PatchLinear for Single-step Forecasting and Anomaly Detection

Model for univariate time series single-step forecasting and anomaly detection.

Requirements

data_provider: Dataset preprocessing
logs: Log directory, storing result logs of each dataset
util: Utility library
- index.py: Calculate the position of the most similar historical subsequence to the current subsequence
- tools.py: Anomaly label adjustment, plotting, rolling calculation of subsequence mean and variance
models: Models of different Methods
layers: Layers that execute the model
scripts: Executable files directory
exp: Main files directory

Dataset
1. Place all datasets in the same directory, such as the ./dataset/ directory
2. The organization of each dataset file is as follows:
  1. KPI
    1. train/*.csv
    2. test/*.csv
    3. The same name sequence in the train and test directories corresponds to the same sequence, and the sequence consists of timestamp, value, label
  2. Yahoo
    1. real/*.csv
    2. synthetic/*.csv
    3. A3/*.csv
    4. A4/*.csv
  3. MSLU: As it is part of the NASA MSL dataset, it can be directly accessed from the NASA dataset
    1. train/*.npy Storing training set sequences
    2. test/*.npy Storing test set sequences
    3. test_label/*.csv Storing corresponding label data
  4. NAB
    1. *.csv Directly place the corresponding sequence data
Adapt the environment with requirement.txt:
```
pip install -r requirement.txt
```
Run the executable files in the corresponding executable files directory to start the algorithm.

You can run the following command to perform single-step forecasting or anomaly detection on the sequences in the corresponding dataset:

sh ./scripts/PatchLinear/Yahoo.sh

The settable parameters are explained in the run_longExp.py file.

We appreciate the following github repo very much for the valuable code base and datasets:

If you have any questions or concerns, please contact me: zhengmh1229@stu.xmu.edu.cn or submit an issue