Questions regarding NASA Data and Orion Pipelines

[github-user-29A]

Hello,

I have a few questions regarding the Orion Library:

• What is the sampling frequency of the NASA data supplied with the library?
• Why are the timestamps for the NASA data as big as they are (is each measurement really six hours apart)?
• Does Orion work with any time series data (both actual data values as well as the timestamps)? Does the data need to have a certain format besides having a timestamp and value column? (a specific sampling frequency for example)
• During testing i input one of my signals (a recording of a ballistocardiographic signal; basically a low-frequency alternative to an electrocardiography) with VERY large artificial outliers (signal values are all <1, the outliers have a value of amplitude+2) into the ARIMA, AER andTADGAN pipelines. AER and TADGAN trained for 5 epochs, ARIMA predicted 1 step.
  While AER and ARIMA found all anomalies, while TADGAN only found 3 out of 10. Is TADGAN not as good as the others?

Thank you in advance.

[sarahmish]

Hello @github-user-29A! Thank you for opening this discussion topic!

NASA Data

We have the notebook used to convert and store the NASA data available here.

• To my knowledge, the sampling frequency of the NASA is unknown. I would refer to the original paper in case it does supply this information.
• In Orion, we synthetically added a timestamp column to the NASA data. The choice of six hours is arbitrary and does not exists in the originally distributed dataset by NASA.

Data Preprocessing

• Additional information about the data need to be set in the hyperparameter field. For example, for sampling frequency, the interval hyperparameter in the primitive mlstars.custom.timeseries_preprocessing.time_segments_aggregate#1 needs to be adjusted for correct pre-processing. Additional documentation can be found here.
• Each pipeline differs from one another and the selection is based on the user's data. It is expected that not all pipelines will be able to find all the anomalies (or even the same subset of anomalies). According to our benchmark, AER currently surpasses other pipelines within some of the public datasets. For more details, see the AER paper which presents a nice summary on the successes and limitations of the pipelines we have in this library in Section 4.

[github-user-29A]

Many thanks for your reply.
Is is possible to run an entire data set through a pipeline?
I have a data set of 40 csv-files and so far I've only figured out how to train and predict on a single file (the detect anomalies and fit functions only take a data frame with time and values). I want to give the pipeline a few more files to train and predict on at once without having to concatinate them into one huge csv.

[sarahmish]

Are these csv files pertaining to the same signals? Or are they different? Here is how I would handle each situation:

1. If the csv files are one signal split into 40 files
In this case, I recommend concatenating these files into a single dataframe for cleaner training of the pipeline. If the csv files are large and you would like to process the data in batches, you can instantiate the Orion object once and call the fit function for each loaded csv file. It would be an iterative process as follows:

orion = Orion(..)

for file in files:
     df = pd.read_csv(file)
     orion.fit(df)

2. If the csv files pertain to multiple signals
Depending on whether these signals have the same characteristics as each other or not I would suggest different approaches. If the have the same characteristics, I would process them in batches similar to what I suggest in the previous point.
However, if their characteristics are different, meaning the intervals between one signal and another, or the range of values are different, then I would suggest creating a separate pipeline for each csv file.

I hope this addresses your question. Let me know if you have any follow-ups!

[github-user-29A]

Hello,

thank you for all your help so far.
In regards to the ARIMA model implemented in Orion. I am still trying to use it to find anomalies in ballistocardiographic data. While ARIMA found the artificial spikes it does not find naturally occuring anomalies within the signals.
Are there any strategies in regards to finding the right amount of parameters to use or is it just a matter of trial and error?

Here is a BCG signal as an example:

Zoomed in Signal:

The data i use is from this link: https://ieee-dataport.org/open-access/bed-based-ballistocardiography-dataset