Data-Efficient Motor Condition Monitoring with Time Series Foundation Models

             

Official implemental code for the paper "Data-Efficient Motor Condition Monitoring with Time Series Foundation Models."

Dataset

Falut Diagnosis

The falut diagnosis raw dataset was collected and arranged by the Korea Advanced Institute of Science and Technology, and the adaptation edition used in this work can be found in

including:  

• 1.0 kW, 1.5 kW and 3.0kW PMSM

Total 576M and 147M time points for current signal and vibration signal respectively

Total 1.9B observations 

Details   

Fault Label  

In 1st PMSM (1.0 kW) as example:

Project architecture

FM4FD/
├── README.md                 
├── requirements.txt
├── baseline
├───├── sota.py            # contains state-of-the-art timeseries classification models     
│   └── trad.py            # contains traditional deep learning models
│   └── other items
├── MOMENT  
├───├── zero_shot.py       # use moment as a feature exractor  
│   └── finetune.py        # finetune moment by LoRA or full-finetuning
│   └── classifier.py      # compare zero_shot without feature exraction
│   └── other items
├── Mantis  
├───├── zero_shot.py       # use mantis as a feature exractor  
│   └── finetune.py        # finetune mantis
│   └── other items  
├── 1.0kW  
│   ├── current           
│   └── vibration  
├── 1.5kW  
│   ├── current           
│   └── vibration  
├── 3.0kW  
│   ├── current           
│   └── vibration  
└── checkpoints/   
    ├── baseline             
    └── fm                   

Key Result Comparison  

Here are the results of SOTA deep learning models and fine-tuned foundation models for fault diagnosis.

     

The SOTA models are from tsai:

• MiniRocket  (Dempster, 2021) (paper) 

• XCM -  An Explainable Convolutional Neural Network (Fauvel, 2021)

  (paper)

• TSPerceiver - Adapted from Perceiver IO (Jaegle, 2021) (paper)

• TSSequencerPlus - Adapted from Sequencer (Tatsunami, 2022) (paper)

• PatchTST -   (Nie, 2022)(paper)  

The foundation models are from MOMENT and Mantis.

Quick Start

First, download the corresponding dataset from the link above.

Then, change the dataset path into yours:

• fault diagnosis employs data_ssl.py in MOMENT and data_ssl.py in Mantis for data pre-processing

After that, you will be able to run the relevant code directly to see the corresponding results. For example, if you'd like to test the MOMENT feature exraction, you can run this command:

python -u MOMENT/zero_shot.py > /your/path/file/output.log 2>&1 &

Inspired by LogME, I've added scores to the model, you can get the model score when zero_shot happens, higher scores indicate the model is more applicable. 

At the same time, to better monitor the fine-tuning process in real time, I have added tensorboard to the MOMENT fine-tuning code, which allows us to run the following command immediately after the fine-tuning is complete:

tensorboard --logdir=<directory_name> --port=6006

For more steps and details about the implementation, you can check out my tutorial.