Rolling Element Bearing Fault Detection (1D CNN)

This project implements a complete pipeline for rolling element ball bearing fault detection using vibration signals stored in .mat files and a 1D CNN in PyTorch.

Project Structure

• utils.py: Scans folders, loads .mat files, extracts vibration signals (Data1_AI_1_Xaxis), segments signals into overlapping windows, and builds NumPy datasets.
• dataset.py: Defines the PyTorch Dataset and helpers to split, scale (StandardScaler), and build DataLoaders.
• model.py: Implements the 1D CNN architecture for fault classification.
• train.py: Entry point for training, evaluation, and plotting.
• requirements.txt: Python dependencies.

Dataset Expectations

• The workspace (or a given --data_dir) must contain .mat files.
• Each .mat file should contain a key named Data1_AI_1_Xaxis with a long 1D vibration signal.
• File names encode the fault type:
  • Normal → label 0
  • BF (Ball Fault) → label 1
  • IF (Inner Race Fault) → label 2
  • OF (Outer Race Fault) → label 3

Matching is case-insensitive and based on substrings in the file name.

Installation

Create and activate a virtual environment, then install dependencies:

pip install -r requirements.txt

Running Training

From the project root:

python train.py --data_dir path/to/your/dataset

Arguments:

• --data_dir (default .): Root directory to scan recursively for .mat files.
• --epochs (default 20): Number of training epochs.
• --batch_size (default 64): Batch size.
• --lr (default 1e-3): Learning rate for Adam optimizer.

Outputs

Running train.py will:

• Print training loss per epoch.
• Print final test accuracy.
• Save a training loss plot as loss_curve.png.
• Save a confusion matrix plot as confusion_matrix.png.