The utility relied on a time-intensive manual review process where engineers coded CCTV footage inconsistently, leading to unreliable data for strategic capital planning and network-wide infrastructure assessments.
.png){kind=small}
ML pipeline for defect detection and asset classification on critical water infrastructure
The utility relied on a time-intensive manual review process where engineers coded CCTV footage inconsistently, leading to unreliable data for strategic capital planning and network-wide infrastructure assessments.
Automated detection was hindered by years of backlog footage featuring extreme variations in lighting, resolution, and frame rates across diverse legacy camera systems, requiring a highly resilient processing pipeline.
A convolutional neural network, trained on 45,000 labelled frames, was designed to recognise standard PACP defect codes. The model provides precise localisation and confidence scores for anomalies like cracks, root intrusions, and structural deformations.
Detected defects were matched to GIS records using distance-along-main metadata extracted from the video stream. This spatial referencing allows the utility to analyse defect density by pipe material, age, and pressure zone for strategic planning.
The system runs on AWS using Docker and Kubernetes to automate the ingestion and processing of new inspection files. A web-based interface allows engineers to verify low-confidence detections, feeding corrections back into the model for ongoing refinement.
To handle diverse pipe materials from brick to PVC, VE3 developed material-conditional routing. This architecture directs video segments to specialist classifiers trained on the unique visual signatures of specific materials, significantly improving detection accuracy.
VE3 developed a custom ETL module to parse proprietary metadata from multiple camera systems. This required reverse-engineering the schemas of legacy vendors to ensure accurate timestamp and distance synchronisation across the entire fleet.
VE3 built an end-to-end pipeline using adaptive histogram equalisation and noise reduction to standardise varied footage quality. This ensured consistent defect detection across both high-definition modern video and low-resolution legacy analogue captures.
Independent Verification & Performance
The model achieved 87.3% accuracy in defect classification during independent engineering validation, providing highly reliable data for PACP-compliant capital investment planning.
Time Reduction & Rapid Triage
Manual review time was reduced by 60%, enabling full triage of inspection footage within 48 hours—a process that previously required several weeks to complete.
.png)
Automated complex feature extraction within high-precision public sector environments, successfully reducing manual workloads by 60% while maintaining rigorous accuracy standards.
.png)
