RNN - Recurrent Neural Networks & Signal Processing

A comprehensive collection of RNN implementations, signal processing utilities, and deep learning models for time series analysis, waveform identification, and regression tasks.

---

📑 Table of Contents

• Project Overview
• Project Structure
• Installation
• Source Modules
  • Signal Analysis Module (NEW)
  • Linear Regression Module
  • Signal Generation Module
• Documentation & Formulas
• Notebooks & Experiments
• Performance Benchmarks
• Requirements
• License

---

🎯 Project Overview

This repository provides a robust framework for:

• Production-ready PyTorch models for regression and time series analysis.
• Advanced Signal Processing: Identify frequency (FFT) and amplitude from 1D arrays.
• Signal Generation: Create synthetic datasets with 17+ waveform types.
• RNN/GRU/LSTM implementations from scratch.
• Video classification using CRNN architectures.

---

📁 Project Structure

RNN-main/
│
├── src/
│   ├── identify_amplitude_frequency.py  # NEW: Signal analysis (FFT, Amplitude)
│   ├── linear_regression.py              # Advanced PyTorch regression model
│   ├── dataset.py                        # Signal generation engine
│   ├── multiple_signal.py                # Legacy signal generation utils
│   └── RNN.py                            # RNN/GRU/LSTM implementations
│
├── docs/
│   └── signal_analysis_formulas.md       # Comprehensive mathematical reference
│
├── Notebook/
│   ├── identification.ipynb              # Signal identification & visualization
│   ├── 01- RNN_Classification.ipynb
│   ├── 02- RNN_Regression.ipynb
│   ├── 03- RNN_vs_GRU_Classification.ipynb
│   └── ... (other experiments)
│
├── datasets/                             # CSV datasets for training
├── linear_output/                        # Model training outputs
└── README.md

---

🚀 Installation

# Install core dependencies
pip install torch torchvision numpy pandas matplotlib scipy scikit-learn jupyter

---

📦 Source Modules

Signal Analysis Module (NEW)

File: src/identify_amplitude_frequency.py

Identify key parameters from 1D signal arrays using Fast Fourier Transform (FFT) and statistical methods.

• Frequency Identification: Automatic detection of dominant frequency components.
• Amplitude Statistics: Peak-to-Peak, RMS, Mean, and Standard Deviation.
• Visualization: Integrated plotting functions for time and frequency domains.

Model Comparison Framework (NEW)

File: src/comparative_analysis.py

A comprehensive benchmarking tool that evaluates multiple regression models for signal parameter prediction:

• Classical ML: Linear Regression, Ridge, SVR, RandomForest, GradientBoosting, and XGBoost.
• Deep Learning: Custom 1D Convolutional Neural Network (PyTorch).
• Features: Automatic extraction of 8+ time and frequency domain features.
• Evaluation: 5-fold cross-validation and ranked summary table (R2, MAE, RMSE).

Linear Regression Module

File: src/linear_regression.py

A production-ready PyTorch regression model featuring:

• 5-layer deep architecture with BatchNorm and Dropout.
• Advanced scheduling (CosineAnnealingWarmRestarts).
• Automatic GPU/CPU detection and data normalization.

---

📚 Documentation & Formulas

Detailed mathematical foundations for all metrics and wave generation can be found in: 👉 Signal Analysis Formulas & Algorithms

Includes formulas for:

• Metrics: RMS, Skewness, Kurtosis, etc.
• Transformation: Discrete Fourier Transform (DFT).
• Generation: Periodic waves, noise models, and decay functions.

---

Signal Parameter Regression (Benchmark Results)

The following scores were achieved on a standardized 256-point synthetic sine wave dataset (500 samples, 50-epoch training for CNN):

Model	Test R²	Test MAE	Test RMSE	Rank
Linear Regression	0.9998	0.0078	0.0104	1
Ridge Regression	0.9998	0.0082	0.0110	2
Gradient Boosting	0.9996	0.0106	0.0137	3
Random Forest	0.9996	0.0109	0.0145	4
XGBoost	0.9994	0.0134	0.0181	5
PyTorch 1D CNN	0.9965	0.0361	0.0431	6
SVR (RBF)	0.9929	0.0448	0.0611	7

---

👤 Author

• Jayesh Pandey

---

📝 License

This project is open source and available under the MIT License.