Validation on synth and real data

This issue adresses the lack of testing and validating of what we've done so far.

In particular, we want to:

• implement a cut-off-value, depending on reconstruction in the nominal case (e.g., avg. reconstruction error for nominal cases),
• train and validate (metrics: Prec, Rec, F1, ROC, AUC and evaluate some / at least one of the metrics depending on anomalous behaviour, e.g. amp_rest_pos) an AE on synth data,
• train and validate (metrics: Prec, Rec, F1, ROC, AUC) an AE on the acquired real data,
• based on the results, evaluate whether the data quality (doubled data) of the past trials is sufficient.
• Plot reconstruction loss over time
• Plot reconstruction loss over time for each feature
• Plot reconstruction over time for each feature - first synth data

We see this validation procedure as a crucial step before acquiring new data or deploying online 😃