Machine Vision QC in Nonwoven Production
Customer claims fell from 4.2% to 0.7% of shipped volume; investment recovered in 14 months from claims reduction alone.
The challenge
Manual inspection at line speed was statistically incapable of catching the defect types European buyers were rejecting shipments for.
Customer claims were averaging 4.2% of shipped volume — replacement shipments, credit notes, and the kind of sustained quality noise that erodes contract positions with European buyers. The root cause was not production; the manufacturing process was capable. The root cause was that inspection was operating below the threshold needed to catch what the process occasionally produced. The gap between 'we inspect everything' and 'we catch everything worth catching' was costing the business in ways that were visible on every invoice but invisible in the inspection process itself. The secondary problem was documentation: EU buyers on public infrastructure contracts increasingly required per-roll quality records. There was no mechanism to produce them.
The system
Decision system built
A machine vision inspection system was installed on the primary production line — two high-speed line-scan cameras spanning the full web width, connected to a GPU-based defect detection processor running a convolutional neural network trained on 12,000 labelled defect images from the company's own quality archive. A beta-gauge sensor array at 50cm cross-web intervals was integrated to surface gramage deviation data. The system flags rolls exceeding a configurable defect density threshold in real time, generates a per-roll defect map and weight profile, and outputs a structured quality record compliant with EN documentation requirements — automatically, at line speed.
System components
Dual high-speed line-scan cameras with full-web coverage (4.4m width)
GPU-based defect detection processor with CNN trained on proprietary defect image library
Beta-gauge cross-web sensor array for real-time gramage deviation detection
Configurable defect density threshold logic with automatic roll flagging
Per-roll automated quality documentation output (defect map, weight profile, dimensional measurement)
EN-aligned quality record generation for EU buyer compliance
How we worked
Engagement scope
Full deployment scoping, camera and sensor selection calibrated to line geometry and defect taxonomy, CNN training on client-owned defect image archive, integration with existing production line controls, documentation output configuration to EN standard requirements, operator training and threshold calibration support.
Timeline
90-day phased deployment: hardware installation and initial model training in weeks 1–6, supervised parallel running (manual + machine vision) in weeks 7–10, full production handover with threshold tuning in weeks 11–12.
Operating model
On-site commissioning team with production engineering involvement at every stage. Threshold sensitivity and defect classification reviewed jointly with quality managers before go-live. Post-deployment model retraining protocol established using flagged borderline cases — the system improves with production volume.
Outcomes
Business impact & measurable results
Customer claims fell from 4.2% to 0.7% of shipped volume; investment recovered in 14 months from claims reduction alone.
Defect detection rate improved from ~45% (manual inspection) to 91% (machine vision) within 90 days of deployment
Customer claims fell from 4.2% to 0.7% of shipped volume — an 83% reduction — with claims reduction alone recovering the full system investment in 14 months
Production line speed subsequently increased by 15%, previously constrained by manual inspection throughput
Automated per-roll quality documentation became a commercial differentiator with EU buyers requiring EN-standard quality records, enabling access to specification-grade tender positions previously unavailable
Governance
Trust, collaboration & governance
Training data drawn exclusively from the client's own defect archive — no generic model assumptions about what their defects look like
Parallel running period before manual inspection was decommissioned, with explicit sign-off criteria agreed in advance
Configurable threshold logic kept in operator hands — production supervisors adjust sensitivity without vendor involvement
Documentation output reviewed against EN standard requirements by the client's compliance team before EU buyer submission
Reframe
Claims reduction paid the investment back; CE-compliant documentation unlocked tender segments previously closed to the business.
Across every engagement, the goal is the same: engineer a system that makes better decisions — faster, more consistently, and at scale — than the process it replaces.
Insights
Related perspectives.
Next steps
Related services
Start a discovery
Most engagements begin with a conversation about context.
We do not send a proposal before we understand the problem. Start by telling us about your decision context — we will identify the highest-leverage intervention areas before any scope is agreed.