orion-ml 0.1.0.dev0

Orion is a machine learning library built for data generated by satellites.

An open source project from Data to AI Lab at MIT.

Orion

Orion is a machine learning library built for telemetry data generated by satellites.

• License: MIT
• Development Status: Pre-Alpha
• Homepage: https://github.com/D3-AI/Orion

Overview

Orion is a machine learning library built for telemetry data generated by Satellites.

With this data, our interest is to develop techniques to:

• identify rare patterns and flag them for expert review.
• predict outcomes ahead of time.

The library makes use of a number of automated machine learning tools developed under "The human data interaction project" within the Data to AI Lab at MIT.

With the ready availability of automated machine learning tools, the focus is on:

• domain expert interaction with the machine learning system;
• learning from minimal labels;
• explainability of model outputs;
• model audit;
• scalability;

Table of Contents

• I. Data Format
  • I.1 Input
  • I.2 Output
  • I.2 Dataset we use in this library
• II. Orion Pipelines
  • II.1 Current Available Pipelines
  • II.2 Leaderboard
• III. Getting Started
  • III.1 Requirements
  • III.2 Install
• IV. Quickstart
• V. Database

Data Format

Input

Orion Pipelines work on time Series that are provided as a single table of telemetry observations with two columns:

• timestamp: an INTEGER or FLOAT column with the time of the observation in Unix Time Format
• value: an INTEGER or FLOAT column with the observed value at the indicated timestamp

This is an example of such table:

timestamp	value
1222819200	-0.366358
1222840800	-0.394107
1222862400	0.403624
1222884000	-0.362759
1222905600	-0.370746

Output

The output of the Orion Pipelines is another table that contains the detected anomalous intervals and that has at least two columns:

• start: timestamp where the anomalous interval starts
• end: timestamp where the anomalous interval ends

Optionally, a third column called score can be included with a value that represents the severity of the detected anomaly.

An example of such a table is:

start	end	score
1222970400	1222992000	0.572643
1223013600	1223035200	0.572643

Dataset we use in this library

For development, evaluation of pipelines, we include a dataset which includes several satellite telemetry signals already formatted as expected by the Orion Pipelines.

This formatted dataset can be browsed and downloaded directly from the d3-ai-orion AWS S3 Bucket.

This dataset is adapted from the one used for the experiments in the Detecting Spacecraft Anomalies Using LSTMs and Nonparametric Dynamic Thresholding paper. Original source data is available for download here. We thank NASA for making this data available for public use.

Orion Pipelines

The main component in the Orion project are the Orion Pipelines, which consist of MLBlocks Pipelines specialized in detecting anomalies in time series.

As MLPipeline instances, Orion Pipelines:

• consist of a list of one or more MLPrimitives
• can be fitted on some data and later on used to predict anomalies on more data
• can be scored by comparing their predictions with some known anomalies
• have hyperparameters that can be tuned to improve their anomaly detection performance
• can be stored as a JSON file that includes all the primitives that compose them, as well as other required configuration options.

Current Available Pipelines

In the Orion Project, the pipelines are included as JSON files, which can be found inside the orion/pipelines folder.

This is the list of pipelines available so far, which will grow over time:

name	location	description
Dummy	orion/pipelines/dummy.json	Dummy Pipeline to showcase the input and output format and the usage of sample primitives
LSTM Dynamic Threshold	orion/pipelines/lstm_dynamic_threshold.json	LSTM Based pipeline inspired by the Detecting Spacecraft Anomalies Using LSTMs and Nonparametric Dynamic Thresholding paper
Mean 24h LSTM	orion/pipelines/mean_24h_lstm.json	LSTM Based pipeline with 24h mean aggregation preprocessing
Median 24h LSTM	orion/pipelines/median_24h_lstm.json	LSTM Based pipeline with 24h median aggregation preprocessing
Sum 24h LSTM	orion/pipelines/sum_24h_lstm.json	LSTM Based pipeline with 24h sum aggregation preprocessing
Skew 24h LSTM	orion/pipelines/skew_24h_lstm.json	LSTM Based pipeline with 24h skew aggregation preprocessing
CycleGAN	orion/pipelines/cyclegan.json	CycleGAN Based pipeline
ARIMA	orion/pipelines/arima.json	ARIMA Based pipeline

Leaderboard

In this repository we maintain this up-to-date leaderboard with the current scoring of the pipelines according to the benchmarking procedure explained in the benchmark documentation.

pipeline	accuracy	f1	precision	recall
CycleGAN	0.781147	0.137234	0.147674	0.18173
LSTM Dynamic Thresholding	0.832052	0.125999	0.178968	0.151298
Dummy	0.818975	0.108436	0.13994	0.133865
Mean 24h LSTM	0.667412	0.0420656	0.0775713	0.0456106
Sum 24h LSTM	0.685844	0.0417817	0.066248	0.033882
ARIMA	0.510343	0.038821	0.0604475	0.0377441
Median 24h LSTM	0.673667	0.0237867	0.0604165	0.0178578
Skew 24h LSTM	0.369548	0.01142	0.0213837	0.00902504

Getting Started

Requirements

Python

Orion has been developed and runs on Python 3.6.

Also, although it is not strictly required, the usage of a virtualenv is highly recommended in order to avoid interfering with other software installed in the system where you are trying to run Orion.

MongoDB

In order to be fully operational, Orion requires having access to a MongoDB database running version 3.6 or higher.

Install

The easiest and recommended way to install Orion is using pip:

pip install orion-ml

This will pull and install the latest stable release from PyPi.

If you want to install from source or contribute to the project please read the Contributing Guide.

Docker

Even thought it's not mandatory to use it, Orion comes with the possibility to be distributed and run as a docker image, making its usage in offline systems easier.

For more details please read the Docker Usage Documentation.

Quickstart

In the following steps we will show a short guide about how to run one of the Orion Pipelines on one of the signals from the Demo Dataset.

1. Load the data

In the first step we will load the S-1 signal from the Demo Dataset.

We will do so in two parts, train and test, as we will use the first part to fit the pipeline and the second one to evaluate its performance.

To do so, we need to import the orion.data.load_signal function and call it twice passing the 'S-1-train' and 'S-1-test' names.

from orion.data import load_signal

train = load_signal('S-1-train')
test = load_signal('S-1-test')

The output will be a table in the format described above:

    timestamp     value
0  1222819200 -0.366359
1  1222840800 -0.394108
2  1222862400  0.403625
3  1222884000 -0.362759
4  1222905600 -0.370746

2. Detect anomalies using a pipeline

Once we have the data, let us try to use the LSTM pipeline to analyze it and search for anomalies.

In order to do so, we will have import the orion.analysis.analyze function and pass it the train and test dataframes and the name of the pipeline that we want to use:

from orion.analysis import analyze

anomalies = analyze(
    pipeline='lstm_dynamic_threshold',
    train=train,
    test=test
)

NOTE: Depending on your system and the exact versions that you might have installed some WARNINGS may be printed. These can be safely ignored as they do not interfere with the proper behavior of the pipeline.

The output of the previous command will be a pandas.DataFrame containing a table in the Output format described above:

        start         end     score
0  1394323200  1399701600  0.673494

3. Evaluate performance

In this next step we will load some already known anomalous intervals and evaluate how good our anomaly detection was by comparing those with our detected intervals.

For this, we will first load the known anomalies for the signal that we are using:

from orion.data import load_anomalies

known_anomalies = load_anomalies('S-1')

The output will be a table in the same format as the anomalies one.

        start         end
0  1392768000  1402423200

Afterwards, we pass the ground truth, the detected anomalies and the original test data to the orion.metrics.accuracy_score and orion.metrics.f1_score functions in order to compute a score that indicates how good our anomaly detection was:

from orion.metrics import accuracy_score, f1_score

accuracy_score(known_anomalies, anomalies, test)  # -> 0.972987721691678

f1_score(known_anomalies, anomalies, test)  # -> 0.7155172413793103

Database

Orion comes ready to use a MongoDB Database to easily register and explore:

• Multiple Datasets based on signals from one or more satellites.
• Multiple Pipelines, including historical Pipeline versions.
• Pipeline executions on the registered Datasets, including any environment details required to later on reproduce the results.
• Pipeline execution results and detected events.
• Comments about the detected events.

This, among other things, allows:

• Providing visibility about the system usage.
• Keeping track of the evolution of the registered pipelines and their performance over multiple datasets.
• Visualizing and browsing the detected events by the pipelines using a web application.
• Collecting comments from multiple domain experts about the detected events to be able to later on curate the pipelines based on their knowledge.
• Reproducing previous executions in identical environments to replicate the obtained results.
• Detecting and keeping a history of system failures for later investigation.

The complete Database schema and usage instructions can be found in the database documentation

History

0.1.0

• First release.