This notebook walks through a practical pipeline for classifying heartbeat audio:

1. Visualize raw signals (normal vs abnormal)
2. Denoise & simplify with rectification and rolling averages (signal envelope)
3. Engineer features:
   • Global stats (mean, std, max of envelopes)
   • Tempo features from librosa
   • Spectrogram features: spectral centroid & bandwidth
4. Train & evaluate a LinearSVC classifier

The emphasis is on feature engineering for time series, using audio as the example domain.

• Data loading

  • CSVs: invariance_normal.csv, invariance_abnormal.csv, audio.csv
  • TXT (converted to CSV in-notebook): normal_full.txt, abnormal_full.txt, etc.
  • HDF5 bundle: audio_munged.hdf5 with keys:
    • h5io/key_data → (time-indexed audio columns)
    • h5io/key_meta → labels and metadata
    • h5io/key_sfreq → sampling frequency (e.g., 2205 Hz)

• Visualization

  • Multi-panel raw waveform plots
  • Envelope plots after rectification + rolling mean
  • Spectrograms (librosa.stft + amplitude_to_db) with overlays

• Feature engineering

  • Envelope stats: mean / std / max per recording
  • Tempo estimates & summary stats
  • Spectral centroid & bandwidth (sequence and summary)

• Model

  • sklearn.svm.LinearSVC with simple train/test split and accuracy

# 1) Create and activate a virtual environment
python -m venv .venv
# Windows
.venv\Scripts\activate
# macOS/Linux
source .venv/bin/activate

# 2) Install dependencies
pip install -r requirements.txt

#3) Git Clone
git clone https://github.com/Joe-Naz01/time_series.git
cd time_series

# 3) Launch Jupyter
jupyter lab   # or: jupyter notebook