Multivariate time series forecasting for unsupervised clustering

[huberl]

I'm working on a tool that extracts anomalies from the IMU and other sensors of a robot without having any labels. My code already extracts anomalous snippets from the sensor data which results in an array of size (#num_snippets, #varying_snippet_length, #num_sensors), where #num_sensors is 10.

In a next step I want to cluster all the obtained snippets without having to label anything. The plan is to first extract features and then select those that are actually useful using tsfresh.

I've already read #678, which suggests to transform this into a forecasting task. Since I have 10 sensors, I would need to forecast 10 time-series at once. Is there any way to do this or am I missing something?

Can I simply perform the feature selection independently? So I would use the time-series data of sensor1 to forecast sensor1 and repeat this for all 10 sensors. In the end I could take the union over all relevant features as shown in this notebook.