Warp-Knitted Fabric Online Defect Inspection System

Overview

Multiaxial warp-knitted fabrics are widely used in key composite parts for wind turbine blades, automotive lightweighting and ships. Fabrics are wide, output is high and downstream requirements are strict, so a single missed defect can mean a scrapped roll or a customer complaint. Most lines still rely on visual inspection:

• Wide fabric, many blind spots: at 2–3.5 m wide, the eye cannot cover the whole width, and attention drops over long shifts.
• Limited coverage: visual sampling covers about 60%–70%, and detection and judgment vary by operator.
• High value per roll: a roll can be worth tens of thousands of yuan, so escaped defects multiply losses with output.
• Harsh environment: glass fiber dust is bad for operator health, making long patrol shifts hard to sustain.

The Warp-Knitted Fabric Online Defect Inspection System replaces manual patrols with machine vision and deep learning. It scans both sides across the full width, identifies defects in real time, raises alarms immediately and keeps a full trace.

Compared with manual inspection

Item	Manual inspection	Online inspection
Coverage	About 60%–70% (sampling)	Continuous full width
Sides	Usually front only	Front and back simultaneously
Consistency	Varies by operator	Uniform algorithm criteria
Fatigue and blind spots	Significant	Continuous, no fatigue
Traceability	Difficult	Meter mark + cross position + image
Quality statistics	Manual records	Automatic statistics, trends and reports

Key advantages

• Two-sided inspection: 6 cameras per side, staggered 200–400 mm along the running direction so the light paths do not interfere; back-side defects are detected as well.
• Full width at high resolution: 6 cameras stitched across a field of about 2.94 m, fitting 2.4–2.85 m fabric, with stable detection of defects ≥1 mm.
• AI recognition: classic image processing screens candidates and lightweight deep learning classifies them; color imaging separates color defects such as dirty yarn and colored threads.
• Speed headroom: line-scan mode reaches about 100 fps while current conditions need only 5 fps, leaving room for faster lines.
• CPU-only inference: a multi-core industrial PC handles real-time inference with lower hardware cost, power and cooling, and no GPU driver or CUDA maintenance.
• Line integration: takes new roll, running and speed signals and outputs stop, audible/visual alarm and lighting control; joints are only recorded by meter mark, without alarms or stops.

Architecture

• Image acquisition: 12 color global-shutter cameras with low-distortion lenses and IP65 housings; 2 high-CRI white LED strobe lights per side, triggered by encoder / line speed.
• Edge computing: an industrial PC takes the front and back data streams on two separate gigabit NICs for pre-processing, AI inference, defect location and storage; the two sides are networked separately so a fault on one does not affect the other; UPS protects against power loss.
• Application and interfaces: live view, defect management, reports and model management in one platform; audible/visual alarms, PLC stop interlock and reserved MES reporting.

Workflow

1. Synchronized capture: encoder / line speed triggers cameras and strobe lights together
2. Pre-processing: channel separation, contrast enhancement, ROI cropping and cross-direction stitching
3. Candidate screening: adaptive thresholding + morphology extract suspicious regions
4. AI classification: a lightweight CNN outputs 10 labels, with size measurement and confidence filtering
5. Response and traceability: graded alarms / stop interlock, with defect data and images stored

10 defect types, graded response

Defects are graded by production impact, and the alarm, prompt and stop action of each grade can be switched in the software.

Grade	Defects	Default response
A severe	Tape, gap, dirty yarn, colored thread	Alarm, optional line stop
B moderate	Stain, fuzz ball	Alarm prompt
C minor	Insect, wood chip, paper scrap	Record and prompt
Process feature	Joint	Meter mark recorded only, no alarm or stop

Software

• Live inspection: 12 camera views side by side with defects overlaid in real time, meter mark and status always visible; Class A defects flash red with a buzzer.
• Defect management: filter by time, type and grade; view original images, annotations and meter marks; confirm or reject by manual review.
• Reports: statistics by shift / day / week / month, defect type distribution and density trends, export to Excel / PDF.
• Model management: training data matched to model versions; start training, deploy and roll back.
• Parameters: defect thresholds, ROI and the response for each grade adjustable without code changes.
• Access and monitoring: operator / engineer / administrator roles with audit logs; online monitoring of cameras, lighting, disk and network.

Delivery and service

Full delivery from site survey and factory acceptance test (FAT) to site acceptance test (SAT), with continued sample collection and model iteration after go-live. 12-month warranty, remote response ≤4 hours during warranty, and key spare parts such as cameras and lights supplied for ≥5 years.

Notes

The above is the standard configuration; width, speed and other parameters can be recalculated for site conditions. Miss and false alarm rates are achieved after training on ≥200 defect samples per type; final values are subject to the contract technical appendix and site acceptance.