Need help with telemanom?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

577 Stars 157 Forks Other 37 Commits 8 Opened issues


A framework for using LSTMs to detect anomalies in multivariate time series data. Includes spacecraft anomaly data and experiments from the Mars Science Laboratory and SMAP missions.

Services available


Need anything else?

Contributors list

Telemanom (v2.0)

v2.0 updates: - Vectorized operations via numpy - Object-oriented restructure, improved organization - Merge branches into single branch for both processing modes (with/without labels) - Update requirements.txt and Dockerfile - Updated result output for both modes - PEP8 cleanup

Anomaly Detection in Time Series Data Using LSTMs and Automatic Thresholding


Telemanom employs vanilla LSTMs using Keras/Tensorflow to identify anomalies in multivariate sensor data. LSTMs are trained to learn normal system behaviors using encoded command information and prior telemetry values. Predictions are generated at each time step and the errors in predictions represent deviations from expected behavior. Telemanom then uses a novel nonparametric, unsupervised approach for thresholding these errors and identifying anomalous sequences of errors.

This repo along with the linked data can be used to re-create the experiments in our 2018 KDD paper, "Detecting Spacecraft Anomalies Using LSTMs and Nonparametric Dynamic Thresholding", which describes the background, methodologies, and experiments in more detail. While the system was originally deployed to monitor spacecraft telemetry, it can be easily adapted to similar problems.

Getting Started

Clone the repo (only available from source currently):

git clone && cd telemanom

Configure system/modeling parameters in

file (to recreate experiment from paper, leave as is). For example: -
train: True
, a new model will be trained for each input stream. If
(default) existing trained model will be loaded and used to generate predictions -
predict: True
Generate new predictions using models. If
(default), use existing saved predictions in evaluation (useful for tuning error thresholding and skipping prior processing steps) -
l_s: 250
Determines the number of previous timesteps input to the model at each timestep
(used to generate predictions)

To run via Docker:

```shell script docker build -t telemanom .

rerun experiment detailed in paper or run with your own set of labeled anomlies in 'labeled_anomalies.csv'

docker run telemanom -l labeled_anomalies.csv

run without labeled anomalies

docker run telemanom ```

To run with local or virtual environment

From root of repo, curl and unzip data:

curl -O && unzip && rm

Install dependencies using python 3.6+ (recommend using a virtualenv):

pip install -r requirements.txt

Begin processing (from root of repo):

# rerun experiment detailed in paper or run with your own set of labeled anomlies
python -l labeled_anomalies.csv

run without labeled anomalies


A jupyter notebook for evaluating results for a run is at

. To launch notebook:
jupyter notebook telemanom/result-viewer.ipynb

Plotly is used to generate interactive inline plots, e.g.:



Using your own data

Pre-split training and test sets must be placed in directories named

. One
file should be generated for each channel or stream (for both train and test) with shape (
). The filename should be a unique channel name or ID. The telemetry values being predicted in the test data must be the first feature in the input.

For example, a channel

should have train/test sets named
with shapes akin to
(3925, 61)
, where the number of input dimensions are matching (
). The actual telemetry values should be along the first dimension

Raw experiment data

The raw data available for download represents real spacecraft telemetry data and anomalies from the Soil Moisture Active Passive satellite (SMAP) and the Curiosity Rover on Mars (MSL). All data has been anonymized with regard to time and all telemetry values are pre-scaled between

according to the min/max in the test set. Channel IDs are also anonymized, but the first letter gives indicates the type of channel (
= power,
= radiation, etc.). Model input data also includes one-hot encoded information about commands that were sent or received by specific spacecraft modules in a given time window. No identifying information related to the timing or nature of commands is included in the data. For example:


This data also includes pre-split test and training data, pre-trained models, predictions, and smoothed errors generated using the default settings in

. When getting familiar with the repo, running the
notebook to visualize results is useful for developing intuition. The included data also is useful for isolating portions of the system. For example, if you wish to see the effects of changes to the thresholding parameters without having to train new models, you can set
to use previously generated predictions from prior models.

Anomaly labels and metadata

The anomaly labels and metadata are available in

, which includes:
  • channel id
    : anonymized channel id - first letter represents nature of channel (P = power, R = radiation, etc.)
  • spacecraft
    : spacecraft that generated telemetry stream
  • anomaly_sequences
    : start and end indices of true anomalies in stream
  • class
    : the class of anomaly (see paper for discussion)
  • num values
    : number of telemetry values in each stream

To provide your own labels, use the

file as a template. The only required fields/columns are
is a list of lists that contain start and end indices of anomalous regions in the test dataset for a channel.

Dataset and performance statistics:


| | SMAP | MSL | Total | | ------------------------------- | :-------: | :------: | :------:|
| Total anomaly sequences | 69 | 36 | 105 | | Point anomalies (% tot.) | 43 (62%) | 19 (53%) | 62 (59%)| | Contextual anomalies (% tot.) | 26 (38%) | 17 (47%) | 43 (41%)| | Unique telemetry channels | 55 | 27 | 82 | | Unique ISAs | 28 | 19 | 47 | | Telemetry values evaluated | 429,735 | 66,709 | 496,444 |

Performance (with default params specified in paper)

| Spacecraft | Precision | Recall | F_0.5 Score | | ----------------- | :-------: | :------: | :------: |
| SMAP | 85.5% | 85.5% | 0.71 | | Curiosity (MSL) | 92.6% | 69.4% | 0.69 | | Total | 87.5% | 80.0% | 0.71 |


Each time the system is started a unique datetime ID (ex.

) will be used to create the following - a results file (in
) that extends
to include identified anomalous sequences and related info - a data subdirectory containing data files for created models, predictions, and smoothed errors for each channel. A file called
is also created that contains parameter settings and logging output during processing.

As mentioned, the jupyter notebook

can be used to visualize results for each stream.


If you use this work, please cite:

  title={Detecting Spacecraft Anomalies Using LSTMs and Nonparametric Dynamic Thresholding},
  author={Hundman, Kyle and Constantinou, Valentino and Laporte, Christopher and Colwell, Ian and Soderstrom, Tom},
  journal={arXiv preprint arXiv:1802.04431},


Telemanom is distributed under Apache 2.0 license.

Contact: Kyle Hundman ([email protected])


We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.