License

MIT License — Free to use, modify, and distribute with attribution.