Weld And Coating Quality Computer Vision

The following discussion illustrates a project that is well suited to the capabilities of an independent consultant in the Umbrex Shipbuilding & marine systems Practice. This is an illustrative example. Umbrex consultants adapt their methodology, timeline, and deliverables to the specific needs of each client.

1) Client Situation

We engaged with a naval and commercial vessel shipyard within the Aerospace & Defense sector seeking to strengthen quality inspection performance by deploying computer vision across radiography, ultrasound, and visual imagery to detect weld defects, coating anomalies, and dimensional deviations. The client operated a multi-program yard with steel and aluminum hull fabrication, modular block assembly, and outfitting activities. Several process and performance gaps indicated a need to modernize Non-Destructive Testing (NDT) and coating inspection with AI-driven detection and human-in-the-loop verification aligned to classification society and defense program requirements.

Key pain points we documented included:

• Inspection throughput and backlog: Interpretation of radiographic testing (RT) and ultrasonic testing (UT, PAUT, TOFD) images was constrained by Level II/III inspector bandwidth, driving a backlog of films and digital radiographs. KPI underperformance included long cycle time from scan capture to disposition (often measured in days) and a growing inspection backlog measured in hull-days.
• Defect escapes and rework: Non-conformances (NCRs) surfaced late in the assembly sequence due to missed indications or inconsistent interpretation, increasing rework hours per hull section and affecting first-time yield. KPIs underperforming included defect escape rate, weld repair ratio, and rework hours per block.
• Coating quality variability: Dry Film Thickness (DFT) compliance, holiday detection, and surface preparation standards varied by contractor and shift. Coating rework ratio and DFT out-of-tolerance percentages were above target ranges, with inconsistent photo documentation and limited traceability to surface prep grades (e.g., ISO 8501) and environmental conditions logged at time of application.
• Dimensional deviations: Visual inspections and metrology checks identified dimensional non-conformities post-assembly (e.g., edge misalignment, undercut visibility, fillet weld size discrepancies), creating schedule churn at key milestones (e.g., block roll-up acceptance). KPI impacts included block acceptance rate on first pass and NCRs per linear meter of weld.
• Data fragmentation: NDT image data resided across legacy film archives, mixed CR/DR repositories, and UT device memory, with incomplete adherence to DICONDE naming conventions. The QMS lacked consistent linkage between NDT images, weld maps, WPS/PQR identifiers, welders’ IDs, and MES production operations, impeding closed-loop learning.
• False calls and variability: Variability in interpretation across inspectors led to higher false call rates for benign indications (e.g., geometry echoes, spurious artifacts), increasing unnecessary repairs and secondary inspections. Inspector utilization was suboptimal due to re-review loops.
• Compliance and audit risk: Audit trails for classification society surveys and defense quality representatives were manual and inconsistent. Model-assisted decisions lacked documented lineage and evidence frameworks acceptable under NAVSEA, ABS, DNV, or Lloyd’s Register oversight.
• Workforce and capability constraints: Experienced Level III NDT personnel were scarce, with retirement risk and limited time for training Level II personnel. Coaching moments were ad hoc, and knowledge capture from experts into standardized decision rules was limited.
• IT/OT integration gaps: Edge environments in fabrication bays and dry docks had intermittent connectivity. Devices from multiple OEMs (e.g., Olympus, GE/Krautkramer, Mistras) lacked standardized APIs into the data lake, and the cyber posture needed alignment with ITAR, DFARS, CMMC, and internal RMF controls for controlled unclassified information (CUI).
• Economic leakage: Cost of Poor Quality (CoPQ) was elevated due to scrap, rework, schedule slips, and customer chargebacks for late discovery of non-conformances. The value case for AI assistance was recognized, but prioritization and execution discipline were not yet defined.

These issues directly related to the capability “Weld And Coating Quality Computer Vision: Deploy computer vision on radiography, ultrasound, and imagery to detect weld defects, coating anomalies, and dimensional deviations, reducing escapes and rework in shipyard inspections.” The scope emphasized accelerating detection, standardizing interpretation against codes (e.g., ISO 5817/17635, NAVSEA, AWS D1.1, ASME IX), and creating an auditable, human-in-the-loop workflow that integrated with MES/QMS for closed-loop corrective and preventive actions (CAPA).

2) Project Objective

Primary objective: We established and piloted an AI, Data & Analytics solution using computer vision to assist NDT and coating inspectors in detecting weld defects, coating anomalies, and dimensional deviations across radiography, ultrasound, and imagery, thereby reducing escapes, compressing interpretation cycle time, and enabling consistent, auditable dispositions.

Secondary objectives included:

• Performance: Reduce inspection interpretation lead time and backlog by introducing triage and prioritization via AI-assisted pre-screening and anomaly flags.
• Economics: Lower rework labor and consumables by earlier detection and standardized disposition aligned with acceptance criteria, and quantify a defensible investment case for wider rollout.
• Compliance: Embed audit-ready evidence capture and traceability meeting classification society and defense quality standards, including secure handling of CUI and adherence to DICONDE for NDT images.
• Customer impact: Increase confidence with naval program offices and class societies through transparent, consistent, and explainable defect identification and record-keeping.
• Digital/data enablement: Create a governed NDT/Coatings data backbone, standardize metadata, and integrate devices and repositories into a lakehouse with model lifecycle management (MLOps).
• Capability-building: Train Level II inspectors and weld/coating supervisors to use AI-assisted tools, apply structured decision rules, and operationalize human-in-the-loop workflows.
• Risk reduction: Implement model risk controls, validation protocols, and rollout safeguards to prevent overreliance on AI and to ensure human oversight for acceptance decisions.
• Time-to-value acceleration: Focus the pilot on high-defect-risk weld classes and coating zones to accelerate learning curves and demonstrate operational feasibility.

3) Methodology and Approach

Workstream 1: Current-State Diagnostic of NDT and Coating Quality Operations

We conducted a comprehensive diagnostic of weld and coating inspection operations across fabrication, assembly, and outfitting areas. Activities included:

• Process mapping of RT, CR/DR, UT (shear wave, PAUT, TOFD), visual inspection, magnetic particle (MT), and penetrant (PT) workflows from scan scheduling through disposition and NCR creation.
• Time-and-motion studies capturing cycle time from image capture to Level II interpretation to Level III verification, including re-review loops and backlog accumulation points.
• Sampling review of historical non-conformances and weld repair records by class (e.g., lack of fusion, porosity, slag inclusion, undercut, root concavity, incomplete penetration) and coating anomalies (e.g., thin DFT, overspray, pinholes/holidays, blistering, adhesion failure, rust back).
• Assessment of acceptance criteria usage (ISO 5817 quality levels, NAVSEA spec references, AWS D1.1), and variation in disposition decisions by inspector and program.
• Device and data landscape inventory: UT sets, RT/CR/DR systems, coating gauges (DFT), holiday detectors, high-resolution cameras; repository structures; DICONDE conformance; and MES, PLM, QMS interfaces.
• Capability maturity baseline for data governance, labeling practices, and training protocols for inspectors, welders, and coaters.

We used structured frameworks for quality system maturity, inspection value stream analysis, and critical-to-quality (CTQ) flowdown from customer and class society requirements. The diagnostic isolated where AI assistance would reduce interpretation variability, compress cycle time, and support auditability without altering the inspector’s accountability.

Workstream 2: Data Architecture, Ingestion, and Labeling Enablement

We designed a data architecture under the AI, Data & Analytics capability group that ingested, standardized, and governed NDT and coating inspection data. Activities included:

• Establishing secure data pipelines from RT/CR/DR systems and UT devices into an on-premises or hybrid lakehouse with DICONDE-compliant metadata, mapped to weld maps, WPS/PQR IDs, welder IDs, heat numbers, and MES operations.
• Implementing device connectors and APIs (where supported) and fallbacks for secure file transfer and automated metadata extraction. We defined data contracts with OEMs to ensure reliable, versioned data artifacts.
• Standing up a labeling environment using CVAT/Label Studio with controlled taxonomies: defect classes per code, severity gradations, dimensional deviation tags, and coating anomaly categories aligned to standards (e.g., ISO 8501 surface prep grades and SSPC/NACE equivalents).
• Curating a training set and holdout set spanning hull materials, thickness ranges, joint configurations (butt, fillet), positions (1G–6G), and coating systems (epoxy, polyurethane, anti-fouling), ensuring representation across programs and environmental conditions.
• Applying data quality gates: rejecting low-contrast radiographs, de-noising UT signals, artifact masking (e.g., backscatter, scatter-to-primary ratio issues), and normalizing image orientation and scaling for downstream models.

This workstream created the data foundation and controlled vocabularies necessary for reproducible model training and future reuse. It addressed data lineage, access controls, and CUI segregation to satisfy defense cybersecurity requirements.

Workstream 3: Model Development for Radiography, Ultrasound, and Visual/Coating Imagery

We developed and evaluated specialized models per modality, with explainability and reviewability tailored for human-in-the-loop workflows:

• Radiography (RT/CR/DR): We built defect localization and classification models using convolutional neural networks and transformer-based detectors (e.g., YOLO-class architectures and segmentation networks such as U-Net/Mask RCNN), trained on DICONDE images. Preprocessing included contrast enhancement, scatter artifact removal, and geometric standardization. We targeted indications such as lack of fusion, porosity clusters, slag inclusions, and cracks, with heatmap overlays for explainability.
• Ultrasound (UT, PAUT, TOFD): We implemented signal and image-based models that operated on B-scan and S-scan data, with algorithms for echo-pattern analysis, time-of-flight features, and multi-channel fusion. We emphasized distinguishing geometry echoes from flaw indications and assessing through-thickness positional estimates to support code-based acceptance checks.
• Visual/coating imagery: We trained computer vision models for weld surface inspection to detect undercut, overlap, excessive reinforcement, arc strikes, and dimensional deviations (e.g., leg length for fillet welds using calibrated references). For coatings, we used anomaly detection for surface defects (pinholes, overspray, sags) and correlated DFT readings to image regions via registration techniques.
• Anomaly detection and weak supervision: We incorporated autoencoders and one-class classifiers where labeled defect data were scarce, combined with weakly supervised learning to leverage inspector notes and NCR narratives.
• Data augmentation and domain adaptation: We augmented training data for material thickness, exposure variations, and transducer settings, and applied domain adaptation to generalize across devices and shifts.
• Validation protocols: We designed defect-class confusion analyses, per-code acceptance threshold simulations, and inter-rater agreement benchmarking against Level II/III interpretations, with extensive documentation in model cards and validation plans.

The model portfolio was aligned to risk reduction goals: accelerate triage, highlight likely indications for review, and standardize evidence presentation for consistent decisions. We did not automate final disposition; instead, we structured model outputs for inspector decision-making and audit trails.

Workstream 4: Workflow, UI/UX, and Human-in-the-Loop Integration

We re-engineered inspector workflows to embed AI assistance without disrupting compliance or accountability:

• Designed a review console presenting modality-specific views, heatmaps, signal traces, and model confidence, with one-click code references and acceptance criteria.
• Built queues for triage based on model-prioritized risk scores, linking back to MES operations and weld maps for spatial context.
• Implemented structured decision capture: indication marking, code-based disposition, and rationale selection, feeding back to the training set via active learning.
• Established escalation rules for Level III review on edge cases, with digital sign-offs and version-controlled result packages for audits and class surveys.
• Created coaching modes where Level III annotations and comments were visible to Level II inspectors for continuous upskilling.

This workstream focused on adoption, interpretability, and consistency. We minimized additional clicks compared to legacy systems and aligned UI nomenclature to the shop floor vernacular and code language inspectors used daily.

Workstream 5: Edge Deployment, MLOps, and Systems Integration

We executed a robust deployment pattern suitable for shipyard environments:

• Edge inference: Deployed models on ruggedized edge devices (e.g., GPU-enabled industrial PCs or embedded accelerators) inside RT and UT interpretation rooms to mitigate connectivity issues, using ONNX/TensorRT optimization for low-latency analysis.
• MLOps and CI/CD: Stood up model registries, versioning, canary rollouts, and shadow modes to compare AI predictions with human interpretations before enabling production assist features.
• Integration: Connected the review console to QMS for NCR creation, MES for operation status updates, PLM for drawings and weld maps, and data lake for centralized archiving and analytics. Implemented OPC UA or vendor APIs where supported.
• Security and compliance: Segmented networks and applied role-based access controls, logging, and encryption aligned to CMMC practices. Ensured DICONDE storage conformed to retention and audit policies, with traceable model versions linked to each inspection decision.

This deployment framework supported time-to-value by minimizing downtime and by enabling controlled, reversible changes governed by change control boards (CCBs) within the QMS.

Workstream 6: Governance, Model Risk Management, and Regulatory Alignment

We established governance for AI-assisted inspections that would withstand regulatory and customer scrutiny:

• Defined a model risk management framework including intended use statements, limitations, known failure modes, and monitoring thresholds.
• Created validation protocols aligned to internal quality procedures and external requirements (e.g., NAVSEA, ABS/DNV classes), including acceptance testing, periodic revalidation, and documentation templates.
• Set up an oversight committee with Quality, Welding Engineering, NDT Level III leadership, IT security, and Program Quality to approve model changes and workflow adjustments.
• Developed an audit trail schema: model inputs, outputs, overrides, rationale codes, and digital signatures stored with inspection records; reports formatted for classification society surveys and defense customer reviews.

This governance foundation reduced adoption risk and provided clear decision rights for rollout sequencing and scale-up.

Workstream 7: Training, Change Management, and Capability Build

We designed and delivered training programs for inspectors, welders, coating teams, and supervisors:

• Curriculum: Use of AI-assisted tools, interpretation of heatmaps and signal overlays, code-based decision rules, and structured feedback entry for active learning.
• Job aids: Quick reference guides mapping defect classes to code clauses, common false-positive patterns, and model confidence interpretation.
• Coaching: Shadow sessions pairing Level II inspectors with Level III mentors using recorded cases to build consistency and capture tacit decision rules.
• Performance management: Defined adoption and proficiency metrics (e.g., review time per scan, agreement rates, override patterns) to guide continuous improvement.

This ensured the solution improved inspector effectiveness without eroding professional judgment or accountability embedded in the QMS.

Workstream 8: Value Case, Pilot Scope, and Scale Roadmap

We quantified the investment case and structured the pilot to de-risk scale-up:

• Value levers: Inspection throughput, backlog reduction, earlier defect detection yielding lower rework scope, standardization lowering false calls, and reduced schedule volatility at block acceptance gates.
• Pilot scope: Selected representative weld classes and coating zones with measurable defect incidence and available labeled data; staged rollout across one fabrication bay and one paint shop line.
• Scale roadmap: Defined expansion by modality and hull area, with readiness criteria including model validation metrics, inspector adoption rates, IT security sign-off, and class society engagement feedback.

The roadmap sequenced investments, tied decision gates to validation evidence, and outlined vendor partnerships and integration milestones.

4) Data Request

To execute the project, we requested the following datasets and system access, with expected time horizons and granularity. We identified common data quality pitfalls and proposed remediation steps in parallel.

• Commercial/program data:
  – Program master schedules and key quality milestones (e.g., weld/NDT gates, block roll-ups) for the last 12–24 months.
  – Customer and classification society requirements mapped to weld categories and coating systems.
  – Change logs for contract quality clauses and concessions.
• Operational data:
  – Weld maps and joint lists by block/hull with attributes: joint type, position, material, thickness, WPS/PQR references, welder IDs, dates.
  – NDT schedules and completion logs by modality, inspector IDs, queue times, and re-review actions over 12–18 months.
  – Coating operations logs: surface preparation grades, environmental conditions (temperature, humidity, dew point), DFT readings by zone, holiday test results, applicator IDs.
• NDT image and signal data:
  – Radiography (CR/DR) images in DICONDE format with metadata; at least 12–24 months of history, focusing on welds with recorded defects and accepted indications.
  – UT data: B-scan, S-scan, PAUT/TOFD files with transducer setup parameters, calibration blocks references, and A-scan where available.
  – Associated interpretations: indication markings, defect codes, severity classifications, dispositions, and Level III concurrence notes.
• Visual/coating imagery:
  – High-resolution images of weld surfaces pre/post-weld, and post-cleaning; calibrated images when available for dimensional checks.
  – Coating application images pre/post-application; evidence of masking, surface cleanliness, and defect annotations.
  – DFT measurement data linked to image regions or zone IDs.
• Quality records:
  – NCRs tied to welds and coating zones, repair instructions, and closure notes; CAPA records and trend analyses.
  – Audit and survey findings from ABS/DNV/Lloyd’s and defense quality representatives.
• Technical/engineering:
  – WPS/PQR library, code references (ISO 5817/17635, AWS D1.1, NAVSEA, ASME IX), acceptance criteria matrices by joint class and program.
  – Drawings and 3D models necessary for weld map context and dimensional evaluation rules.
• People and governance:
  – Inspector qualifications (Level II/III), training records, and shift schedules.
  – QMS procedures for NDT/Coatings, document control, and change control.
• Systems landscape:
  – MES, PLM, ERP, QMS system details, data models, and integration points.
  – Existing DICONDE archives, image repositories, and UT device management systems.
  – Network topology for bays/docks, connectivity constraints, and security policies relevant to CUI.

Granularity and horizons: We targeted record-level data for at least 12–24 months to ensure seasonal and program variability coverage. Common pitfalls included inconsistent DICONDE metadata, image orientation mismatches, missing linkage between images and weld maps, incomplete UT setup metadata, and unstructured NCR narratives. We addressed these with data cleaning, standardization scripts, controlled vocabularies, and labeling playbooks.

5) Questions for Client

• What acceptance criteria and codes govern the target weld classes and coating systems (e.g., ISO 5817 quality levels, NAVSEA, AWS D1.1), and where is there program-specific tailoring?
• Which hull areas and joint types contribute most to rework and schedule volatility, and what is the prioritized scope for the initial pilot?
• What are the current inspector staffing levels by shift (Level II vs. Level III), and where do interpretation backlogs cause downstream delays?
• How does the QMS define decision rights for NDT and coating dispositions, and what documentation is required for internal and classification society audits?
• Which RT/CR/DR and UT devices are in scope, and what interfaces or export formats (including DICONDE) are available for data extraction?
• What cybersecurity requirements apply (ITAR, CMMC level, RMF/ATO boundaries), and how should CUI be segregated and logged within the AI tooling?
• How are NCRs initiated and routed today, and what integration points exist with MES/QMS to enable closed-loop CAPA?
• What constraints exist for edge compute in interpretation rooms (power, space, network), and what is the preferred deployment pattern (on-prem, hybrid)?
• Which classification society representatives or defense quality authorities need to review the proposed AI-assisted workflow and validation protocols?
• What is the tolerance for false positives during pilot triage, and how should we gate the progression from shadow mode to assist mode?
• Are there known reliability issues with image quality (e.g., film digitization artifacts, UT calibration drift) that should inform data quality gates?
• What training and change management support will inspectors and supervisors need, and who will champion adoption on each shift?
• How should we measure success in the pilot (e.g., interpretation lead time, agreement rates, re-review reduction) without asserting production impact at this stage?
• What capital and operating budget is available for devices, edge hardware, and platform licenses to support scale-up?
• What are the legal and union considerations related to the use of AI assistance in inspection and quality roles?

6) Interview Guide for Subject Matter Experts

NDT Level III (Radiography and Ultrasonic)

• Which defect types most frequently challenge Level II interpretations, and what visual or signal cues distinguish flaws from geometry or artifacts?
• How do you apply acceptance criteria under borderline conditions, and what additional evidence do you seek before dispositioning?
• Where do device setup parameters (e.g., PAUT focal laws, RT exposure settings) most influence detectability and false indications?
• What are common sources of image quality degradation, and how do you mitigate them in practice?
• Which documentation elements must accompany an interpreted indication to satisfy audits and customer reviews?
• How would you structure escalation thresholds and overrides in a human-in-the-loop AI workflow?

Welding Engineer

• Which WPS/PQR combinations and joint configurations exhibit higher defect incidence, and what process controls are in place?
• How do you translate code requirements into shop floor decision rules for inspectors and welders?
• What dimensional deviations at the weld (e.g., mismatch, reinforcement) most affect downstream assembly and acceptance?
• How can inspection feedback be used to adjust WPS parameters or welder training in a closed-loop fashion?
• What metadata about welds (e.g., heat input, interpass temperature) would enhance predictive defect detection?
• Where do you see risks in using AI-assisted detection relative to welding procedure compliance?

Coatings Superintendent/Inspector

• Which coating systems and surface prep steps most commonly drive rework or NCRs?
• How do you document DFT, holiday tests, and environmental conditions today, and where are traceability gaps?
• What visual cues in images indicate likely adhesion issues or pinhole clusters before failure manifests?
• How would you prefer AI to present suspected anomalies for quick validation and disposition?
• What variance in applicator technique impacts quality, and how could the system flag at-risk zones early?
• How should coating anomaly categories be defined to align with standards and the QMS?

Quality Manager (QMS/AS9100)

• What are the documentation and sign-off requirements for inspections, NCRs, and CAPA, and how must AI-assisted steps be recorded?
• How do you manage change control for procedures affecting NDT and coating inspections?
• What audit findings or customer feedback have highlighted gaps in traceability or consistency?
• Which metrics does leadership review for inspection performance and quality costs, and how are targets set?
• How will you govern model changes and periodic validation within the QMS?
• What training and competency records must be maintained for inspectors using AI-assisted tools?

IT/OT Architect and Cybersecurity Lead

• What is the current data flow from NDT devices to repositories, and where are the integration pain points?
• What are the requirements for handling CUI and DICONDE data at rest and in transit, including logging and access controls?
• How reliable is network connectivity in fabrication bays and interpretation rooms, and what edge deployment constraints exist?
• Which identity, secrets, and certificate management systems should be used for AI services?
• What is the path to authorization (ATO) for new software components in the yard’s environment?
• How will you monitor and incident-manage AI components as part of OT operations?

Program Quality/Classification Society Liaison

• What evidence packages and formats do ABS/DNV/Lloyd’s or defense quality reps expect for inspections and non-conformances?
• How do you propose engaging external stakeholders on AI-assisted inspection methods and validation protocols?
• What concerns have surveyors raised about digitization of inspection records, and how were they addressed?
• What notification or approval processes are required before changing inspection workflows?
• Which programs are most receptive to piloting AI assistance, and where is the risk tolerance lowest?
• How should model documentation be structured to align with survey expectations?

Production Superintendent (Fabrication/Assembly)

• Where do inspection delays create critical path impacts in the production schedule?
• Which blocks or zones experience the most rework due to late defect discovery?
• How should inspection priority be set to prevent downstream schedule slips?
• What constraints (crane time, bay handoffs) influence when and how inspections can occur?
• What feedback loop from inspection to production would be most actionable for supervisors and welders?
• What adoption concerns do you anticipate from crews regarding AI-assisted inspection?

7) Timeline

We executed a 12-week plan structured around discovery, data enablement, model development, pilot deployment, and validation, with clear decision gates tied to AI-assisted weld and coating inspection readiness.

• Weeks 1–2: Discovery and Diagnostic
  – Conducted process mapping, backlog analysis, and KPI baselining for RT, UT, and coating inspections.
  – Completed device and data landscape inventory; confirmed governance and cybersecurity requirements.
  – Decision gate: Confirmed pilot scope (weld classes, coating zones), target modalities, and stakeholders’ alignment.
• Weeks 2–4: Data Architecture and Labeling Setup
  – Built ingestion pipelines from RT/CR/DR and UT devices into the lakehouse with DICONDE compliance.
  – Established labeling environment and taxonomies; curated initial training and holdout datasets.
  – Implemented data quality gates and augmentation strategy.
  – Decision gate: Approved datasets and labeling standards for model training.
• Weeks 4–7: Model Development and Validation Protocols
  – Developed modality-specific models (RT, UT, visual/coating), implemented preprocessing and explainability overlays.
  – Drafted validation protocols: inter-rater comparisons, acceptance simulations, and error analysis plans.
  – Decision gate: Authorized shadow-mode deployment based on offline validation thresholds.
• Weeks 6–9: Workflow Integration and Edge Deployment
  – Deployed review console in shadow mode in interpretation rooms; integrated with MES/QMS for read-only links.
  – Configured MLOps (model registry, telemetry) and security controls; trained pilot users.
  – Decision gate: Transition plan from shadow to assist mode contingent on user acceptance and validation metrics.
• Weeks 9–11: Assist-Mode Pilot and Human-in-the-Loop Tuning
  – Enabled AI-assisted triage for selected welds and coating zones; captured overrides and rationale.
  – Held daily huddles with Level II/III inspectors to refine thresholds and UI usability.
  – Decision gate: Governance committee review for expanded scope and integration for NCR initiation.
• Week 12: Validation, Governance Setup, and Handoff
  – Finalized validation reports, model cards, workflow SOPs, and auditor-ready documentation.
  – Defined scale roadmap with phase criteria, training plan, and budgetary estimates.
  – Handoff of playbooks, runbooks, and dashboards to Quality and IT/OT teams.

Critical path items included access to high-quality labeled NDT data, Level III availability for protocol design and validation, and security approvals for edge software deployment. Decision gates ensured readiness and compliance before each progression.

8) Deliverables

• Current-State Diagnostic and KPI Baseline
  – A detailed narrative of inspection workflows, bottlenecks, backlog drivers, and quality cost levers, including baseline metrics such as interpretation cycle time, false call rates, NCR rates, and DFT compliance variability.
• Data Architecture and DICONDE Integration Blueprint
  – A technical blueprint covering ingestion from RT/UT devices, metadata standards, DICONDE mapping, data lakehouse schemas, access controls, and retention policies for audit-ready storage.
• Labeling Taxonomy and Playbook
  – Controlled vocabularies for defect classes, severity scales, dimensional deviation categories, and coating anomalies, plus labeling SOPs and quality checks to ensure consistent annotations.
• Modality-Specific Model Cards and Validation Protocols
  – Documented model purpose, training data characteristics, preprocessing steps, performance metrics, known limitations, and validation methods aligned to code acceptance criteria and inter-rater benchmarks.
• Inspector Review Console and Workflow SOP
  – UI guide and SOP for AI-assisted triage and review, including evidence capture, escalation rules, override logging, and digital sign-off procedures integrated with QMS.
• Edge Deployment and MLOps Runbook
  – Configuration guides for edge hardware, model serving, CI/CD pipelines, monitoring and telemetry, rollback procedures, and asset lifecycle management.
• Security and Compliance Control Set
  – Cybersecurity architecture, CUI handling procedures, role-based access controls, logging standards, and documentation to support internal IT security reviews and ATO processes.
• Human-in-the-Loop Governance Framework
  – Model risk management plans, change control workflows, oversight committee charters, periodic revalidation schedules, and audit trail schemas.
• Training Curriculum and Job Aids
  – Training materials for Level II/III inspectors, welding and coating supervisors, and quality managers, including case libraries and quick reference guides for defect dispositions.
• Pilot Validation Report
  – Evidence summaries from shadow and assist modes, user acceptance feedback, error analyses, and recommendations for threshold settings and next-phase expansion.
• Integration Design with MES/QMS/PLM
  – Interface specifications, data mappings, and event triggers enabling NCR creation, operation status updates, and traceable linkage to weld maps and drawings.
• Investment Case and Scale Roadmap
  – Phased rollout plan with cost estimates, resource plan, partner/OEM engagement strategy, and decision gates for expanding modalities and shop areas.

9) Industry Insights

Deploying computer vision in shipbuilding and marine systems for weld and coating quality sits at the intersection of NDT digitization, AI adoption in safety-critical environments, and defense/commercial regulatory rigor. Several trends informed our approach and the design decisions taken during this project.

• Shift from film to digital NDT: Shipyards continue to transition from film radiography to computed/digital radiography and advanced ultrasonics (PAUT/TOFD). This unlocks data for analytics but introduces variance in device configurations and metadata. DICONDE is increasingly the standard for archival and interoperability, and its disciplined use is a foundational enabler for AI.
• Inspector scarcity and skill transfer: Level III expertise is constrained as experienced inspectors retire. AI assistance that captures tacit decision rules and provides explainable overlays can extend expertise and standardize interpretation, provided that model risk and human oversight are formally governed.
• Classification society and defense acceptance: ABS, DNV, Lloyd’s, and defense quality authorities expect audit-ready evidence, consistent application of standards, and clear accountability. AI usage is acceptable when positioned as an assistive tool, with documented validation, traceable decisions, and human sign-off retained.
• Digital thread and closed-loop quality: Leading yards link inspection data to weld maps, WPS/PQRs, materials, and production operations in MES/PLM. This digital thread enables faster feedback to welding procedures, welder training, and coating process controls, as well as predictive quality analytics targeting at-risk zones.
• Edge-first deployment: Connectivity variability and sensitive data constraints drive edge inference for low-latency review and secure handling of CUI. Robust MLOps practices, version control, and telemetry are required to support operational reliability and regulatory audits.
• Explainability and human factors: In high-consequence industries, inspectors require clear visualization of model rationales and fast ways to confirm or override suggestions. UI/UX design must mirror code language, present evidence efficiently, and minimize click burden to drive adoption.
• Standards-aligned taxonomies: Using defect classes and acceptance levels aligned to ISO 5817/17635, AWS, ASME, NAVSEA, and coating standards allows trained models and labeling practices to scale across programs without rework. Consistency in taxonomies and SOPs improves inter-rater agreement and audit defensibility.
• Ecosystem and OEM integration: UT and RT OEMs vary in their data export capabilities and APIs. Early engagement and data contracts avoid vendor lock-in and ensure sustained access to high-fidelity data. Partnerships with labeling/tool vendors and cloud/on-prem providers streamline time-to-value.
• Cybersecurity and compliance posture: CMMC progression, DFARS clauses, and internal RMF requirements shape where and how AI tools can operate. Designing for principle of least privilege, encryption, and comprehensive logging is as critical as model accuracy in this sector.
• What “good” looks like: Mature implementations show standardized DICONDE metadata, integrated inspection queues linked to MES/QMS, inspector adoption of assistive tools with consistent override patterns, and periodic revalidation workflows. KPIs typically targeted include reduced interpretation lead time, lower re-review rates, improved agreement with Level III, and earlier identification of at-risk zones for proactive intervention.
• Near-term disruptions: Transformer-based vision models, self-supervised learning on large unlabeled NDT datasets, and synthetic data generation are increasing performance where labeled data are scarce. Yards are also piloting private 5G for reliable device connectivity and AR-enabled guidance for inspection and repair. These trends will amplify the benefits of well-governed AI in inspection when combined with robust data foundations.

For clients in Shipbuilding & marine systems, the imperative is to combine AI, Data & Analytics with rigorous quality governance, strong integration into MES/QMS/PLM, and disciplined change management. The approach we executed created a compliant, inspector-centered workflow that leveraged computer vision on radiography, ultrasound, and imagery to assist detection of weld defects, coating anomalies, and dimensional deviations. This positioned the client to reduce escapes and rework risk through earlier, more consistent, and audit-ready inspection decisions while building the internal capabilities to scale across hulls, programs, and modalities.