Modeling Manufacturing Observations

Licentiate thesis, 2026

Neural time-series methods hold increasing promise for improving how man
ufacturing systems operate, enabling more accurate monitoring, earlier fault
detection, and better-informed operational decisions. Human-in-the-loop archi
tectures extend this further, pairing machine learning with human expertise to
refine and guide how analytical pipelines perform in practice.
Building such pipelines, however, surfaces a range of challenges. Industrial
systems record machine states, buffer levels, and event logs for operational
supervision, not to produce clean training sequences. The data entering neural
time-series pipelines are often irregular in time and incomplete across variables,
while most neural methods assume regularly sampled and densely observed
inputs.
The aim is to relate how observation conditions arise in manufacturing
systems to where neural methods encode assumptions about time and absence,
and thereby ground method selection in the observation process for human-in
the-loop (HITL) decision-support settings.
A simulation study of a discrete manufacturing system shows how ordinary
production-structure mechanisms generate temporally uneven, asynchronous,
and dependent traces under controlled assumptions. A structured survey
examines how neural methods for irregular and incomplete time series differ in
their treatment of time and absence within the modeling pipeline. The main
contribution is methodological: comparison should begin from the observation
process, treating irregular sampling and missingness as distinct properties with
different implications for representation, evaluation, and method selection in
HITL decision-support settings.