TimeGAN

Note on Training Constraints

Due to hardware limitations, training this GAN (Generator + Discriminator) for more epochs or with larger batch sizes is not feasible here. As a result, the model does not fully achieve maximum accuracy or fidelity. This code, however, demonstrates how to generate synthetic HR-like signals and train a basic GAN under constrained conditions.

Project Overview

This project demonstrates:

1. Reading HR signals from WFDB files in the Stress Recognition in Automobile Drivers dataset.
2. Loading data into a BatchLoader that normalizes, detrends, and windows the HR signals for GAN training.
3. Building a GAN with an LSTM-based Generator and Discriminator, then combining them into a single GAN model.
4. Training the GAN over multiple epochs, comparing generated (fake) HR windows to real HR windows.
5. Generating new HR signals after training and plotting these generated signals.
6. Optional display of the first 600 seconds of actual HR data for each driver.

Dataset Note
The same dataset (Stress Recognition in Automobile Drivers) was also used in the accompanying LSTM Autoencoder project, where anomaly detection was performed on HR signals. Here, we adapt a GAN approach to synthesize new HR-like data from this very dataset.

Main Steps

1. BatchLoader

   • Reads each .dat file’s HR channel and trims it to a maximum duration (max_duration_hours).
   • Interpolates missing points, normalizes the range [0, 1], applies detrend, and creates windows of size window_size.
   • Returns training batches of shape (batch_size, window_size, 1).

2. GAN Components

   • Generator (LSTM-based): Takes random noise (window_size, 1) as input and outputs a synthetic signal of the same shape.
   • Discriminator (LSTM-based): Predicts whether a given (window_size, 1) signal is real or fake.

3. Training

   • Discriminator trains on real vs. fake windows each iteration.
   • Generator is trained via the GAN model (frozen discriminator weights) to fool the Discriminator.

4. Generation & Plotting

   • After training, we generate synthetic signals by sampling random noise and feeding it into the Generator.
   • We also plot the actual HR data (first 600 seconds) for each driver in the dataset to visualize real signals.

Usage

• Requirements: wfdb, numpy, pandas, matplotlib, tensorflow, scipy, sklearn.
• Paths: Point data_dir to your directory of .dat files.
• Parameters: Adjust batch_size, window_size, max_duration_hours, and epochs as needed.
• Run:
  1. Install WFDB: !pip install wfdb (if in Jupyter/Kaggle).
  2. Execute each cell in sequence.

Dataset Link

• Kaggle: Stress Recognition in Automobile Drivers
• Same data is used in the LSTM Autoencoder project for anomaly detection on HR signals.

Examples

Original HR

Synthetic Signals