This project contains research code related to my master's thesis "An improved transfer learning approach for Intelligent Fault Diagnosis with application to ball bearings".

This repo contains the PyTorch code for the proposed FRAN-X model. The model can perform fault detection/isolation of ball bearings faults using vibrations collected from industrial machinery.

The proposed model can perform fault isolation of unlabeled datasets. This is achieved using state-of-the-art transfer learning techniques and semi-supervised training.

Run training in a docker container leveraging all GPUs

docker run --rm -v $(pwd):/workspace/ --gpus all -it pytorch-gpu python trainer_main.py --accelerator gpu --source CWRUA --target CWRUB --num_classes 4 --batch_size 128 --save_embeddings false --alpha 0.01 -n debug --learning_rate 1e-3

Run training in the cloud using Grid.AI

grid run --datastore_name "cwru" \
trainer_main.py --data_dir "/datastores/" --accelerator cpu --autorestore false  \
    --grid_search true --save_embeddings false --experiment_name grid_search --learning_rate 1e-3 --max_epochs 80

This project is heavily inspired on the paper Unsupervised Cross-domain Fault Diagnosis Using Feature Representation Alignment Networks for Rotating Machinery For training/testing we are using the CWRU dataset from Case Western Reserve University, Cleveland, Ohio

Transfer against different loads

• CWRU
  • Load: 0HP for CWRUA, 3HP for CWRUB
  • Sensor: DE
  • Faults: B IR OR N
  • Fault diameter: any
  • OR position: any

Classify: 4 classes (B IR OR N)

Transfer against different sensors/severities

• CWRU
  • Load: 0 to 3HP
  • Sensor: DE, FE
  • Faults: B IR OR N
  • Fault diameter: 007,014,021
  • OR position: any

Classify: 4 classes (B IR OR N) e.g. train on FE007, test on DE014