AIOps Anomaly Detection 🤖

This repository contains an approach to Tsinghua University's AIOps challenge in fall semester 2018. The goal was to identify anomalies among 26 different key performance indicators (KPIs).

Figure: Trends of two distinct KPIs with highlighted anomalies

Prerequisites

• Python >= 3.6
• Pytorch >= 0.4

Data preprocessing

Temporal KPI data is cleaned, resampled and augmented using the preprocess.py script.

python preprocess.py <source> <destination>

Training

Train the model by calling the train.py script. You should take a look at its content first and select the appropriate model from the models submodule. Available models are ConvModel, FullyConnected and Inception. Hyperparameters can also be changed in this file.

Submission

Take a look at the submit.py script if you need to run a prediction pass against a test dataset.