Quality Leads Bridge Inline Traffic-Light Patrols and Edge Computer Vision

The Traffic Light System is a random inline inspection method that places Green, Yellow or Red cards at the workstation. Fibre2Fashion describes checks of semi-finished pieces from the current bundle once every hour or two, with one inline checker covering 15 or 20 workstations. Its checks include stitches per inch, seam width, sewing allowance, trim placement and measurements.[S37] The source's example assigns Green when no rework is found, Yellow at two or more defective items out of ten and Red at five or more out of ten. It also states that factories alter sample sizes and thresholds to suit their quality targets.[S37] Those thresholds are examples, not a universal standard.

GFT's illustrative commissioning protocol uses four patrols per eight-hour shift and a seven-piece sample from the current bundle. Green means zero defects in that sample, Yellow means one or two non-critical defects and Red means three or more defects, or one critical defect. Cameroon's standard working week provides the shift context, while the actual route must respect legally applicable working-time and break requirements.[R16] The seven-piece rule is deliberately labelled as a local starting design. The quality lead must revise it against buyer requirements, defect risk and observed process capability rather than treating it as a sourced industry norm.

Randomness needs operational discipline. The inspector should draw pieces from different positions in the current bundle without letting the operator preselect them, and patrol times should vary enough to avoid a predictable performance for the camera or card. The inspector records the bundle and style identifiers so that a Red result can be traced to all pieces produced since the last accepted check. A Green card means only that the sampled pieces contained no observed defect under that inspection method. It does not prove that the station is incapable of producing a defect.

Yellow is a coaching and verification signal. The inline inspector identifies the exact deviation, the operator corrects the next piece, and the line supervisor checks whether the cause lies in method, material or machine setup. Red is a containment signal. The operation stops, the suspect population is quarantined, and the line supervisor, zone mechanic and quality lead decide who must act. GFT's illustrative restart rule requires five consecutive defect-free pieces after correction. A card cannot be cleared merely because production is behind target.

The protocol also needs protection against social display. A station card communicates process status; it must not become a public ranking of personal worth or a reason for intimidation. The quality lead audits whether inspectors apply the same defect definitions across operators and shifts. Recurring Yellow or Red results trigger training, method engineering or maintenance, not automatic punishment. This is consistent with the source's emphasis on immediate corrective action and clarification of quality requirements at the operation.[S37]

Exhibit 26 assigns the decision rights for GFT's proposed patrol and escalation routine. The thresholds are illustrative and intentionally distinct from the source example.

The table separates detection from disposition. An inline inspector can place a card and preserve evidence, but the quality lead controls release after a Red event. The line supervisor restores flow only after the technical cause and the quarantined population have been addressed. The zone mechanic owns machine correction where relevant, while the industrial-engineering technician owns an unrealistic method or work standard. This prevents production urgency from silently overriding quality authority.

A final limitation matters: patrol sampling cannot guarantee zero defects between checks. Its value lies in earlier warning, clear containment and organisational learning. Where a critical defect can cause injury, the control plan must require immediate stop and containment on discovery, regardless of sample counts or the next scheduled patrol.

A usable quality record must distinguish the number of defects from the number of garments affected. Defects per Hundred Units (DHU) counts every recorded nonconformity, divides by units inspected and multiplies by 100. ISO 2859-1 uses the same fundamental distinction between nonconforming items and nonconformities per 100 items, and notes that one item may contain more than one nonconformity.[S0701] First-Pass Yield (FPY), in GFT's internal Quality Log definition, counts units that pass without repair, stain removal or stitch alteration and divides them by all units presented for first inspection. DHU measures defect frequency; FPY measures the flow of units through the process without a second touch.

An illustrative inspection of 250 units that finds 20 individual defects produces DHU of 20 ÷ 250 × 100 = 8.0. If 232 units pass on their first presentation, FPY is 232 ÷ 250 × 100 = 92.8 percent. The two results are not complements. Eighteen units failed first pass in this example, but twenty defects were counted because some units carried more than one defect. Adding DHU and FPY would therefore be a category error. The Quality Log must preserve both measures with the same unit population, inspection boundary and time period.

The minimum useful record contains more than a total. Each event needs a timestamp, line, workstation, style, bundle or roll identifier, defect code, severity, inspection stream and disposition. Human observations and camera alerts use the same defect taxonomy, but the source stream remains visible so that detection performance can be compared honestly. A vision event also retains its model version and a reference to the review image or bounding-box metadata. A corrective-action record then names the accountable owner and the evidence required before closure.

This structure allows a Pareto review to direct engineering attention. The daily huddle examines the leading recurring codes at the affected stations. The weekly review separates immediate containment from corrective action on the cause and preventive action that stops recurrence on another line. A repair to one garment closes a disposition task, not a CAPA. Closure requires evidence that the cause has been removed or controlled across a defined later production population.

Exhibit 27 shows a constructed defect mix for a greenfield shirt line. It is a planning illustration of how a Quality Log can change as recurring causes are removed, not an observed GFT ramp.

The exhibit demonstrates why a falling DHU total is not enough to explain improvement. The quality lead should be able to state which codes declined, what physical change produced the decline and whether inspection coverage stayed constant. A change in sampling intensity, product mix or defect definitions can mimic progress. Equally, an initial rise after vision deployment may mean better detection rather than worse production. Interpretation must begin with the denominator and the measurement system before celebrating the trend.

The Quality Log also creates the evidence needed for fair supervision. Operators should not be blamed for a defect cluster caused by incorrect machine timing, off-tolerance cut panels or an ambiguous specification. The station record allows the quality lead to test those causes with the mechanic and industrial engineer. Quality governance then becomes an engineering routine rather than a search for a culpable operator.