Thesis

A governed real-time architecture for cycle-level digital representation, quality monitoring and decision support in cyclic manufacturing : injection moulding as a case study

Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2026
Thesis identifier
  • T18230
Person Identifier (Local)
  • 202163118
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • Cyclic manufacturing processes exhibit repeated operating patterns, yet small variations within and between cycles can affect product quality and undermine the reliability of offline prediction models. This thesis proposes, implements and evaluates a governed real-time architecture for cycle-level digital representation, quality monitoring and decision support, using injection moulding as the experimental case study. The production cycle is treated as the traceable unit connecting acquired sensor signals, digital representations, predictive outputs, released decisions, subsequent quality confirmation and model-management actions. Two predictive representations were evaluated: an event-anchored feature vector aligned to the global maximum cavity pressure, and a full-curve representation retaining temporal process information. These were integrated into an architecture comprising a fixed Champion model, an updateable Challenger, drift-aware monitoring, a conservative final decision policy, human oversight, controlled retraining conditions and persistent evidence records. Evaluation used 903 historically labelled cycles and two governed runtime runs of 850 cycles each, representing stable and retraining-enabled operation. The best offline configurations achieved accuracy of 0.966, F1-score of 0.963 and ROCAUC of 0.994. Runtime evaluation examined classification performance, model disagreement, drift, processing latency, decision-loss sensitivity and a recorded retraining event. These findings emphasise traceability and governance alongside predictive accuracy. The principal contribution is an implemented and evaluated governed real-time architecture that integrates cycle representation, quality prediction, runtime monitoring, traceable decision release and controlled model updating within a single operational framework, rather than treating these functions as isolated problems. This integration preserves explicit boundaries between prediction, decision release and model-update authority, without claiming autonomous control of machine actuators. Although the architecture provides a basis for cycle-level decision support in other cyclic manufacturing processes, transfer requires process-specific adaptation and independent validation.
Advisor / supervisor
  • Qin, Y. (Yi)
Resource Type
DOI

Relations

Items