🎯 Sound-Based Drone Fault Classification – Multi-Task Convolutional Neural Network

📌 Overview

This project addresses the classification of drone faults using sound recordings from multiple microphones.
Instead of building three separate models for:

1. ⚙ Fault type
2. 🧭 Maneuvering direction
3. 🚁 Drone model type

We implement a single Multi-Task Learning (MTL) Convolutional Neural Network (CNN) that learns all three tasks simultaneously.

---

🚀 Why Multi-Task Learning?

Traditional ML pipelines train one model per task:

• 🛠 A model for fault classification
• 🛠 A model for direction classification
• 🛠 A model for drone model classification

❌ Drawbacks

• Redundant feature extraction
• Higher compute cost
• No benefit from shared knowledge between tasks

✅ Our Solution

• 📡 Shared CNN feature extractor
• 🎯 Three task-specific heads
• 🔄 Knowledge transfer across tasks for better generalization

---

📂 Dataset

• 🎤 .wav audio files from mic1 and mic2
• 🚁 Drone models: A, B, C
• 📚 Train, validation, and test sets

🎵 Feature Extraction

• MFCCs (n_mfcc=48) extracted with librosa
• Mean pooling across time frames
• Output: 48-dim vector per audio file

🏷 Label Preparation

• Extracted:
  • Drone model
  • Maneuvering direction
  • Fault type
• Encoded with LabelEncoder

📑 Dataset Assembly

• Combined mic1 & mic2 data
• Stored features + labels in DataFrames

---

🧠 Model Architecture

🏗 Shared Feature Extractor

1. Conv1D (1→64) → BatchNorm → ReLU → MaxPool
2. Conv1D (64→128) → BatchNorm → ReLU → MaxPool
3. Conv1D (128→256) → BatchNorm → ReLU → MaxPool
4. Conv1D (256→128) → BatchNorm → ReLU

🎯 Task-Specific Heads

• Fault Classification (9 classes)
  Linear → Dropout(0.4) → Linear
• Direction Classification (6 classes)
  Linear → Dropout(0.4) → Linear
• Model Classification (3 classes)
  Linear

---

⚙ Training Setup

• Loss: CrossEntropyLoss per task
• Optimizer: Adam (lr=0.001)
• Batch Size: 32
• Epochs: 40
• Device: GPU if available

Total Loss: = loss_fault + loss_direction + loss_model