A Tokyo industrial conglomerate sequences predictive maintenance and computer-vision QC, cutting downtime 64%

Context

Your industrial conglomerate runs 14 plants across Japan and South-East Asia, producing precision components for automotive Tier-1 customers. Unplanned downtime on the four highest-value lines averaged 87 hours a quarter, costing the group an estimated JPY 2.1B (US$14M) per year in scrap, late-delivery penalties, and overtime recovery. Visual quality inspection on the same lines required 84 inspectors across two shifts, with a 3.1% defect-escape rate that triggered three customer complaints in the previous fiscal year. The Group COO, an engineer by training, had read four McKinsey reports on predictive maintenance and computer-vision QC. He wanted to know which of the two to start with, in his plants, with his data.

Challenge

Three plant constraints. First, machine telemetry was inconsistent: two plants ran modern OPC-UA, four ran a 2011 SCADA, the rest were a mix. Second, the QC inspectors were a 25-year average tenure team. Replacing them with cameras would have been a labour-relations crisis. Third, the IT-OT divide inside the group was wide enough that the plant-floor PLC engineers and the Tokyo data team had not collaborated on a shared project in five years.

Approach

We ran a Cynefin-tagged sequencing exercise first: predictive maintenance on legacy SCADA was complicated (analyzable, expert-bound), while computer-vision QC across 14 plants was complex (emergent, experiment-bound). We tackled them differently. The 5-phase mentor model: scan, sequence, prove, scale, hand-over.

Scan (3 weeks) inventoried telemetry quality across all four target lines and shadowed inspectors at two plants. Sequence (1 week) produced a written go/no-go: predictive maintenance first on the two OPC-UA plants, computer-vision QC as a parallel pilot at one plant where the inspector union representative agreed to co-design the workflow.

Prove (12 weeks) deployed a vibration-and-temperature predictive model on six critical assets per plant, and a vision-augmented inspection workflow where the camera flagged candidates and the inspector confirmed. The inspector kept final say. Scale (16 weeks) extended predictive coverage to 22 assets per plant and rolled vision-augmented inspection to three more lines. Hand-over (parallel from week 14): two of your plant data engineers and one OT specialist took ownership.

Results

Unplanned downtime on the two predictive-maintenance plants fell from 87 hours per quarter to 31 hours, a 64% reduction, recovering an estimated JPY 1.3B (US$8.7M) per year. The vision-augmented workflow cut defect-escape rate from 3.1% to 0.6%, an 81% reduction, eliminating customer-complaint events for two consecutive quarters. Inspector headcount remained unchanged: the team reallocated to root-cause analysis and supplier-quality engineering, raising upstream defect catch by 34%. Time per inspection dropped from 47 seconds to 19 seconds, allowing the same headcount to absorb a 22% line-speed increase without overtime.

Three Tier-1 automotive customers cited the QC improvement in their annual supplier scorecards, advancing the conglomerate from B to A grade with two of them.

Lessons

The Cynefin sequencing call (complicated vs complex) saved the program from being one large failed initiative. Inspector co-design with union representation removed the labour-relations risk that had killed two prior vendor pitches. Telemetry quality, not model quality, set the ceiling on predictive-maintenance value at every plant.