Time-series-Anomaly-Detection-transformer

Transformer based Time-series Anomaly detector model implemented in Pytorch.

This is an implementation of transformer based time-series anomaly detector, which consists of two-stage strategy of time-series prediction and anomaly score calculation.

Requirements

Ubuntu 16.04+
Python 3.5+
Pytorch 0.4.0+
Numpy
Matplotlib
Scikit-learn
argparse
visdom
pickle

Dataset

OmniAnomaly/ServerMachineDataset at master · hdjkfhkj/OmniAnomaly (github.com)

We build on this public dataset with anomaly enhancements

Transformer-based Multi-Step Prediction Model

0. Architecture

Example of usage

0. File structure

AlignmentTimeSeries: This fold includes the process of our exploration of time series alignment methods and the experimental results of time series alignment.
Anomaly_detection: This folder contains the anomaly detection models for the general structure, including models based on LSTM and models based on Transformer.
Anomaly_detection_VFL: This folder contains the anomaly detection models for vertical federated learning, both LSTM based models and Transformer based models.
DistriTailNoise2Th: This folder contains the implementation of the anomaly threshold detection algorithm and related experimental explorations.

1. Time-series prediction: Train and save prediction model on a time-series trainset.

CUDA_VISIBLE_DEVICES=1  python 1_train_predictor_FL.py --data SMD_1_3_10dim_E   --filename  machine-1-3_10dim.pkl   --epochs  1100 --emsize 32   --lr 0.002 --bptt 200 --prediction_window_size 50   --log_interval 5 --batch_size 128 --weight_decay 0.0 --model transformer --index 1st

2. Anomaly detection: Fit multivariate gaussian distribution and calculate anomaly scores on a time-series testset

CUDA_VISIBLE_DEVICES=1  python 2_2_anomaly_detection_FL.py   --data SMD_1_3_10dim_E   --filename machine-1-3_10dim.pkl   --prediction_window_size 50 --index 1st

Result

Contact

If you have any questions, please open an issue.