zhuyiche / time-series-anomaly

Anomaly detection on time series using Deep Learning techniques

Time Series Anomaly Detection with Deep Learning

Intro

The goal of this project is to design an assumption-free anomaly detection tool that can take in any time series data (one-dimensional for now) and flag out datapoints that are significantly higher/lower than usual.

Problem: Given a time series $X = [x_0, x_1, … , x_n] \in \R^n$, return a series of time indices $i$ which are anomalous.

Description

LSTMs

Currently, there are two predictors in this repo. The first is a stateful LSTM (see lstm_main.py). The LSTM works fine for time-series with shorter seasonality, but will have issues with long-term seasonalities.

The stateful LSTM takes one input at the time and predict k time-steps ahead. The state is retained for the next input (hence stateful).

Currently, the model works quite well on the power dataset, with 1 hour sampling (week seasonality of $7*24$ time-steps)

but falls short when the seasonality gets too long (e.g. 30 Mins sampling with $7224$)

1D Causal Convolution

We use a wave-net like structure to capture patterns from very long timeseries.

This is extremely fast (per epoch time of ~10s with GPU), but output can be easily thrown off with large anomalies. May need to rethink training strategy (cross-validation of some kind). Also, the anomalies does not feature as strongly as LSTM for now.

Check out the .html plotly plots to see full results.

Other similar detectors

See STL, or for a more sophisticated Seaonsal Hybrid ESD (Twitter's).

While there has been attempts to use other deep learning techniques, e.g. autoencoder.

Examples

To be completed. In the meanwhile, refer to atrous_conv_exp_03.ipynb.

TODO:

1. Differentiate between point anomalies and structural anomalies (e.g. one whole day missing vs sudden spike). Possible idea is to use a CUMSUM algorithm for the structural ones (since prediction error fits normal distribution well), then use a higher threshold for point anomalies.

Dependencies

TODO

Known errors with jupyter notebook

If you encounter a long error message for the plotly diagrams, upgrade nbformat. pip install --upgrade nbformat.

If you are still getting an error Try updating jupyter notebook. The current For newer version of jupyter, you might get an error.

IOPub data rate exceeded. The notebook server will temporarily stop sending output to the client in order to avoid crashing it. To change this limit, set the config variable `--NotebookApp.iopub_data_rate_limit`.

To fix this, run with jupyter notebook --NotebookApp.iopub_data_rate_limit=10000000000.

History

1. The conventional STL techniques work brilliant on this dataset.
2. LSTMs have issues capturing long term dependencies