🛸 deep-space-sentinel

AI-powered anomaly detection and autonomous decision support system for deep space missions — combining Isolation Forest, One-Class SVM, PyTorch Autoencoder, and LSTM models on NASA's CMAPSS turbofan dataset to predict engine failures and trigger mission-critical autonomous responses.

---

📌 Overview

deep-space-sentinel simulates an onboard AI system capable of monitoring spacecraft engine telemetry in real time, detecting degradation anomalies, predicting Remaining Useful Life (RUL), and issuing autonomous mission decisions — from routine monitoring all the way to emergency shutdown protocols.

Built on NASA's CMAPSS FD001 dataset, this project explores how multi-model ensemble anomaly detection and deep learning can support autonomous decision-making in environments where ground-station communication latency makes human-in-the-loop responses impractical.

---

🗂️ Project Phases

Phase	Description
Phase 1	EDA, Feature Engineering, Rolling Statistics
Phase 2	Statistical Baseline, Isolation Forest, One-Class SVM, PyTorch Autoencoder, LSTM Classifier, LSTM RUL Regressor, Model Evaluation
Phase 3	Autonomous Decision Layer — Rule-Based Logic, AI Risk Scoring, Decision Confidence Estimation
Phase 4	Interactive Visualization Dashboards (Plotly + Matplotlib)

---

📊 Dataset

NASA CMAPSS Turbofan Engine Degradation Simulation Dataset (FD001)

• 100 training engines, 100 test engines
• 21 sensor readings + 3 operational settings per cycle
• Ground-truth RUL values provided for test set
• Failure threshold: RUL ≤ 30 cycles

🔗 Kaggle Link

Place the following files in the project root before running:

train_FD001.txt
test_FD001.txt
RUL_FD001.txt

---

🧠 Models

Anomaly Detection

Model	Approach
Z-Score Baseline	Statistical thresholding on normalized sensor means
Isolation Forest	Ensemble-based unsupervised anomaly detection
One-Class SVM	Kernel-based boundary learning on normal data
PyTorch Autoencoder	Reconstruction-error anomaly scoring with early stopping

Sequence Models

Model	Task
LSTM Classifier	Binary failure prediction from 30-cycle sensor sequences
LSTM Regressor	Remaining Useful Life (RUL) regression (capped at 125 cycles)

---

⚙️ Autonomous Decision Layer

The decision system integrates outputs from multiple models to produce a unified risk score (0–100) and a corresponding autonomous action:

Risk Level	Score Range	Action
LOW	0–19	Nominal Operations
MODERATE	20–39	Increase Monitoring Frequency
ELEVATED	40–59	Switch to Redundant Subsystem
HIGH	60–79	Alert Ground Station + Safe Mode
CRITICAL	80–100	Emergency Shutdown + Alert Earth

Decision confidence is also estimated based on inter-model agreement and margin from decision boundaries.

---

📈 Dashboards

Three interactive/static dashboards are generated:

• mission_control_dashboard.html — Risk scores, anomaly scores, risk level distribution, decision confidence (Plotly)
• live_telemetry_dashboard.html — Per-engine sensor trends and RUL curves (Plotly)
• model_summary.png — Comparative model performance bar charts (Matplotlib)

---

🛠️ Tech Stack

• Python 3.10+
• PyTorch — Autoencoder, LSTM Classifier, LSTM Regressor
• scikit-learn — Isolation Forest, One-Class SVM, metrics
• Pandas / NumPy — Data processing and feature engineering
• Plotly — Interactive mission dashboards
• Matplotlib / Seaborn — Static plots and model evaluation visuals

---

📋 Requirements

numpy
pandas
matplotlib
seaborn
scikit-learn
torch
plotly
jupyter