Best approach/general advice for predicting anomaly in timeseries

Hello all,
I’d like to know your opinion and best practices for detecting a condition of “failure” in a mechanical/electrical motor that is monitored by a series of sensors.

In other words I need to create a unsupervised model that given unknown Inputs in a sequence can establish a category that we are approaching a fail condition
For the past months I’ve study several unsupervised approaches for clustering (hierarchical clustering, k-means) and also RNN with LSTM, GRU etc.
I’ve also made some Principal component analysis, k-means clustering on some simulated failure conditions for a small dataset where artificial anomaly where created to observe failure condition

Given that the engine would normally work well 99.x% of time and the failure condition normally arise with different pattern compared to the simulated one, what would be the best approach to collect a good sample data? How frequent? For how long? For how many motor?

What ML technique would you use to categorize an anomaly in these time-series?

Thanks for any insight

© Copyright 2013-2019 Analytics Vidhya