Infrastructure for Unsupervised Quality Pattern Discovery

Unsupervised pattern discovery specifically targets patterns that existing documented procedures and alarm thresholds have not defined — it finds what humans haven't yet formalized. The system needs documented baseline quality procedures (ISO/IATF-mandated SOPs, SPC limits, defect taxonomies) to establish what 'normal' looks like, but does not require deeply connected or queryable documentation. The value of this capability is discovering the undocumented; over-formalization of what constitutes anomalies would constrain the unsupervised approach.

Unsupervised ML for quality pattern discovery requires automated, high-volume capture of time-series quality metrics, contextual variables (shift, equipment, operator, weather), and multi-dimensional defect data. SPC systems and MES must automatically log measurement data from equipment without human intervention. Manual or periodic capture creates temporal gaps that break time-series continuity — the system cannot detect gradual bimodal distribution development if data arrives in weekly batches. Automated capture from operational workflows is the minimum for this use case.

Multi-dimensional unsupervised ML requires formal ontology defining entities (Product, Line, Shift, Operator, DefectType, ProcessParameter) and their relationships to enable cross-dimensional pattern discovery. Without mapped relationships — DefectType.occurrence linked to Line.equipment_age AND Shift.operator — the system processes isolated time series and misses cross-dimensional correlations. Formal structure enables the 'unusual combinations of defects' use case that simple relational database schemas cannot support.

Pattern discovery models must query QMS for defect and inspection data, pull SPC time series, access MES for production context (volumes, shifts, equipment), and push anomaly alerts to dashboards and quality engineers. API access to QMS, SPC systems, and MES covers the core data ingestion and alert distribution pipeline. Fully unified access is not required — the model can tolerate some latency as pattern detection operates over days and weeks rather than seconds.

Unsupervised models require near real-time baseline recalibration as process capability improves, new products launch, or equipment changes. A pattern that was anomalous six months ago may be the new normal after a process improvement. If baseline definitions lag process changes by weeks, the system generates stale alerts based on outdated normal distributions — alerting quality engineers to process improvements rather than actual emerging issues. Near real-time sync of process change events to model retraining is essential.

Discovering hidden correlations between quality metrics and external factors (weather, supplier batch, shift patterns) requires API-based connections between QMS, MES, SPC systems, and contextual data sources. Point-to-point integrations cover the core data ingestion pipeline. Full integration platform is not required — the pattern discovery model can tolerate batch-API pulls from each system rather than real-time unified streaming, given that pattern shifts manifest over hours and days.