Bayesian Deep Learning for Remaining Useful Life Estimation of Machine Tool Components

Bayesian and frequentist deep learning models for remaining useful life (RUL) estimation are evaluated on simulated run-to-failure data. Implemented in PyTorch, developed and tested on Ubuntu 18.04 LTS. All the experiments were run on a publicly available Google Compute Engine Deep Learning VM instance with 2 vCPUs, 13 GB RAM, 1 NVIDIA Tesla K80 GPU and PyTorch 1.2 + fast.ai 1.0 (CUDA 10.0) framework.

Requirements

Anaconda Python >= 3.6.4 (see https://www.anaconda.com/distribution/)

Installation

Clone or download the repository, open a terminal in the root directory and run the following commands:

conda env create -f environment.yml

conda activate bayesian-deep-rul

Now the virtual environment bayesian-deep-rul is active. To deactivate it, run:

conda deactivate

When you do not need it anymore, run the following command to remove it:

conda remove --name bayesian-deep-rul --all

Dataset

The models were tested on the four simulated turbofan engine degradation subsets in the publicly available Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset. Check datasets/CMAPSS/README.md for instructions on how to download the dataset.

Usage

Open a terminal in the root directory, activate the virtual environment and run one of the following commands:

• sh train.sh to train the selected model. Parameters can be modified by editing train.sh

• sh evaluate.sh to evaluate the selected model. Parameters can be modified by editing evaluate.sh

• sh run_experiments.sh to replicate the experiments on the C-MAPSS dataset

TensorBoard

Open a terminal in the root directory, activate the virtual environment and run tensorboard --logdir . to monitor the training process with TensorBoard. If you are training on a remote server, connect through SSH and forward a port from the remote server to your local computer (gcloud compute ssh <your-vm-name> --zone=<your-vm-zone> -- -L 6006:localhost:6006 on a Google Compute Engine Deep Learning VM instance).

Results

Training and evaluation logs of the experimental results are provided for verification. Run results/results.ipynb in Jupyter Notebook to check the results by yourself.

Contact

[email protected]