One-Page FRD — POC for Image-based Defect Checking
(Automotive Electronics)
1) Purpose / Background
Automotive electronics assemblies (PCB/ECU sub-assemblies) require consistent defect
detection (presence, orientation, solder/placement quality). This POC validates feasibility of
vision + AI to detect defects on 1–2 component types under controlled imaging, and to
quantify ROI drivers (inspection time reduction, scrap avoidance, escaped defect reduction).
2) POC Scope (1–2 Components)
Component A (Recommended): Connector / Pin-Header / Terminal Block
● Defects: missing pin/terminal, bent/misaligned pin, incomplete insertion, wrong
connector variant (optional).
Component B (Recommended): Polarized SMD (Diode/Electrolytic Capacitor/IC) OR
Solder-Joint Region
● Defects: missing component, polarity/orientation error, offset/misalignment; optional
stretch: solder bridge / insufficient solder (if imaging supports).
In-scope Line Stage: post-placement / post-reflow inspection station (offline or inline mock).
Out of scope (POC): full-board coverage, multi-SKU generalized model, AOI replacement
certification, root-cause analytics.
3) Goals & Success Criteria (Feasibility + ROI)
Primary Goals
1. Demonstrate detection on target components with stable performance.
2. Validate practical deployment constraints: lighting, camera placement, cycle time,
operator workflow.
POC Acceptance Criteria
● Recall (defect detection): ≥ 95% on “critical defects” (missing, polarity error, missing
pin).
● Precision / False Reject Control: ≤ 3–5% false rejects (configurable threshold).
● Latency: inference + decision ≤ 500 ms per image (or within takt constraint).
● Data Sufficiency: model performance stable across ≥ 3 production sessions / shifts
(lighting + operator variance).
● Operational Fit: operator can review flagged cases in UI in ≤ 10 seconds per case.
(Exact thresholds can be tuned per plant risk appetite; above are typical POC-grade targets.)
4) Users & User Stories
Primary Users
● Quality Inspector / Operator
● Process Engineer / Quality Engineer
● Manufacturing Manager
User Stories
● As an operator, I want the system to automatically flag defective components so I only
review exceptions.
● As a quality engineer, I want defect images + labels saved for traceability and model
iteration.
● As a manager, I want a daily summary of detection counts and false calls to judge ROI.
5) Functional Requirements (POC)
FR-1 Image Capture
● Capture high-res RGB images of defined ROIs for Component A and B.
● Store raw images + metadata: timestamp, line/station ID, PCB ID (barcode if available),
operator ID (optional).
FR-2 Inference & Decisioning
● Run AI model(s) to output:
○ defect class (e.g., missing pin / bent pin / polarity error / missing component /
misalignment)
○ confidence score (0–1)
○ bounding box / segmentation mask (where applicable)
● Apply configurable thresholds per defect type to classify: PASS / FAIL / REVIEW.
FR-3 Review Workflow (Human-in-the-loop)
● UI shows image, overlay (box/mask), predicted defect + confidence.
● Operator actions: Confirm Defect / Mark False Alarm / Reclassify (from a controlled list).
● Store operator feedback as ground truth for retraining.
FR-4 Reporting (POC-level)
● Simple dashboard/export (CSV is fine):
○ total inspected, fails, review count
○ defect distribution
○ false reject / false accept estimates (from sampled audits)
○ top recurring defect images
FR-5 Audit / Traceability
● For every decision, log: model version, threshold config, image hash, outcome, reviewer
action.
6) Non-Functional Requirements
● Reliability: station runs continuously for a shift with auto-restart on crash.
● Security: images stored in secured folder/bucket; role-based access for UI.
● Maintainability: model/version config via YAML/JSON; easy rollback.
● Deployment: local edge PC (GPU optional) + optional sync to central server.
7) Data & Model Approach (POC)
Data Needed
● For each component type:
○ ~500–2,000 “good” samples (to learn normal variance)
○ 50–200 samples per defect type (even synthetics/seeded defects acceptable in
POC)
● If real defects are rare: use seeded defect boards + controlled perturbations.
Model Strategy
● Component A (Pins): detection + keypoint/geometry checks OR segmentation-based
missing/bent detection.
● Component B (Polarized SMD): classification on ROI + orientation verification; optional
anomaly detection if defect labels scarce.
8) Assumptions & Dependencies
● Stable imaging setup: fixed camera mount, controlled lighting, consistent focal distance.
● Ability to identify board/ROI consistently (fixture or fiducial alignment).
● Access to at least one SME for labeling and defect taxonomy.
9) Risks & Mitigations
● Reflection / glare / solder shine: add dome light / polarizer; control exposure.
● Defect scarcity: seeded defect creation + active learning loop.
● SKU drift: limit POC to 1 PCB variant; plan generalization later.
● False rejects hurt throughput: start with REVIEW bucket; tighten thresholds gradually.
10) POC Deliverables
1. Imaging station spec (camera + lens + lighting) + ROI definition for 1–2 components
2. Trained model v1 + inference service (containerized)
3. Review UI + feedback capture
4. POC report: metrics, confusion matrix, root causes of misses, ROI estimate & next-step
plan
If you tell me which exact connector type (e.g., 8/16-pin header, FFC, terminal block) and the
second component you want (polarized cap vs IC vs solder joint), I’ll tailor the defect list + ROI
thresholds + data plan to that choice.