See it, Think it, Sorted: Large Multimodal Models are Few-shot Time Series Anomaly Analyzers

This is the official implementation of TAMA in the following paper: See it, Think it, Sorted: Large Multimodal Models are Few-shot Time Series Anomaly Analyzers.

1. Environment Setup

pip install -r requirements.txt

2.Add your API keys

Before you get started, it is necessary to get API keys of LLMs. In our framework, you should create a .yaml file called api_keys.yaml in BigModel/ directory. The format is shown below:

openai:
  api_key: 'Your API Keys'
chatglm:
  api_key: 'Your API Keys'

3. Prepare for datasets

We use datasets across four domains to evaluate our framework, the detailed information is shown below:

Dataset	Domain	Source
UCR	Industry	Current Time Series Anomaly Detection Benchmarks are Flawed and are Creating the Illusion of Progress
NASA-SMAP	Industry	Detecting Spacecraft Anomalies Using LSTMs and Nonparametric Dynamic Thresholding
NASA-MSL	Industry	Detecting Spacecraft Anomalies Using LSTMs and Nonparametric Dynamic Thresholding
NormA	Industry	Unsupervised and scalable subsequence anomaly detection in large data series
SMD	Web service	Robust Anomaly Detection for Multivariate Time Series through Stochastic Recurrent Neural Network
Dodgers	Transportation	Dodgers Loop Sensor
ECG	Health care	TSB-UAD: an end-to-end benchmark suite for univariate time-series anomaly detection
Synthetic	-	We have uploaded this dataset to Google Drive. ( Link )

It is recommended to create a directory data before downloading datasets. The file tree should be like:

./data/
├── Anomaly_Classification -> /nas/datasets/Anomaly_Classification
├── Dodgers
├── ECG
├── NASA-MSL
├── NASA-SMAP
├── NormA
├── SMD
├── UCR
│   ├── 135_labels.npy
│   ├── 135_test.npy
│   └── ...
└── synthetic_datasets

4. Run

# convert sequence data into image
python3 make_dataset.py --dataset UCR --mode train --modality image --window_size 600 --stride 200
python3 make_dataset.py --dataset UCR --mode test --modality image --window_size 600 --stride 200
# convert sequence data into text
python3 make_dataset.py --dataset UCR --mode train --modality text --window_size 600 --stride 200
python3 make_dataset.py --dataset UCR --mode test --modality text --window_size 600 --stride 200

# Image-modality
python3 main_cli.py --dataset UCR --normal_reference 3 --LLM 'GPT-4o'
# Text-modality
python3 main_cli_text.py --dataset UCR --normal_reference 1 --LLM 'GPT-4o'

5. Results Analysis

# evaluation
python3 evaluation.py
# ablation study
python3 ablation_eval.py

The quantitative results across seven datasets:

Tip

Feel free to explore our other papers in the field of time series! [Time Series Forecasting] [Time Series Anomaly Detection]