✈ S.P.E.C.T.R.A

Sensor Prognostics & Edge Computing for Telemetry-based Real-time Analysis

Intelligent Edge Computing for UAV Predictive Maintenance in Real Time

A real-time Prognostics and Health Management (PHM) system for autonomous UAVs, powered by BiLSTM+Attention deep learning models and an agentic decision layer for predictive maintenance.

---

🚀 Overview

S.P.E.C.T.R.A is an end-to-end UAV health monitoring system that processes live telemetry from three critical sensor subsystems — IMU (Inertial Measurement Unit), ATT (Attitude), and BAT (Battery) — through deep learning models to predict component failures before they happen.

The system features a stunning real-time web dashboard that visualizes sensor health, degradation trends, Remaining Useful Life (RUL), and automated maintenance recommendations through an intelligent agentic decision layer.

Key Capabilities

Feature	Description
🧠 BiLSTM + Self-Attention	Bidirectional LSTM with additive attention for temporal anomaly detection
📊 Real-time Dashboard	Flask + SSE-powered live monitoring with smooth gauge animations
⚡ Agentic Decision Layer	Context-aware fault diagnosis with dynamic urgency classification
🔄 Health Index Tracking	EMA-based normalized health indices with degradation rate monitoring
🛡️ Fault Injection Engine	Simulated motor degradation and power stress for validation
📈 RUL Estimation	Linear extrapolation-based Remaining Useful Life computation

---

🎮 Live Demo

▶ Launch S.P.E.C.T.R.A Demo on GitHub Pages

The demo replays pre-recorded inference data from the BiLSTM+Attention models with the same smooth dashboard animations as the live system.

---

🏗️ System Architecture

┌────────────────────────────────────────────────────────┐
│                   SENSOR DATA LAYER                    │
│  IMU.csv (Acc/Gyr)  │  ATT.csv (Angles)  │  BAT.csv   │
└──────────┬───────────┴──────────┬─────────┴────┬───────┘
           │                     │               │
     ┌─────▼─────┐        ┌─────▼─────┐   ┌─────▼─────┐
     │  Feature   │        │  Feature   │   │  Feature   │
     │Engineering │        │Engineering │   │Engineering │
     │ 10 feats   │        │ 11 feats   │   │  9 feats   │
     └─────┬──────┘        └─────┬──────┘   └─────┬──────┘
           │                     │                 │
     ┌─────▼─────┐        ┌─────▼─────┐   ┌─────▼─────┐
     │  BiLSTM   │        │  BiLSTM   │   │  BiLSTM   │
     │+Attention │        │+Attention │   │+Attention │
     └─────┬──────┘        └─────┬──────┘   └─────┬──────┘
           │                     │                 │
     ┌─────▼─────────────────────▼─────────────────▼─────┐
     │              HEALTH INDEX ENGINE                   │
     │  EMA Smoothing → Degradation Rate → RUL → Status  │
     └──────────────────────┬────────────────────────────┘
                            │
     ┌──────────────────────▼────────────────────────────┐
     │           AGENTIC DECISION LAYER                  │
     │  Severity Ranking → DR Tiebreak → Dynamic Faults  │
     │  Context-Aware Actions → Confidence Scoring        │
     └──────────────────────┬────────────────────────────┘
                            │
     ┌──────────────────────▼────────────────────────────┐
     │          S.P.E.C.T.R.A LIVE DASHBOARD             │
     │  Flask + SSE → Smooth Gauges → HI Chart → RUL Bar │
     └───────────────────────────────────────────────────┘

---

📂 Project Structure

S.P.E.C.T.R.A/
├── dashboard.py         # 🖥️  Live monitoring dashboard (Flask + SSE + embedded HTML)
├── final_model.py       # 🧠  Main PHM engine — BiLSTM+Attention + Agentic layer
├── lstm_imu.py          # 📡  Standalone IMU model training & evaluation
├── lstm_att.py          # 🧭  Standalone Attitude model training & evaluation
├── lstm_bat.py          # 🔋  Standalone Battery model training & evaluation
├── IMU.csv              # 📊  IMU sensor telemetry dataset
├── ATT.csv              # 📊  Attitude sensor telemetry dataset
├── BAT.csv              # 📊  Battery sensor telemetry dataset
├── index.html           # 🌐  GitHub Pages static demo (pre-recorded data)
├── spectra_architecture.png  # 📋  System architecture diagram
├── requirements.txt     # 📦  Python dependencies
└── README.md            # 📖  This file

---

⚡ Quick Start

Prerequisites

• Python 3.10+
• pip

Installation

# Clone the repository
git clone https://github.com/ansu647/SPECTRA-UAV-PHM.git
cd SPECTRA-UAV-PHM

# Install dependencies
pip install -r requirements.txt

Run the Live Dashboard

python dashboard.py

This will:

1. Load sensor data from CSV files
2. Load/train BiLSTM+Attention models (cached as .pt files after first run)
3. Launch a Flask server at http://localhost:8050
4. Auto-open the S.P.E.C.T.R.A dashboard in your browser
5. Stream real-time inference results with smooth animations

Run the Terminal-Based PHM Engine

python final_model.py

Outputs an ANSI-colored terminal dashboard with live health tracking, RUL estimation, and maintenance recommendations.

Train Individual Sensor Models

python lstm_imu.py    # Train IMU model
python lstm_att.py    # Train Attitude model
python lstm_bat.py    # Train Battery model

---

🖥️ Dashboard Features

Component	Description
Circular Gauges	Per-sensor health index with WARNING/CRITICAL threshold markers
Health Chart	Real-time 3-line chart tracking all sensor HI values
RUL Bar	Composite worst-case Remaining Useful Life with shimmer animation
Agentic Panel	Dynamic fault detection, priority sensor, fastest-degrading sensor
Per-Sensor Strip	Live HI + degradation rate snapshot for all subsystems
System Status	Animated NORMAL → WARNING → CRITICAL badge with glow effects
Trend Badge	STABLE / DEGRADING system-wide trend indicator

---

🧠 Model Architecture

BiLSTM + Self-Attention

Input (seq_len=15, features=N)
       │
  ┌────▼────┐
  │ BiLSTM  │  2 layers, hidden=64, dropout=0.3
  │ (→ + ←) │  Output: (batch, seq, 128)
  └────┬────┘
       │
  ┌────▼──────────┐
  │Self-Attention │  Additive (Bahdanau) attention
  │  W·tanh(V·h)  │  Learns temporal importance
  └────┬──────────┘
       │
  ┌────▼────┐
  │   FC    │  128 → 64 → output_dim
  │ Layers  │  ReLU + Dropout
  └─────────┘

Engineered Features per Sensor

Sensor	Features	Count
IMU	AccX/Y/Z, GyrX/Y/Z, vibration magnitude, ZCR(Acc), ZCR(Gyr), jerk	10
ATT	Roll, Pitch, Yaw, DesRoll, DesPitch, DesYaw, roll_err, pitch_err, yaw_err, total_track_err, Δyaw	11
BAT	Volt, Curr, CurrTot, EnrgTot, Temp, power, SoC, ΔVolt, ΔTemp	9

---

🔧 Configuration

Key parameters in final_model.py:

Parameter	Default	Description
SEQ_LEN	15	LSTM input sequence length
EMA_ALPHA	0.08	Health Index smoothing factor
EPOCHS	15	Training epochs
BATCH_SIZE	32	Training batch size
FAULT_INJECT	True	Enable/disable fault simulation

Dashboard config in dashboard.py:

Parameter	Default	Description
STEP_DELAY	0.55s	Interval between inference steps
REPLAY	True	Loop dashboard after completing all steps
PORT	8050	Flask server port

---

📊 Health Status Classification

Status	Condition	Action
✅ NORMAL	HI < 1.20, DR < 0.003	Routine monitoring
⚠️ WARNING	HI > 1.20 or (HI > 1.10 & DR > 0.003)	Schedule inspection
🚨 CRITICAL	HI > 1.55 or (HI > 1.35 & DR > 0.006)	Immediate grounding

---

🛠️ Tech Stack

• Deep Learning: PyTorch (BiLSTM + Self-Attention)
• Data Processing: Pandas, NumPy, scikit-learn
• Dashboard Backend: Flask + Server-Sent Events (SSE)
• Dashboard Frontend: Vanilla HTML/CSS/JS + Chart.js + Canvas gauges
• Design System: Orbitron + Inter + JetBrains Mono fonts, glassmorphism UI

---

📜 License

This project is licensed under the MIT License — see the LICENSE file for details.

---

Built with 🧠 Deep Learning & ✈️ Aviation Engineering

S.P.E.C.T.R.A — Keeping UAVs flying safe, one prediction at a time.