Computer Vision Development in Stockholm

Manufacturing / Industrial Equipment and Components

A Stockholm-headquartered industrial manufacturing group (SEK 120B annual revenue, 35,000 employees, manufacturing industrial equipment, components, and tools across 22 factories in Sweden, Finland, Germany, and China — products used in mining, construction, and infrastructure worldwide) needed computer vision to automate quality inspection of precision-machined components, reduce field failures, and improve manufacturing efficiency. (1) Inspection complexity: the group manufactured 45,000 different component types — from small precision parts (5mm tolerance) to large structural assemblies (5-metre welded frames). Manual visual inspection required inspectors trained on specific product families, with knowledge of what constituted acceptable vs. rejectable variations for each component type. 180 quality inspectors across 22 factories, with new inspector training requiring 6-9 months per product family. (2) Field failure cost: precision components that passed visual inspection but contained undetected defects could fail in the field — a drill bit component failing underground, a compressor part failing in a remote oil platform, or a construction equipment component failing on a job site. Each field failure cost an average of SEK 850,000 (replacement component, emergency shipping, field technician, lost production for the customer, and warranty/goodwill costs). The group averaged 340 field failures per year attributable to inspection escapes — SEK 289M annually. (3) Surface inspection limitations: many defect types (micro-cracks, sub-surface porosity, stress corrosion) were invisible to standard visual inspection under normal lighting. Advanced inspection methods (dye penetrant, magnetic particle, ultrasonic) caught these defects but were too slow and expensive for 100% inspection — they were applied as sampling methods on 5-10% of production. Between the 90-95% not receiving advanced inspection and the human variability of visual inspection, the group estimated that 3-5% of components shipped with undetected quality issues. (4) Consistency across factories: quality standards were documented in inspection procedures, but enforcement varied across 22 factories in 4 countries. Swedish factories consistently outperformed Chinese and German factories on quality metrics — not due to equipment differences but due to inspector skill and experience variations. Quality consistency was essential for the brand's reputation for reliability. (5) Throughput constraint: inspection was a production bottleneck — components queued for inspection at several high-volume factories. Adding inspection capacity meant hiring and training inspectors (6-9 month ramp-up) or adding shifts. The group needed to increase inspection throughput without compromising quality.

We developed and deployed a visual inspection platform over 16 weeks. (1) Multi-modal inspection stations: camera systems combining standard visual inspection with enhanced detection methods — high-resolution cameras (5-micron pixel resolution for surface defect detection), structured light (3D surface scanning for dimensional verification and surface topology), UV fluorescence (detecting dye penetrant indications from automated dye application — bringing advanced inspection to 100% of production rather than 5-10% sampling), and thermal imaging (detecting sub-surface anomalies through thermal response patterns). Automated material handling: components automatically positioned, rotated, and imaged from multiple angles — each component inspected in 15-45 seconds depending on size and complexity (vs. 3-8 minutes for manual inspection). (2) Product-family models: deep learning models trained per product family — understanding what constitutes acceptable variation for each component type. Transfer learning: base models trained on general surface inspection, then fine-tuned on specific product families using the group's historical quality data (3.2 million annotated inspection images). New product onboarding: new components added through automated training from 200-500 sample images — reducing new product inspection readiness from 6 months (manual training) to 2 weeks (model training). Defect classification: 35 defect categories covering surface, dimensional, and material defects — each classified by type, severity, and likely root cause. (3) Predictive quality: inspection data correlated with manufacturing parameters (machine settings, tool wear, material lot, operator, environmental conditions) — identifying which process conditions produced which defect types. Tool wear prediction: visual inspection detecting progressive quality degradation correlated with tool wear — recommending tool change before defects exceeded tolerance (extending useful tool life while preventing defective production). Process drift detection: systematic quality trends identified from inspection data — alerting manufacturing engineers to process changes before they produced out-of-specification components. (4) Cross-factory standardisation: identical inspection systems across all 22 factories — ensuring consistent quality assessment regardless of location. Centralised model management: models trained and validated centrally, then deployed to all factories simultaneously — eliminating the factory-to-factory variation that manual inspection introduced. Quality benchmarking: real-time quality comparison across factories — identifying best practices and quality gaps for continuous improvement. (5) Field failure correlation: linking field failure reports back to manufacturing inspection data — determining whether the field failure was an inspection escape (detectable defect missed) or a previously undetectable failure mode. Root cause feedback: new failure modes discovered in the field were added to the inspection model — expanding detection capability based on real-world failure data.

Detection rate: from 82% (human visual inspection average across factories) to 96.5% (AI multi-modal inspection). Enhanced detection: UV fluorescence and thermal imaging detecting defects previously found only through sampling — 100% inspection replacing 5-10% sampling for advanced methods. Field failures: from 340/year to 85/year (75% reduction) — SEK 217M annual savings. Customer perception: quality complaints reduced by 62% — reinforcing the brand's reliability reputation in markets where equipment failure has severe consequences. Throughput: inspection time from 3-8 minutes to 15-45 seconds per component — eliminating inspection as a production bottleneck. Inspection capacity increased 8x without additional stations or staff. Cross-factory consistency: quality variation between factories reduced by 80% — Chinese and German factories achieving quality levels previously only reached by Swedish factories. Rework: defect identification at earlier process stages (predictive quality) reduced average rework cost per defect by 45% — catching issues before value-adding operations. Tool life optimisation: visual-based tool wear prediction extended average tool life by 12% while reducing tool-related defects by 35%. Inspector redeployment: 180 inspectors reduced to 65 (retained for complex products, audit, and system management) — 115 positions redeployed through attrition and retraining per consultation with Swedish trade unions. Annual value: SEK 217M (field failure reduction — 255 fewer field failures at SEK 850K average cost), SEK 145M (throughput improvement — eliminated inspection bottleneck enabling additional production capacity), SEK 85M (cross-factory quality — standardised quality across 22 factories reducing customer complaints and warranty costs), SEK 52M (tool and process optimisation — extended tool life and reduced rework through predictive quality), SEK 38M (inspector efficiency — reduced inspection staffing while improving detection) = SEK 537M (approximately USD 52M). Development cost: SEK 5.8M (approximately USD 560K). Payback: under 4 days.