Ion Mill Predictive Maintenance System

A full-stack predictive maintenance platform built using the PHM 2018 Data Challenge dataset. The system continuously ingests industrial ion mill telemetry, performs anomaly detection and Remaining Useful Life (RUL) estimation, stores telemetry in PostgreSQL, and visualizes equipment health through a React dashboard, and is fully containerized using Docker.

---

Live Dashboard Demo (After Running 4 Machines)

Dashboard Overview (Before Running Any Machine)

---

Overview

Industrial ion milling equipment generates large volumes of sensor telemetry during operation. Unexpected failures can lead to costly downtime, production losses, and increased maintenance costs.

This project combines machine learning, backend engineering, containerization, and data visualization to create an end-to-end predictive maintenance platform capable of monitoring multiple tools simultaneously and estimating equipment degradation in real time.

---

Key Features

Fleet Monitoring

• Monitor multiple ion mill tools simultaneously
• Compare equipment health across the fleet
• Identify tools requiring maintenance attention
• View real-time operational status

Anomaly Detection

• Isolation Forest based anomaly detection
• Real-time anomaly scoring
• Automated health score generation
• Early warning alerts for abnormal behaviour

Remaining Useful Life Prediction

• Time-to-Failure (TTF) estimation
• Machine-aware predictive model
• Fleet-wide degradation modeling
• Maintenance planning support

Telemetry Ingestion

• Concurrent multi-tool simulation
• Streaming industrial telemetry
• Historical sensor storage in PostgreSQL

Containerized Deployment

• Dockerized React frontend
• Dockerized Django backend
• Dockerized PostgreSQL database
• One-command deployment with Docker Compose

---

System Architecture

PHM 2018 Dataset
        │
        ▼
Multi-Tool Telemetry Simulator
        │
        ▼
Django REST API
        │
        ├── PostgreSQL Storage
        ├── Isolation Forest Anomaly Detection
        └── TTF Prediction Engine
        │
        ▼
React Fleet Dashboard
        │
        ▼
Maintenance Engineers

---

Technology Stack

Frontend

• React
• Vite
• Axios
• Recharts

Backend

• Django
• Django REST Framework

Database

• PostgreSQL

Machine Learning

• Scikit-learn
• Isolation Forest
• HistGradientBoostingRegressor
• Pandas
• NumPy

Deployment

• Docker
• Docker Compose

---

Dataset

This project uses the PHM 2018 Data Challenge dataset containing telemetry collected from industrial ion milling equipment.

Example telemetry signals include:

• Ion Gauge Pressure
• Flowcool Pressure
• Flowcool Flowrate
• Etch Beam Voltage
• Etch Beam Current
• Rotation Speed
• Source Usage
• Auxiliary Source Timers
• Process Duration

The dataset also provides:

• Failure labels
• Time-to-Failure targets
• Tool identifiers
• Process metadata

---

Machine Learning Pipeline

Anomaly Detection

An Isolation Forest model monitors incoming telemetry and identifies abnormal operating conditions.

Outputs:

• Anomaly Score
• Health Score
• Anomaly Flag

Remaining Useful Life Prediction

A fleet-wide HistGradientBoostingRegressor model predicts Time-to-Failure using sensor telemetry and machine-specific context.

Features used:

• Etch Source Usage
• Auxiliary Source Timers
• Etch Beam Current
• Flowcool Pressure
• Rotation Speed
• Step Duration
• Ion Gauge Pressure
• Etch Beam Voltage
• Rotation Angle
• Machine Identifier

Outputs:

• Predicted Time-to-Failure
• Remaining Useful Life Estimate

---

Model Evaluation

Several models were evaluated during development.

Model	Dataset Scope	MAE	Model Size
Random Forest (Baseline)	Single Tool	~1.29M	491 MB
Random Forest (Extended Features)	Single Tool	~295	8.9 GB
HistGradientBoosting	Single Tool	~7,337	726 KB
Production HistGradientBoosting	14 Tools	~38,896	730 KB

The final deployment model was selected based on scalability, inference speed, deployment efficiency, and fleet-wide generalization performance.

---

API Endpoints

Equipment List

GET /api/equipment/

Equipment Health

GET /api/equipment/{id}/health/

Recent Readings

GET /api/equipment/{id}/readings/

Fleet Health

GET /api/fleet-health/

Alerts

GET /api/alerts/

Sensor Data Ingestion

POST /api/sensor-readings/

---

Running with Docker

Build

docker compose build

Start

docker compose up

Stop

docker compose down

---

Current Capabilities

• Multi-tool monitoring
• Fleet health visualization
• Anomaly detection
• Health score generation
• Fleet-wide TTF prediction
• Historical telemetry storage
• Concurrent telemetry simulation
• PostgreSQL persistence
• Containerized deployment

---

Future Improvements

• Failure-specific prediction models
• Explainable AI dashboards
• Real-time streaming infrastructure
• Cloud deployment
• Authentication and access control
• Advanced maintenance scheduling
• Fleet-level analytics

---

Author

Aryan Sheikh

MIT Manipal

B.Tech Computer Science (AI & ML)