Michal Wrobel

2025-01

Implemented real-time SPC dashboard using OPC-UA data feed from CNC machines. Machine data streams to ClickHouse, SPC calculations run in Python, visualized in Grafana.

Challenge: OPC-UA Python library (opcua-asyncio) had sparse documentation. Spent significant time reverse-engineering the server's node addressing scheme from vendor documentation and community GitHub issues before data ingestion worked reliably.

Learned: OPC-UA is a standard, but vendor implementations vary substantially in how nodes are organized. Always request the server's address space export from the machine vendor before starting integration — saves days of discovery work.

2024-08

Delivered SPC training program for 18 new operators and technicians — statistical process control fundamentals, control chart interpretation, run rules, and appropriate responses to out-of-control signals.

Challenge: Most operators were confusing specification limits with control limits — a common and dangerous misunderstanding. A process within spec but out of control is still signaling a problem. Developed visual analogies and hands-on exercises to reinforce the distinction.

Learned: SPC training must start with conceptual understanding before introducing formulas. If operators understand why control limits come from process data rather than specs, the rest follows logically.

2024-03

Partnered with data scientist to build a predictive model for dimensional drift in disk machining. Provided 3 years of historical Minitab SPC records as training data.

Challenge: Three years of SPC data were stored in 14 different Minitab project files with inconsistent variable naming and unit conventions — some in millimeters, some in micrometers. Data cleaning took 3 weeks before modeling could begin.

Learned: Retroactive data science projects on quality data almost always hit data quality issues. Standardized naming conventions and centralized data storage from day one would have saved weeks of cleaning effort.

2023-11

Performed Weibull reliability analysis on HPT component vibration fatigue failure data collected over 5 years of service. Estimated characteristic life (eta) and shape parameter (beta) for failure mode planning.

Challenge: Sample size was only n=12 confirmed fatigue failures — insufficient for tight confidence intervals. The 90% confidence interval on characteristic life spanned almost an order of magnitude. Collaborated with Data Science team to explore Bayesian analysis with prior data from similar components.

Learned: Weibull analysis with small samples requires explicit communication of uncertainty. Presenting point estimates without confidence intervals to decision-makers leads to overconfident failure predictions.

2023-07

Conducted full MSA (Gauge R&R) study for CMM measurement of turbine blade airfoil geometry — 3 operators, 10 parts, 2 replicates, following AS13100 methodology.

Challenge: Gauge R&R results showed 38% of total variation attributed to measurement system — far above the 10% threshold for acceptable measurement systems. Root cause analysis identified CMM fixture variability as the dominant source. Redesigned the fixture.

Learned: Fixture design is often the forgotten variable in CMM measurement quality. A well-calibrated CMM with a poor fixture produces poor MSA results regardless of machine accuracy.

2023-03

Implemented SPC monitoring for turbine disk machining — Xbar-R charts for 6 critical dimensions including bore diameter, face flatness, and slot positions. Set control limits from 3 months of baseline data.

Challenge: Data collection was entirely manual — operators recorded measurements on paper every 2 hours. Transcription errors and delayed entries made the SPC charts unreliable. Triggered a project to automate data collection from CNC machine controllers.

Learned: SPC is only as reliable as the data feeding it. Manual data entry introduces errors that can trigger false alarms or, worse, mask genuine out-of-control conditions.