PyTorch 1D-CNN for Motor Current Signature Analysis (MCSA)

This project trains a 1D CNN on three-phase currents to classify motor condition and optionally predict fault severity.

Environment Setup

1. Clone the repository:

   git clone <repository-url>
   cd csa-induction-motor-pytorch

2. Create a Python virtual environment:

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies:

   pip install --upgrade pip
   pip install -r requirements.txt

4. Verify PyTorch installation:

   python -c "import torch; print(torch.__version__)"

5. Prepare data directory:

   mkdir -p data/simulink

6. Place your Simulink CSV files in data/simulink/ following the naming conventions above.

Simulink CSV Data

• Place your CSVs under data/simulink.
• Each CSV contains columns: Time,Torque,RPM,Ia,Ib,Ic (order flexible; headers used).
• Filename conventions (case-insensitive):
  • healthy_1..6.csv → class Healthy, load 1..6
  • BearingFault_1..6.csv → class Bearing, severity 1..6
  • PhaseImbalance_1..6.csv or StatorShort_1..6.csv → class StatorShort, severity 1..6
  • RotorFault_1..6.csv or BrokenRotorBar_1..6.csv → class BrokenRotorBar, severity 1..6
• The dataset infers labels and severity from filenames. If no number is present, severity defaults to 0 (ignored in training).

Training

python train.py --config config.yaml

OR if you want customized training:

python "path/to/PyTorch 1DCNN MCSA/train.py" --data-csv "path/to/PyTorch 1DCNN MCSA/data/simulink" --save-dir "path/to/PyTorch 1DCNN MCSA/runs/exp" --epochs 80 --batch-size 32 --severity-head --severity-loss-weight 0.3 --include-aux

• --severity-head: enables multi-task prediction with a 6-class severity head.
• --severity-loss-weight: scales the severity loss term (default 0.3).
• --include-aux: includes RPM and Torque (and vibration/rpm when available) as auxiliary channels.
• Defaults like window_length, hop_length, scheduler, etc., are in config.yaml.

Evaluation

python "path/to/PyTorch 1DCNN MCSA/evaluate.py" --checkpoint "path/to/PyTorch 1DCNN MCSA/runs/exp/best.pth" --data-csv "path/to/PyTorch 1DCNN MCSA/data/simulink" --save-dir "path/to/PyTorch 1DCNN MCSA/runs/sim_eval"

• Saves confusion_matrix.png. If severity head is enabled, also saves severity_hist.png.

Prediction

python "path/to/PyTorch 1DCNN MCSA/predict.py" --checkpoint "path/to/PyTorch 1DCNN MCSA/runs/exp/best.pth" --data-csv "path/to/PyTorch 1DCNN MCSA/data/simulink" --topk 3

• Prints top-k class labels and confidences per window.
• Prints severity_pred only when the top predicted class is not Healthy.

Label Mapping

• Classes: [Healthy, Bearing, BrokenRotorBar, StatorShort].
• Severity head: 6 classes representing severity 1..6. Healthy windows are ignored in severity loss.

Notes

• Windows are created by sliding over time series (window_length, hop_length from config.yaml).
• Normalization uses z-score by default; scaler computed from training split.
• For Windows, num_workers defaults to 0 in notebooks to avoid pickling issues.