Core Concepts

Modular abstraction layer

The abstraction layer sits between raw sensor signals and the HMM. It is composed of independent feature families that the expert switches on or off through the feature plan. Each family produces engineered columns the HMM consumes as Gaussian observations.

The five built-in families are:

• Statistical — rolling means, standard deviations, min/max windows.

• Temporal — first differences, smoothed differences, time-since-event.

• Stability — flags marking sensor plateaus or drifts.

• Interaction — cross-sensor combinations (ratios, sums).

• Event — boolean predicates encoding expert rules (e.g. (T > 70) & (T_stable_flag == 1)).

Supervised vs unsupervised HMM

The pipeline supports two modes:

• Supervised — initial transition, start, and emission parameters are seeded from labeled training data, then refined via Baum-Welch. Use this when you have ground-truth state labels for at least some batches.

• Unsupervised — a vanilla Gaussian HMM is fit and its discovered states are aligned to the expert’s state list using the Hungarian algorithm. Use this when labels are scarce or absent.

Both modes decode test sequences with the Viterbi algorithm and produce the same downstream artifacts.

From state sequence to event log

The decoded per-timestamp state sequence is converted into intervals (case_id, state, start, end). Brief flickers shorter than a configurable threshold (default 2 seconds) are filtered out to remove spurious transitions. The resulting interval log is the event log used by downstream process-mining tools.