Nested Two Stage Variance Component Decomposition for Industrial Continuous Web Sett Auditing
Nested two-stage variance decomposition isolates between-roll drift from within-roll position gradients to enforce contract sett tolerances in bulk fabric orders.

Structure
Thread counts in woven and knitted webs drift over the course of a production run. Loom mechanics, yarn feeding, and finishing tension all introduce physical variance. An industrial web audit requires quantifying warp end and weft pick counts under standardized atmospheric conditioning; quoting a nominal target such as 110 ends per inch and 90 picks per inch means very little without defined statistical limits.
Woven fabric sett directly drives structural mass, tensile breaking strength, air permeability, and total finished yield per kilogram of yarn, so unmonitored shifts quickly create commercial defects.
Process instability shows up across several operational timescales. Gradual drift develops roll to roll as yarn lots change or sizing mixes vary. Cyclical swings occur inside individual rolls from beam eccentricity, let-off friction spikes, or take-up roll slip.
Across the web width, stenter clip chains pull bows and skews into the drying fabric. A count taken at a single physical point conflates all of these factors: long-term mill drift, roll-to-roll variance, local spatial banding, and gauge measurement error.
Standard ISO 7211-2 pick counts evaluated without multi-point spatial sampling obscure within-roll density gradients exceeding four percent.
Pinpointing the mechanical origin of this variation requires auditing the manufacturing sequence directly. Modern air-jet looms run past 800 picks per minute, where small mechanical pulses build into longitudinal density waves. On circular knitting frames, feeder-to-feeder tension imbalances produce periodic course-density bands across the tube.
Downstream, jet dyeing and stenter frames release tension or induce thermal shrinkage, shifting web geometry yet again. When a mill quality lab relies on a single pick-glass count, it risks passing fabric that later skews, shrinks unevenly, or shades across panels in the cutting room.
Conflating these distinct variance sources causes buyers to misread production issues. Scrapping an entire twenty-roll shipment over one swatch below minimum pick density penalizes the weaver for what might only be isolated stenter pinlet slippage. Conversely, approving a lot based on a single compliant cut from the roll head leaves the apparel plant exposed to shading and sizing splits if the center thins out.
A functional audit distinguishes broad lot drift from localized spatial swings, preventing panel mismatch before cut goods reach sewing lines.

Sampling
Hierarchical sampling gives structure to selecting inspection points across bulk shipments. In this framework, primary units are individual rolls pulled at random from a dye lot. Secondary units are longitudinal points along each selected roll, typically head, middle, and tail cuts taken at predetermined meter marks.
Tertiary units consist of replicate counts taken across the left, center, and right lanes at each cut. This matches the physical production hierarchy: positions exist inside specific rolls, and lane replicates exist inside specific longitudinal positions.
A balanced two-stage nested sampling plan provides the mathematical foundation needed for variance component analysis. The inspector draws a rolls from a delivery of N rolls. Along each selected roll, the inspector marks b longitudinal stations spaced at least ten metres apart to eliminate local spatial autocorrelation.
Across each station, c replicate counts of warp ends and weft picks are recorded across the fabric width. Balancing sample counts across all tiers keeps the calculations straightforward and maximizes statistical power for hypothesis testing.
| Sampling Stage | Physical Entity | Sample Count | Spacing / Conditioning Rule | Primary Target Variance |
|---|---|---|---|---|
| Stage 1 (Primary) | Fabric Rolls | a = 5 rolls per lot | Randomized selection across weaving beams | Between-roll variance (σ2B) |
| Stage 2 (Secondary) | Web Positions | b = 4 cuts per roll | Minimum 15-metre longitudinal separation | Within-roll position variance (σ2W) |
| Stage 3 (Tertiary) | Replicate Counts | c = 3 reads per cut | Left, Center, Right across usable width | Measurement gauge error (σ2e) |
Sample locations also require strict conditioning control. Swatches taken for destructive analysis require 24 hours of flat relaxation in an ISO 139 standard atmosphere at 20 degrees Celsius and 65 percent relative humidity. When using optical counters mounted on inspection frames, running web tension must be accounted for: web tension pulls warp ends taut and reduces weft pick density per unit length during continuous movement.
Measuring fabric under line tension on an uncalibrated re-winder distorts the baseline and artificially inflates the tertiary error term.
Destructive sampling from multiple internal roll points routinely encounters operational resistance. Cutting swatches from interior positions splits continuous export rolls into short-length remnants of lower commercial value, prompting mills to offer end-cuts instead. But swatches taken exclusively from roll tails capture edge-cooling artifacts from stenter heat-setting and tension adjustments made during roll changes.
Data gathered exclusively from roll ends presents an artificially volatile picture of fabric sett that fails to represent the stable middle portion of the production run.

Arithmetic
Mathematical decomposition of variance in a two-stage nested statistical model separates the total observed variance of web sett into additive, independent orthogonal components. Let Yijk represent the measured thread count for replicate k at position j within roll i. The linear statistical model takes the explicit structural form:
Yijk = μ + αi + βj(i) + εk(ij)
The term μ denotes the overall mean fabric sett across the entire population. The random variable αi represents the deviation of roll i from the mean, assumed normally distributed with zero mean and variance σ2B, representing between-roll variance. The random variable βj(i) represents the deviation of position j nested within roll i, assumed normally distributed with zero mean and variance σ2W, representing within-roll position variance.
The random variable εk(ij) represents the residual measurement and micro-spatial error, assumed normally distributed with zero mean and variance σ2e. The total variance of any individual measurement equals the direct sum σ2total = σ2B + σ2W + σ2e.
Calculating the sum of squares proceeds through classical Analysis of Variance partitioning. The total sum of squares partitions into three distinct quadratic forms representing the structural tiers of the sampling design:
SSTotal = sumi=1a sumj=1b sumk=1c (Yijk – barY. )2
SSRolls = bc sumi=1a (barYi. – barY. )2
SSPositions(Rolls) = c sumi=1a sumj=1b (barYij. – barYi. )2
SSError = sumi=1a sumj=1b sumk=1c (Yijk – barYij.)2
Dividing each sum of squares by its corresponding degrees of freedom yields the Mean Squares (MS). For a balanced design with a rolls, b positions per roll, and c replicates per position, the degrees of freedom are dfB = a – 1 for rolls, dfW = a(b – 1) for positions within rolls, and dfe = ab(c – 1) for residual measurement error. Equating observed Mean Squares to their Expected Mean Squares (EMS) derives the individual variance components directly:
E = σ2e + cσ2W + bcσ2B
E = σ2e + cσ2W
E = σ2e
| Source of Variation | Degrees of Freedom | Sum of Squares | Mean Square | Expected Mean Square | Calculated Variance Component |
|---|---|---|---|---|---|
| Between Rolls (αi) | 4 | 14.40 | 3.600 | σ2e + 3σ2W + 12σ2B | σ2B = 0.239 (πcks/in)2 |
| Positions within Rolls (βj(i)) | 15 | 10.95 | 0.730 | σ2e + 3σ2W | σ2W = 0.187 (πcks/in)2 |
| Residual Measurement Error (εk(ij)) | 40 | 6.80 | 0.170 | σ2e | σ2e = 0.170 (πcks/in)2 |
| Total Variation | 59 | 32.15 | — | — | σ2total = 0.596 (πcks/in)2 |
Consider a practical audit scenario evaluating a nominal 90 picks per inch cotton twill lot. The sample parameters use a=5 rolls, b=4 longitudinal positions per roll, and c=3 counting head replicates per position, providing N=60 total observations. The analysis yields MSRolls = 3.600, MSPositions = 0.730, and MSError = 0.170.
Solving the expected mean square equations step-by-step establishes the numeric components:
σ2e = MSError = 0.170
σ2W = fracMSPositions – MSErrorc = frac0.730 – 0.1703 = 0.187
σ2B = fracMSRolls – MSPositionsbc = frac3.600 – 0.73012 = 0.239
The total variance sums to 0.596 (πcks/inch)2, corresponding to a total standard deviation of 0.772 picks per inch. Between-roll variance contributes 40.1 percent of total variation. Within-roll position variance accounts for 31.4 percent.
Measurement error contributes 28.5 percent. When missing data occurs due to damaged web edges or lost swatches, classical ANOVA expected mean squares lose orthogonality. Under unbalanced conditions, restricted maximum likelihood (REML) estimation replaces the ANOVA algebraic partitioning.
REML uses iterative convergence on the linear mixed model likelihood function to yield unbiased numerical variance estimates without requiring equal group sizes.
Expressing variance components as percentage contributions isolates the dominant operational failure mode across the chain. A high percentage contribution from σ2B points to loom setup inconsistency, sizing batch variations, or mixed yarn lots across weaving beams. A high percentage contribution from σ2W indicates mechanical instability on the finishing line, such as fluctuating stenter overfeed or erratic take-up tension.
A high contribution from σ2e highlights improper operator technique, bad lighting, or poorly calibrated camera line-scanners.
Hypothesis testing evaluates whether between-roll and within-roll variance components differ significantly from zero. Testing within-roll position effects uses the test statistic F = MSPositions / MSError against an F-distribution with dfW and dfe degrees of freedom. In the worked twill example, F = 0.730 / 0.170 = 4.29.
Comparing this figure to the critical value F0.05, 15, 40 = 1.92 confirms that within-roll position variation is statistically significant at the 95 percent confidence level. Testing between-roll variation uses F = MSRolls / MSPositions = 3.600 / 0.730 = 4.93, exceeding the critical threshold F0.05, 4, 15 = 3.06. Both structural layers present real variance that demands mechanical correction on the factory floor.

Frame
Physical measurement of web sett on inspection tables relies on accurate hardware and strict operator protocols. Manual glass counting uses standardized pick glasses with ground optical lenses and calibrated apertures conforming to ISO 7211-2. The operator lays the fabric flat without applying tension, aligns the travel pointer with a warp yarn, and counts individual picks across a fixed distance of 25.4 millimetres or 50 millimetres.
Counting errors scale rapidly when fabric sett exceeds 120 threads per inch or when complex weave structures like satin or Jacquard obscure yarn boundaries.
Automated optical inspection systems mounted directly on inspection machinery eliminate operator eye fatigue and accelerate data acquisition. High-resolution line-scan cameras frame the moving web under high-intensity backlight or low-angle darkfield LED illumination. Fast Fourier Transform algorithms and spatial frequency analysis process the captured optical intensity profile, converting spatial peaks into real-time thread counts.
Optical systems operate continuously at line speeds reaching 100 metres per minute. Continuous measurement generates thousands of data points per roll, requiring algorithmic downsampling into discrete spatial clusters to match the two-stage nested statistical architecture.

What Distinguishes Stage One Variance from Stage Two Variation?
Stage one variance isolates macroeconomic changes that occur between distinct fabric rolls across a manufacturing contract. These variations stem from differences in raw material yarn lots, warp sizing formulations, beam winding tensions, loom setting calibrations, or separate jet-dyeing machine batches. Stage two variation isolates localized dimensional instability within a single continuous roll along its longitudinal length or across its transverse width.
Stage two shifts occur due to thermal gradients inside stenter drying chambers, mechanical slippage on take-up rollers, uneven web tension during batching, or localized relaxation during cooling.
A rigorous measurement system analysis (MSA) quantifies the gauge repeatability and reproducibility (Gauge R&R) component within the residual variance term σ2e. Repeatability assesses the variation obtained when one inspector uses the same pick counter on the same fabric location multiple times. Reproducibility quantifies the variation introduced when different inspectors or different camera stations measure the identical site.
If the combined gauge R&R variance exceeds 10 percent of total tolerance width, the auditing system lacks adequate resolution to distinguish real fabric density shifts from instrument noise.
- Instrument Calibration verifying optical magnification factors against certified optical reticles under standard illumination.
- Tension Isolation applying mechanical web clamps to isolate the measurement region from winder drive forces.
- Environmental Equilibrating holding swatches inside controlled conditioning spaces prior to benchmark optical scanning.
- Operator Blind Replicates presenting masked roll positions to human auditors to prevent observational expectation bias.
- Data Compression Protocol grouping raw line-scan spatial frequency logs into standardized ten-metre block averages.
Environmental conditions on the mill floor directly corrupt physical count accuracy if unmanaged. Relative humidity fluctuations alter hydrophilic fiber dimensions, causing rayon and cotton webs to expand or contract during inspection. A five percent shift in ambient humidity shifts cotton fabric picks per inch by up to 0.8 percent through moisture regain variations.
Inspection frames installed directly adjacent to wet processing areas experience localized microclimate shifts as stenter exhausts cycle on and off. Enclosing the measurement zone or performing audit cuts inside a dedicated conditioning room remains mandatory for audit integrity.
Unresolved measurement discrepancies occur when high-speed automated line-scan system data is compared directly to static manual glass counts. Optical camera systems read thread density on webs traveling under active mechanical tension, whereas laboratory technicians take manual counts on fully relaxed swatches. The tension difference creates a systemic bias between online mill logs and offline audit reports.
The precise mathematical transformation required to map dynamic online spatial counts to fully relaxed ISO 7211-2 state across varied fiber blends remains an open technical dispute between textile machinery builders and independent testing laboratories.

Yield
Decomposing variance components provides direct financial protection when structuring commercial fabric supply agreements. Standard specification sheets state nominal sett along with symmetric tolerance limits, such as 90 picks per inch plus or minus 3 picks. Without nested variance decomposition, a mill can meet the nominal target on average across a lot while producing individual rolls that exhibit extreme internal density gradients.
Extreme density gradients cause linear weight shifts, leading to garment weight non-compliance and shade banding after garment washing.
A buyer utilizes calculated variance components to construct exact Statistical Process Control (SPC) acceptance criteria. Total standard deviation σtotal = sqrtσ2B + σ2W + σ2e dictates the upper and lower specification limits that a production process can consistently maintain. Process capability indices, specifically Cp and Cpk, measure the mill’s ability to manufacture within buyer tolerances:
Cp = fracUSL – LSL6σtotal
Cpk = minleft( fracUSL – μ3σtotal, fracμ – LSL3σtotal right)
When an audit yields a Cpk below 1.33, the manufacturing process generates unacceptable defect rates. The nested variance model identifies exactly which parameter requires capital intervention to restore capability. If σ2B dominates, the mill must implement tighter raw yarn control, standardize beam preparation, or unify dyehouse batch profiles.
If σ2W dominates, the mill must overhaul stenter tension drives, repair worn take-up roll grips, or recalibrate drying chamber temperature controllers.
Process capability indices calculated without isolating within-roll variance systematically overestimate mill compliance by up to thirty percent.
Fabric purchasing contracts link monetary penalties directly to thread count non-compliance. Pick density directly governs raw yarn consumption per metre of fabric. A weaver that consistently runs two picks per inch below specification saves approximately 2.2 percent in raw yarn mass while staying within loose single-point inspection tolerances.
Over a 100,000-metre contract, this density reduction steals structural integrity from the cloth and alters garment drape. A nested variance audit detects systematic negative density bias even when individual measurements fall within broad nominal limits.
Structuring lot acceptance sampling plans around nested variance metrics protects buyers from accepting sub-standard goods. Traditional ISO 2859-1 single sampling plans evaluate rolls as simple pass-or-fail units based on visual defect counts. Continuous web sett auditing requires statistical operating characteristic (OC) curves based on the non-central t-distribution or F-distribution.
Setting the buyer risk (β risk) at 5 percent for receiving fabric with more than 1.5 percent out-of-spec yardage establishes the required sample size of rolls a and interior positions b.
Standard purchase order contract terms must explicitly incorporate nested variance component thresholds to remain legally enforceable. A robust contract addendum defines sampling intensity, standard testing conditions, maximum acceptable variance ratios, and financial restitution scales:
“Fabric lot acceptance requires a process capability index Cpk ge 1.33 calculated across a minimum two-stage nested sampling design of 5 rolls (a=5) and 4 positions per roll (b=4). If the calculated between-roll variance component σ2B exceeds 0.35 (πcks/inch)2 or the within-roll variance component σ2W exceeds 0.25 (πcks/inch)2 under ISO 7211-2 conditioned testing, the buyer reserves the right to reject the entire dye lot or apply a linear invoice price reduction equal to the percentage deficit in total fabric mass.”

