Computer Vision Development in Dublin

Pharmaceutical / Manufacturing

An Irish pharmaceutical manufacturing site (producing injectable biologics, 4.2 million vials annually, HPRA-inspected, supplying EU and global markets) operated a manual visual inspection process for finished product vials. Inspection requirements: each vial was inspected for particulate matter (visible and sub-visible contamination), container integrity (cracks, chips, and seal defects), fill volume (visual assessment against reference marks), and label quality (print clarity, alignment, and content). The site employed 32 visual inspectors working in 3 shifts (24/7 operation). The problems: (1) Detection consistency -- inspection accuracy declined measurably across shifts: first-hour detection rates averaged 97.2%, but final-hour rates dropped to 91.4% due to inspector fatigue. The site's quality team estimated that approximately 0.8% of defective vials passed inspection during high-fatigue periods. (2) Throughput limitation -- manual inspection limited line speed. The site was commissioning a new filling line with 40% higher output, and the manual inspection bottleneck would require adding 12 additional inspectors (difficult to recruit and train) or accepting reduced line speed. (3) Inspector training and retention -- training a visual inspector to the required standard took 6 months, and annual turnover was 22%. The site continuously operated with 3-4 inspectors in training who could not yet inspect independently. (4) Documentation -- each inspection required documentation (inspector ID, batch number, inspection time, any rejected vials). Paper-based documentation consumed approximately 15% of inspection time and created data entry errors that quality assurance corrected during batch record review. (5) The HPRA expected that any automated inspection system would be validated to demonstrate equivalent or superior performance to manual inspection, with documented evidence of defect detection capability.

We developed a validated automated visual inspection system over 16 weeks: (1) Multi-Camera Inspection Station -- designed a 5-camera inspection station integrated into the filling line: two cameras for particulate detection (backlighting the vial and capturing images while the vial was spun to mobilize particles), one camera for container integrity (high-resolution imaging of the vial body, neck, and crimp seal), one camera for fill volume (side-profile imaging with reference mark detection), and one camera for label inspection (high-resolution label reading with OCR verification against batch specifications). (2) Defect Detection Models -- trained convolutional neural networks for each defect category: particulate detection (trained on 45,000 annotated vial images including clean vials and vials with various particle types and sizes), container defect detection (trained on 22,000 images including intact vials and vials with cracks, chips, and seal irregularities), fill volume assessment (trained on 15,000 images covering the acceptable fill range and common underfill/overfill conditions), and label quality verification (OCR with verification against batch-specific print specifications). (3) GAMP 5 Validation -- the system was validated following GAMP 5 software category 5 (custom application) requirements: design qualification (DQ) documenting requirements and design decisions, installation qualification (IQ) verifying hardware and software installation, operational qualification (OQ) testing system performance under normal operating conditions with known defect samples, and performance qualification (PQ) comparing automated inspection against manual inspection on production batches. PQ included a statistical study comparing defect detection rates: 10,000 vials inspected by both human inspectors and the automated system. (4) Electronic Batch Records -- inspection results were recorded electronically with each vial's inspection outcome, defect classification, inspection timestamp, and camera images of rejected vials. This electronic documentation eliminated paper records and provided complete traceability for HPRA inspection. (5) Reject Handling -- defective vials were automatically diverted to a reject station with categorized collection bins (particulate, container defect, fill volume, label). Reject vials were retained for quality review, enabling correlation analysis and root cause investigation for recurring defect types.

Defect detection rate improved: the automated system achieved 99.1% detection rate in PQ validation -- compared to 94.6% average for human inspectors (and 91.4% during high-fatigue periods). The improvement was most significant for sub-visible particulate detection, where the automated system detected particles that human inspectors consistently missed under fatigue. Throughput increased: the automated inspection station operated at the new filling line's full speed, eliminating the manual inspection bottleneck. The site achieved the 40% output increase without adding inspection headcount. False positive rate: the automated system initially rejected 2.4% of good vials (false positives), which was reduced to 1.1% through model refinement in the first 3 months of production. This compared to approximately 0.6% for human inspectors. The quality team accepted the higher false positive rate given the significantly better true positive (defect detection) rate. Inspector redeployment: of the 32 inspectors, 18 were redeployed to other quality functions (incoming material inspection, in-process checks, and quality review roles), 8 were transitioned to inspection system operation and reject review roles, and 6 departed through natural attrition. Documentation efficiency: electronic batch records eliminated 100% of paper inspection documentation, reducing batch record review time by 35%. The HPRA inspected the automated system during a routine GMP inspection and accepted the validation evidence and system performance data. The site estimated annual value at EUR 3.2M: EUR 1.4M in throughput improvement (40% higher output without inspection bottleneck), EUR 0.9M in inspection quality (reduced defective product reaching market, avoiding potential recalls), EUR 0.5M in labour efficiency (smaller inspection team), and EUR 0.4M in documentation efficiency (electronic records replacing paper). The system was approved for replication at the company's second Irish manufacturing site.