Manufacturing Case Study1

The Solution

• We started by listening to the machines before we built anything for the humans. The first three weeks were spent in the historian — pulling two years of sensor history, identifying the machines that had failed, and working backward through their pre-failure telemetry to understand what the data had been showing before each breakdown. Those failure signatures became the foundation of every detection model we built. Connecting the machine data • Built MQTT-based real-time data ingestion pipelines pulling live sensor streams from all 38 CNC machines across both plants — vibration (tri-axial), spindle motor temperature, coolant pressure, spindle load percentage, and cycle time — into a centralised AWS Redshift warehouse, with a 30-second refresh cycle for critical monitoring metrics • Built a historical backfill pulling 24 months of OSIsoft PI historian data into the same warehouse — establishing the baseline behavioural signature for every machine under normal operating conditions and cataloguing the pre-failure telemetry patterns from the 11 unplanned failures recorded in the prior two years • Integrated the CMMS (Maintenance Connection) maintenance work order history — linking every recorded failure, repair action, and scheduled service event to the sensor data from the same time window, giving the model labelled failure events to learn from • Built dbt transformation models standardising sensor data across 38 machines that had been installed at different times with slightly different calibration configurations — one agreed methodology for vibration RMS calculation, one temperature normalisation approach, one definition of a "degrading" spindle load trend, documented formally Building the predictive maintenance intelligence layer • Built machine-specific anomaly detection models using Isolation Forest architecture on 18 features per machine — identifying sensor readings that were statistically unusual relative to each machine's own normal operating envelope, rather than applying a single threshold across all machines • Built a failure precursor signature library from the 11 historical failures — identifying the specific combination and sequence of sensor deviations that had consistently preceded each failure type (spindle bearing degradation, coolant pump cavitation, tool holder runout, servo drive thermal event) and building pattern-matching logic to detect those signatures in current data • Built a machine health score model updating every hour — aggregating current anomaly signals across all monitored sensors into a single 0–100 health index per machine, with the score trajectory over the prior 7 days shown alongside it so the maintenance team could see deterioration trends, not just point-in-time readings • Built a condition-based maintenance recommendation engine generating a prioritised weekly maintenance work list — replacing the fixed 90-day calendar with a dynamic schedule driven by actual machine health scores, flagging machines that had deteriorated to the intervention threshold and suppressing unnecessary services on machines showing stable health Building the OEE and operations dashboard • Built a Tableau dashboard suite with four views: a plant operations director summary showing live OEE across all four lines, a machine health monitoring view with real-time health scores and
• trending for all 38 machines, a predictive maintenance work order queue showing flagged machines ranked by health score and failure risk, and a cross-plant spare parts availability tracker • Built an OEE calculation engine pulling shift production data from the MES (Plex Manufacturing Cloud) alongside the sensor data — computing Availability, Performance, and Quality components from actual machine run-time, cycle time, and scrap records, eliminating the manual shift-supervisor spreadsheet • Built a failure prediction alert system — any machine whose health score drops below 65 triggers an amber alert to the plant maintenance lead and the operations director; any machine below 45 triggers a red alert with the predicted failure mode, recommended intervention, and estimated time-to-failure based on the current degradation trajectory • Configured an automated Monday morning plant health briefing covering the prior week's OEE by line, the current health status of all 38 machines, the week's scheduled condition-based maintenance actions, and any machines that had entered the amber alert zone over the weekend