santobedi / battery_fault_detection_nc_github

This is the official code repository for our Nature Communication paper entitled "Realistic fault detection of Li-ion battery via dynamical deep learning approach"

Environment requirement

We recommend using the conda environment.

# basic environment
CUDA 10.2  # Must use this specific version. Please follow https://developer.nvidia.com/cuda-10.2-download-archive. 
python 3.6

# pytorch version
pytorch==1.5.1

# run after installing correct Pytorch package
pip install --no-index torch-scatter -f https://pytorch-geometric.com/whl/torch-1.5.0+cu102.html
pip install --no-index torch-sparse -f https://pytorch-geometric.com/whl/torch-1.5.0+cu102.html
pip install --no-index torch-cluster -f https://pytorch-geometric.com/whl/torch-1.5.0+cu102.html
pip install --no-index torch-spline-conv -f https://pytorch-geometric.com/whl/torch-1.5.0+cu102.html
pip install torch-geometric==1.5.0
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirement.txt
pip install tensorflow==2.6.2

Dataset preparation

Download

Download from the link in our paper and unzip them. Please make sure the data structure is like the following.

    |--data
        |--battery_brand1
            |--label
            |--train
            |--test
            |column.pkl
        |--battery_brand2
            |--...
        |--battery_brand3
            |--...
    

File content

Each pkl file is a tuple including two parts. The first part is charging time series data. column.pkl contains column names for the time sequence data. The second part is meta data which contains fault label, car number, charge segment number and mileage.

The label folder contains car numbers and their anomaly labels.

Generate path information for five-fold validation

In our paper, we provide experiments of training with different brands. To facilitate the organization of training and test data, we use 1) a python dict to save car number-snippet paths information, which is named as all_car_dict.npz.npy, and 2) a dict to save the randomly shuffled normal and abnormal car number to perform five-fold training and testing, which is named as ind_odd_dict*.npz.npy. By default, the code is running on the first brand. So our code is now running on ind_odd_dict1.npz.npy.

The details of running experiments on other brands requires modification on some code, which is illustrated in this readme file later.

To build the all_car_dict.npz.npy and ind_odd_dict*.npz.npy, run

cd data

Run five_fold_train_test_split.ipynb and then you get all the files saved in nature code\five_fold_utils\. (Running each cell of the five_fold_train_test_split.ipynb may take a few minutes. If not, please check the data path carefully.)

The cell output of each cell contains randomly shuffled ind_car_num_list and ood_car_num_list. You may print it out to see the car numbers you are using.

Run DyAD(ours)

Setting another brand: By default, we are using brand 1. To run experiments on other brands, one should manually change the variable ind_ood_car_dict_path in DyAD/model/dataset.py. For example, if you want to use brand 2, then you should use ind_ood_car_dict_path='../five_fold_utils/ind_odd_dict2.npz.npy',. (An easy way to do so is to use Ctrl+F to search the name of the variables.)

train

Please check the model_params_battery_brand*.json files carefully for hyperparameter settings. model_params_battery_brand1/2/3.json are used to train vehicles of brand1/2/3 separately. And use fold_num to do the five-fold training and testing. To start training, run

cd DyAD
python main_five_fold.py --config_path model_params_battery_brand1.json --fold_num 0

If you want to fully run the five-fold experiments, you should run five times with different --fold_num. After training, the reconstruction errors of data are recorded in save_model_path configured by the params_fivefold.json file.

AutoEncoder & SVDD

Setting another brand: By default, we are using brand 1. To run experiments on other brands, one should manually change the variable ind_ood_car_dict in AE_and_SVDD/traditional_methods.py similarly as above.

train

To start training, run

cd AE_and_SVDD
python traditional_methods.py --method auto_encoder --normalize --fold_num 0
python traditional_methods.py --method deepsvdd --normalize --fold_num 0

If you want to fully run the five-fold experiments, you should run five times with different --fold_num.

LSTM-AD

Setting another brand: By default, we are using brand 1. To run experiments on other brands, one should manually change the variable ind_ood_car_dict in Recurrent-Autoencoder-modify/agents/rnn_autoencoder.py and Recurrent-Autoencoder-modify/datasets/battery.py similarly as above.

train

To start training, run

cd Recurrent-Autoencoder-modify
python main.py configs/config_lstm_ae_battery_0.json

where ...0.json means the first fold in five-fold validation.

If you want to fully run the five-fold experiments, you should run five times with different config files (config_lstm_ae_battery_*.json where * can be 1/2/3/4).

GDN

Setting another brand: By default, we are using brand 1. To run experiments on other brands, one should manually change the variable ind_ood_car_dict in GDN_battery/datasets/TimeDataset.py and GDN_battery/main.py similarly as above.

train

cd GDN_battery
bash run_battery.sh 3 battery 1 0 20

where 3 is the gpu number, battery is the dataset name, 1 is the fold number (also can be 0/2/3/4) 0 means use all data to train and 20 is epoch number. For details, please see the .sh file.

Calculated ROC score

For all the mentioned algorithms, we calculated the AUROC values with jupyter-notebooks in notebooks. For each notbook file, the suffix threshold represents the calculation of robust fault scores, and the suffix threshold_no represents the calculation of average fault scores.

Necessary modification: Since the save path may be time dependent and machine dependent, one needs to change the path information in each jupyter notebook. One should also modify the path of the saved reconstruction error if one is using different brands.

Data availability

The datasets are available at links below https://1drv.ms/u/s!AiSrJIRVqlQAgcjKGKV0fZmw5ifDd8Y?e=CnzELH237 or https://disk.pku.edu.cn:443/link/37D733DF405D8D7998B8F57E4487515A238.

Code Reference

We use partial code from

https://github.com/yzhao062/pyod
https://github.com/d-ailin/GDN
https://github.com/PyLink88/Recurrent-Autoencoder