Suitable model architecture for categorical, sequential data


Suppose you have a dataset, containing the log data of a set of complex devices, e.g. turbofans.

For each device, the log consists of a time sequence of categorical events.

Dataset = [[device_1, [('Err_2: voltage high', timestamp), ('Warn_1: temperature rising', timestamp), ...]],
           [device_2, [('Warn_6: rpm > 500', timestamp), ('Inf_5: system X activated', timestamp), ...]],

After some time, a device fails and needs to be replaced, which is also available in the log.
Now, the goal is to predict the failure of devices based on the above log dataset, and to identify, which sequence of events is important for the failure prediction.

Since the data has a temporal structure, I thought about using LSTMs.
However, the usual input of LSTMs consists of a fixed set of features (i.e. like a continuous temperature) over time.

In the above case though, the input consists of a sequence of single, individual features at a given point in time (each log entry). I.e., the log of device_1 may have totally different log events (-> features) than the log of device_2.

So I'm wondering, if LSTMs are still a sensible model architecture for the given problem, or if other types of models are better suited (keeping interpretability in mind as well).


Posted 2020-04-28T14:44:34.640

Reputation: 101

No answers