Statistical Improvement, Flow, and Constraints

Six Sigma, Design for Six Sigma, Theory of Constraints, QRM and POLCA, TPM and OEE, and Zero Defect Manufacturing.

4.4 Six Sigma, DMAIC and Statistical Engineering

4.4.1 Theoretical basis

Six Sigma is a project-based variation-reduction method organised around the DMAIC sequence — define, measure, analyse, improve, control — with a supporting role infrastructure of trained specialists and a portfolio governance layer. Its statistical content is not novel; the design of experiments, regression, hypothesis testing and control charting on which it draws predate it by decades. Its innovation is organisational: it packages statistical method inside a project structure with defined roles, financial accountability and executive sponsorship.

Linderman, Schroeder, Zaheer and Choo (2003) supplied the missing theoretical underpinning by reading Six Sigma through goal theory. Their argument is that the method works, to the extent it works, because it imposes specific and challenging goals on well-structured tasks with defined feedback — the precise conditions under which goal-setting theory predicts strong performance effects. That reading also predicts where it fails: on ill-structured problems where the goal cannot be specified in advance, goal specificity impairs rather than improves performance.

4.4.2 Mechanism

The mechanism is sequential elimination of candidate causes under statistical control of error rates. The measure phase establishes that the measurement system is adequate to detect the effect of interest — a step whose omission invalidates everything downstream. The analyse phase reduces a large candidate set to a small vital set using observational data. The improve phase establishes causality experimentally, typically by fractional factorial design, and identifies operating settings. The control phase installs a mechanism that prevents reversion, which is where most projects fail.

4.4.3 Evidence

Grade A, with an important qualification about mechanism. Swink and Jacobs (2012) conducted an event study comparing 200 Six Sigma-adopting firms against matched controls, using several matching procedures on pre-adoption return on assets, industry and size. They report strong evidence of a positive effect on return on assets, together with a small positive effect on sales growth. Critically, the return-on-assets gain arose mostly from significant reductions in indirect costs; significant improvements in direct costs and in asset productivity were not evident.

This finding deserves more attention than it receives. It implies that the dominant realised benefit of Six Sigma adoption in the studied population was in overhead functions — rework administration, quality assurance labour, warranty and scrap handling, transactional processes — rather than in the direct conversion process that Six Sigma rhetoric emphasises. For a device manufacturer, where documentation, deviation management and investigation labour are a large and growing overhead, this is not a disappointing result; it points at the highest-yield target.

4.4.4 Implementation protocol (project level)

1. Define. Write a charter with a defect defined at the unit level, a measurable baseline, a target expressed as a defect-rate or capability delta, and a validated financial model. Reject charters whose problem statement contains a proposed solution.

2. Measure. Conduct measurement system analysis before collecting improvement data. For an automated optical inspection system, this means an attribute agreement analysis against a truth panel of known-defective and known-conforming parts, not merely a gauge repeatability and reproducibility study on a continuous gauge. Establish baseline capability with a rational subgrouping scheme that separates within-batch and between-batch variation.

3. Analyse. Stratify, then use multi-vari or components-of-variance analysis to locate the dominant variance source (within-part, part-to-part, cavity-to-cavity, batch-to-batch, time-to-time) before hypothesising causes. Skipping this step is the most common cause of wasted experimental effort.

4. Improve. Screen with a resolution IV fractional factorial, then characterise with a response surface where curvature is present. In a validated process, run screening on a development or qualification asset where possible, and reserve production-asset experimentation for confirmation.

5. Control. Install the control mechanism in a hierarchy of preference: eliminate the failure mode by design; then error-proof; then automate the control; then chart it; then procedure it. Charting is fourth, not first. Update the process control plan, the failure mode and effects analysis, and the risk file.

6. Close. Verify sustainment at 60 and 180 days against the control metric, with finance confirmation of realised benefit.

4.4.5 Worked example — ophthalmic

Problem: base-curve radius on a mid-power soft lens exhibits process capability of Cpk = 0.94 against a bilateral specification, producing a dimensional rejection rate of roughly 3,100 parts per million at final inspection.

Measure: attribute agreement analysis on the dimensional station returns 91 percent agreement against the truth panel, with disagreement concentrated near the lower specification limit — meaning a material fraction of the observed rejection is measurement-driven. The measurement system is corrected before any process work proceeds, and the baseline rejection rate is restated.

Analyse: a nested components-of-variance study decomposes total variation into cavity-to-cavity (54 percent), batch-to-batch in hydration (26 percent), within-cavity across time (14 percent) and residual (6 percent). The dominant term is cavity-to-cavity, which redirects the project from the process-parameter hypothesis that the team began with toward tooling.

Improve: a 2^(5-1) fractional factorial on monomer dose, cure profile, tool temperature, demould delay and hydration ionic strength, run on the qualification tool set, identifies a tool-temperature by cure-profile interaction that accounts for most of the cavity-dependent shift. A response-surface follow-up locates a settings region within which the cavity effect falls below the resolution of the measurement system.

Control: the settings region is written into the process control plan as a design space rather than as a point setting, which — as discussed in 4.5 — allows movement within the space without a new change-control event. Capability at 180 days is Cpk = 1.62. The regulatory work, not the statistics, occupies the majority of the elapsed project time, consistent with the nine-month cycle documented by McGrane and colleagues (2022).

4.4.6 Statistical engineering: the discipline above the toolbox

Hoerl and Snee have argued for two decades that the field has an unfilled layer between statistical thinking (strategic, qualitative) and statistical methods (tactical, quantitative), and have proposed statistical engineering to occupy it — defined as the study of how best to use statistical concepts and methods, integrated with information technology and other disciplines, to achieve enhanced results (Hoerl and Snee 2012; Hoerl and Snee 2014). The practical content is a discipline for attacking large, unstructured, high-consequence problems for which no single technique suffices: problem framing, strategy selection, sequencing of methods, and integration of results across studies.

This matters in device manufacture more than in most sectors, because the characteristic hard problems — an intermittent particulate excursion, a supplier-correlated yield shift, a field complaint signal with no in-process correlate — are precisely the ill-structured class on which the goal-specific DMAIC structure performs worst. Statistical engineering is the correct frame for that class. Practically, it means: build a strategy before selecting tools; expect to run several linked studies rather than one; and treat the sequence of studies, not any individual study, as the unit of design.

4.4.7 Limitations and failure modes

• Solution-in-the-charter. A charter that names the fix converts the project into an implementation exercise with statistical decoration.

• Measurement system neglect. The single most common technical failure. In automated-inspection-dominated processes it is also the most consequential, because the inspection system output feeds every downstream model.

• Portfolio drift toward the easy. Project selection gravitates to tractable low-value problems when governance is weak. Risk-file-driven selection is the counter.

• Ill-structured problems forced into DMAIC. Predicted by the goal-theoretic reading (Linderman et al. 2003) and observed constantly. Route these to statistical engineering.

• Benefit inflation. Absent finance validation at 180 days, reported savings systematically exceed realised savings.

4.4.8 Primary metrics

Process capability indices by critical-to-quality characteristic; defects per million opportunities; rolled throughput yield; cost of poor quality as a percentage of cost of goods sold, split into internal failure, external failure, appraisal and prevention; project cycle time by phase (to expose the validation bottleneck); finance-validated benefit at 180 days as a proportion of claimed benefit at closure.

4.5 Design for Six Sigma, Robust Design and the Design Space

4.5.1 Theoretical basis

The unifying idea is that the cost of removing a variation problem rises by roughly an order of magnitude at each lifecycle stage — design, process development, production, field — while the freedom to remove it falls. Design for Six Sigma addresses the design stage using an IDOV or DMADV sequence, deploying quality function deployment to translate user needs into engineering characteristics, robust parameter design (in the Taguchi tradition) to select settings that minimise sensitivity to uncontrollable noise, and tolerance analysis to allocate tolerance where it is cheapest to hold.

The regulated-industry expression of the same idea is Quality by Design, formalised for pharmaceuticals in the ICH Q8(R2), Q9, Q10 and Q13 framework. Its central construct is the design space: the multidimensional combination of material attributes and process parameters demonstrated to provide assurance of quality. Movement within an approved design space is not considered a change requiring regulatory notification. The device sector has no exact statutory analogue, but the underlying logic — establish a characterised operating region rather than a point setting, and demonstrate assurance of quality across it — maps directly onto ISO 13485 process validation and is increasingly reflected in how firms structure their validation packages.

4.5.2 Mechanism

Robust design works by exploiting interactions between control factors and noise factors. If a control factor interacts with a noise factor, then some settings of the control factor make the response less sensitive to that noise. Finding those settings reduces output variance without controlling the noise itself — which is the only economically viable route when the noise is ambient humidity, incoming monomer lot variation, or operator technique.

The design space works by a different and, in a validated environment, more powerful mechanism: it converts what would be a sequence of change-control events into a single characterisation event. A process operated at a fixed validated point must raise a change for every parameter adjustment. A process operated within a characterised, validated region can adjust freely within that region. Over a product lifecycle of ten to fifteen years, the difference in cumulative regulatory effort is large.

4.5.3 Evidence

Grade B. Direct controlled evidence for Design for Six Sigma as such is limited, and the literature is dominated by case reports. The stronger evidence comes from the pharmaceutical Quality by Design corpus, where the regulatory framework has forced systematic adoption and generated a substantial applied literature on design-space definition, critical quality attribute identification, and the use of process analytical technology for real-time control. Regulatory agencies including FDA and EMA have promoted the approach explicitly through the Process Analytical Technology initiative, and the accumulated implementation experience — spanning conceptual frameworks through to industrial application — constitutes a coherent, if largely non-comparative, evidence base.

4.5.4 Implementation protocol

1. Define the quality target product profile: the performance, safety and usability attributes the device must deliver, expressed quantitatively with acceptance criteria.

2. Identify critical quality attributes by risk assessment against the ISO 14971 risk file — the attributes whose variation materially affects safety or performance.

3. Build the cause-and-effect matrix linking material attributes and process parameters to each critical quality attribute; classify parameters as critical, key or non-critical with documented rationale.

4. Characterise experimentally. Use screening designs to identify the significant subset, then response-surface or mixture designs to map the response over the candidate region. Include noise factors explicitly as design factors where possible, or use an inner-array/outer-array structure.

5. Define the design space as the region over which assurance of quality is demonstrated, with the demonstration evidence — not merely the boundary — documented.

6. Define the control strategy: which parameters are controlled by equipment, which by procedure, which by in-process monitoring, and which by end-product testing. Prefer control by design over control by inspection at every opportunity.

7. Validate across the design space rather than at a single point, and structure the validation report so that the space, not the point, is the validated object.

8. Establish continued process verification: ongoing capability monitoring that detects drift toward a space boundary before non-conformance occurs.

4.5.5 Worked example — ophthalmic

A new silicone hydrogel lens family is in development. The quality target product profile specifies oxygen permeability, modulus, water content, base-curve and diameter tolerances, surface wettability and a cosmetic acceptance standard. Risk assessment identifies modulus, water content and surface wettability as critical quality attributes with direct clinical significance.

A cause-and-effect matrix links eleven process parameters to these attributes. A resolution IV screening design across all eleven reduces the significant set to four: monomer ratio, cure temperature profile, hydration bath ionic strength, and surface-treatment plasma dose. A central composite design over these four, with incoming monomer lot and ambient humidity carried as noise factors in an outer array, maps the response surfaces and identifies a region in which modulus variance is minimised and is comparatively insensitive to lot variation.

The team defines the design space as a region rather than a point, validates across its corners and centre, and writes the control strategy so that cure temperature and plasma dose are equipment-controlled with automated verification, ionic strength is monitored in-process with a feedback loop, and monomer ratio is controlled at goods-inward by supplier specification and lot verification. Over the following three years, seven process adjustments that would each have constituted a change-control event under a point-setting validation are executed within the space as routine operational moves. That is the return on the additional development effort, and it is realised in regulatory throughput rather than in unit cost.

4.5.6 Limitations and failure modes

• Experimental cost. Full characterisation of a multi-parameter space is expensive in material and time. Sequential experimentation with early stopping, and the use of prior process knowledge to bound the region, are the standard mitigations.

• Scale and transfer. A design space established at development scale may not transfer to production scale. Scale-dependent parameters must be identified explicitly and the space re-characterised or bracketed at scale.

• Over-claiming the space. Defining a wide space on thin evidence is a regulatory liability. The space is only as wide as the demonstrated assurance.

• Organisational separation. Where development and manufacturing are separate organisations with separate objectives, design-space thinking is systematically under-invested because the cost falls on development and the benefit accrues to manufacturing.

4.5.7 Primary metrics

Predicted versus realised process capability at launch; count of critical quality attributes with a characterised design space; post-launch change-control events per product per year; development cycle time; proportion of critical parameters controlled by design or automation rather than by procedure; cost of quality in the first eighteen months post-launch.

4.6 Theory of Constraints: Drum-Buffer-Rope and Buffer Management

4.6.1 Theoretical basis

Goldratt’s Theory of Constraints rests on a single structural claim: in any system with a goal, throughput is determined by a small number of constraints, frequently one, and any improvement not made at the constraint yields no system improvement. From this follow the five focusing steps — identify the constraint, decide how to exploit it, subordinate everything else to that decision, elevate the constraint, and return to step one without allowing inertia to become the new constraint.

Drum-buffer-rope is the scheduling implementation. The drum is the constraint schedule, which sets the system rhythm. The buffer is time-based protection placed before the constraint (and before shipping, and before assembly points where constraint output converges with non-constraint output) sized to absorb upstream disruption. The rope is the material release mechanism that ties release into the system to constraint consumption, preventing the accumulation of work-in-process at non-constraints. Buffer management — monitoring buffer penetration as a real-time priority and diagnostic signal — is the operational control layer and, in practice, the most valuable single element.

4.6.2 Mechanism

Two mechanisms operate. The first is protective capacity allocation: by concentrating protection at the constraint rather than distributing it, the same service level is achieved with materially less total inventory. The second is diagnostic: buffer penetration statistics identify which upstream resources actually threaten throughput, converting improvement-project selection from an opinion-driven process into a data-driven one. Many organisations obtain more value from the second mechanism than from the scheduling change itself.

4.6.3 Evidence

Grade B, with a strong caveat. Mabin and Balderstone (2003), in the reference quantitative synthesis of Theory of Constraints applications published in the International Journal of Operations and Production Management, assembled results from a large body of documented implementations and reported substantial mean improvements in operational and financial performance, with lead-time and cycle-time reductions of the order of two-thirds and inventory reductions of the order of one-half among the reported cases.

They also report that, despite extensive searching, they found no published reports of failure. That statement should be read as a direct measurement of publication bias rather than as evidence of method infallibility, and the authors themselves flag the heterogeneity and poor standardisation of measurement across the reported cases. The defensible conclusion is that the effect direction is well established and the effect can be large, while the published magnitudes should be treated as an upper envelope. Subsequent work, including data-envelopment-analysis studies of drum-buffer-rope in engineer-to-order aerospace settings and action research in make-to-order environments, supports transferability beyond the repetitive-manufacturing settings of the original cases.

4.6.4 Implementation protocol

1. Identify the constraint empirically: the resource with the highest utilisation adjusted for load, or — more reliably — the resource in front of which queues persistently form. Do not identify it by capacity calculation alone; theoretical capacity and effective capacity diverge sharply where changeover and quality losses are present.

2. Exploit before elevating. Remove non-value time at the constraint: changeover, minor stoppages, inspection performed at the constraint that could be performed elsewhere, quality losses whose scrap consumes constraint time. Verify that only good material reaches the constraint — scrapping a part after the constraint destroys irreplaceable throughput.

3. Set the constraint schedule (the drum) and hold it. Schedule stability at the constraint is worth more than local optimisation elsewhere.

4. Size the constraint buffer from measured upstream variability, typically at an initial value of two to three times the mean upstream lead time, then tune from observed penetration data.

5. Implement the rope: release material into the system only at the rate the constraint consumes it, offset by the buffer time.

6. Operate buffer management with a red-yellow-green zone discipline. Use penetration frequency by cause as the improvement-project selection input.

7. Elevate only after exploitation is exhausted. Capital added to an unexploited constraint buys back capacity that was already available.

4.6.5 Worked example — ophthalmic

In a lens plant, sterilisation autoclaves are the constraint: they are capital-intensive, cycle-time-fixed, and shared across product families. Utilisation is 94 percent, and the plant is contemplating a fourth autoclave at significant capital cost and with an installation-qualification and performance-qualification burden measured in months.

Exploitation analysis, conducted before the capital decision, finds three recoverable losses. Load-density analysis shows a mean fill of 78 percent of validated capacity, driven by batch sequencing that mixes incompatible cycle recipes. Changeover between validated cycle recipes consumes 46 minutes per transition, with an average of 4.1 transitions per day. And 2.4 percent of autoclave throughput is consumed by material that is subsequently rejected at post-sterilisation inspection for defects created upstream — constraint time spent on product that was already scrap.

Countermeasures: campaign scheduling by cycle recipe to reduce transitions from 4.1 to 1.6 per day; load-pattern standardisation raising mean fill to 91 percent; and relocation of the cosmetic inspection gate from post-sterilisation to pre-sterilisation, so that defective product never consumes constraint capacity. The last intervention requires a change-control assessment, because inspection sequence is part of the validated control strategy — but it recovers capacity at zero capital cost. The combined recovered capacity exceeds the increment a fourth autoclave would have provided, and the capital decision is deferred.

Buffer management is then installed on the pre-sterilisation queue, with penetration events coded by cause. Over one quarter, the coded data identify hydration-line stoppages as the dominant threat to constraint feeding, redirecting the maintenance programme accordingly. This diagnostic output, obtained essentially for free, is typical of the method.

4.6.6 Limitations and failure modes

• Wandering constraints. In high-mix environments the constraint moves with product mix, and a single drum is then the wrong model. Mix-dependent constraint identification, or a simplified pull system, is required.

• Local-efficiency conflict. Subordination means non-constraints deliberately run below capacity. Where the measurement system rewards local utilisation, subordination is silently defeated. The measurement system must be changed first or the implementation will not survive contact with the monthly review.

• Publication-bias-inflated expectation. Setting expectations from published case magnitudes invites disappointment; set them from the plant’s own measured buffer and utilisation data.

• Constraint at a validated step. Where the constraint is a validated process such as sterilisation, exploitation options that alter validated parameters carry a change-control cost that must be included in the comparison against elevation.

4.6.7 Primary metrics

Throughput at the constraint (good units per constraint-hour); constraint utilisation decomposed into productive, changeover, idle-starved, idle-blocked and scrap-consumed; buffer penetration frequency by zone and by coded cause; due-date performance; inventory dollar-days; throughput-dollar-days.

4.7 Quick Response Manufacturing and POLCA

4.7.1 Theoretical basis

Quick Response Manufacturing, developed by Suri, is a company-wide strategy whose single objective function is lead-time reduction, on the argument that lead time is a proxy for a large class of system inefficiencies and that attacking it directly attacks them jointly. Its intellectual distinctiveness lies in taking queueing theory seriously as a design tool: it foregrounds the nonlinear relationship between utilisation and queue time, and it explicitly rejects the high-utilisation objective that conventional cost accounting imposes. The associated organisational form is the QRM cell — a cross-functional, co-located, multi-skilled team with end-to-end ownership of a product family segment — and the associated control mechanism is POLCA (paired-cell overlapping loops of cards with authorisation), a hybrid push-pull system designed for the high-variety case in which kanban fails because each part number would require its own card loop.

The lineage runs from time-based competition through to QRM; Godinho Filho and colleagues trace this evolution and situate QRM as the operational culmination of the time-based tradition. Suri’s further construct, the manufacturing critical-path time, provides the measurement definition: the typical calendar time from receipt of an order through to first delivery, assuming the order starts from nothing.

4.7.2 Mechanism

The core mechanism is deliberate capacity slack. Queueing theory gives waiting time as increasing without bound as utilisation approaches one; at 95 percent utilisation with realistic variability, queue time dominates process time by an order of magnitude. QRM therefore targets planned utilisation in the region of 70 to 85 percent on shared resources — a prescription that conventional efficiency accounting reads as waste, and which is the principal reason the method is under-adopted despite its analytical soundness.

POLCA supplies the second mechanism: it authorises material movement between cells only when the downstream cell has capacity, using card loops between cell pairs rather than per-part-number loops. This makes work-in-process control feasible in environments with thousands of routings, where kanban is unusable.

4.7.3 Evidence

Grade B for the queueing foundation, which is analytically established rather than empirical, and grade C for the implementation evidence, which is dominated by case studies. A frequently cited implementation study in Sustainability (2021) reports that a POLCA-integrated QRM framework applied in a precision manufacturing environment produced improved production scheduling with significant reductions in lead time and work-in-process. Exploratory transnational surveys of QRM knowledge and application across Brazil, Europe and the United States (Godinho Filho and colleagues) find awareness of QRM principles to be uneven and adoption to be shallow relative to lean, and both the authors of those surveys and later reviewers note that practical empirical studies of lead-time reduction — QRM in particular — remain comparatively scarce. The honest summary is that the theory is strong, the reported cases are favourable, and the controlled evidence base is thin.

4.7.4 Implementation protocol

1. Measure manufacturing critical-path time for the target product families using order-level time stamps, decomposed into process time, queue time, move time and wait-for-decision time. Expect queue and wait to dominate; if they do not, QRM is not the right method.

2. Form product-family focused cells with cross-trained teams and end-to-end ownership of a defined routing segment. Ownership, not layout, is the operative variable.

3. Set planned utilisation targets on shared resources in the 70–85 percent band and change the measurement system so that idle capacity at a non-constraint is not treated as a variance.

4. Implement POLCA loops between cell pairs, sizing card quantities from measured cell lead times and target work-in-process.

5. Attack the wait-for-decision component directly. In device environments this is dominated by quality disposition, deviation review and batch record review, and it is frequently the largest single term in manufacturing critical-path time.

6. Extend upstream and downstream: QRM is company-wide by design, and gains confined to fabrication are typically swamped by office-side lead time.

4.7.5 Worked example — ophthalmic

A custom and specialty lens operation — toric, multifocal and made-to-order prescriptions — has a manufacturing critical-path time of 19.4 days against a market expectation of 7. Decomposition shows 2.1 days of process time, 9.8 days of queue, 1.4 days of transport and staging, and 6.1 days of wait-for-decision, of which quality disposition and batch record review account for 5.2 days.

The finding reframes the problem entirely. A conventional lean programme would attack the 9.8 days of shop-floor queue and would leave the 6.1-day administrative tail untouched. The QRM diagnosis directs effort at both: cells are formed for the toric and multifocal families with dedicated surfacing, inspection and packaging, and POLCA loops are installed between surfacing, coating and inspection cells. Simultaneously, batch record review is restructured around review-by-exception with electronic batch records, and quality disposition authority for defined non-critical deviations is delegated to a qualified cell-level reviewer under a documented procedure — a change requiring quality-system revision and, in a device context, careful justification against the design history file, but one that addresses the dominant term.

The illustrative arithmetic: queue falls from 9.8 to 3.2 days through cell formation and utilisation reduction; wait-for-decision falls from 6.1 to 1.9 days through review-by-exception; total manufacturing critical-path time falls to approximately 8.6 days. The larger absolute gain comes from the administrative intervention, which is the characteristic QRM result and the one most often missed.

4.7.6 Limitations and failure modes

• Accounting conflict. Deliberate spare capacity conflicts with absorption costing and with utilisation-based performance measurement. Without a change to the measurement system the method is rejected by the controller before it is rejected by operations.

• Thin empirical base. Adoption should be framed as a designed change with a defined evaluation, not as an established best practice.

• Cell formation cost. Duplicated equipment across cells raises capital intensity; the trade against lead-time benefit must be computed, not assumed.

• Wrong environment. In genuinely high-volume, low-variety flow, conventional lean and drum-buffer-rope will outperform QRM at lower cost.

4.7.7 Primary metrics

Manufacturing critical-path time by product family, decomposed into process, queue, move and decision components; on-time delivery; work-in-process by cell; planned versus actual utilisation on shared resources; quote-to-delivery time for made-to-order products.

4.8 Total Productive Maintenance and Overall Equipment Effectiveness

4.8.1 Theoretical basis

Total Productive Maintenance is a plant-wide system for maximising equipment effectiveness through operator involvement, structured around eight pillars of which autonomous maintenance, planned maintenance, focused improvement and quality maintenance carry most of the operational load. Its measurement construct, overall equipment effectiveness, is the product of availability, performance rate and quality rate, and decomposes plant losses into six categories: breakdowns, setup and adjustment, minor stoppages and idling, reduced speed, start-up rejects and process defects.

The theoretical claim is that equipment deterioration is largely a function of neglected basic conditions — cleaning, lubrication, bolting, inspection — and that these are best owned by the operator, who has both the highest observation frequency and the strongest incentive. Maintenance specialists are then redeployed from reactive repair to condition-based and improvement work.

4.8.2 Mechanism

Two mechanisms. First, detection latency: an operator who cleans and inspects a machine daily detects incipient failure days or weeks earlier than a scheduled inspection regime, converting unplanned downtime into planned downtime. Second, variance reduction: the largest component of process-time variability in most plants is unplanned stoppage, so availability improvement tightens the entire flow-control problem addressed at layer L2. This is the availability coupling described in 3.2, and it is why Total Productive Maintenance is a prerequisite rather than a parallel initiative.

4.8.3 Evidence

Grade B. Total Productive Maintenance appears as one of the four practice bundles in Shah and Ward (2003), whose joint contribution to operational performance is established at grade A; its individual contribution is not separately isolated there. The dedicated literature is large but methodologically uneven, consisting predominantly of single-plant before-and-after studies. Representative reported results include an overall equipment effectiveness improvement of approximately 7.4 percent with concomitant productivity and quality gains in a documented plant study, and multi-case automotive evidence associating improved effectiveness and reduced production cost with substantial revenue and profit growth over a three-year horizon. Reported baseline effectiveness values in the surveyed literature range widely, from roughly 15 to 60 percent, which is itself informative: the distribution of starting points is very broad, and reported improvement magnitudes are strongly conditioned on where a plant starts.

More recent work has begun to formalise the integration of Total Productive Maintenance with Industry 4.0 instrumentation, proposing standardised frameworks structured around modular technological integration, phased deployment, organisational readiness and evaluation through overall equipment effectiveness together with mean time between failures and mean time to repair.

4.8.4 Implementation protocol

1. Establish honest overall equipment effectiveness measurement before improvement. Define the loss taxonomy, automate data capture where possible, and resist the near-universal temptation to define planned downtime expansively — an effectiveness figure that excludes inconvenient losses is a reporting instrument, not a control instrument.

2. Run initial cleaning and inspection on pilot equipment with operators and maintainers together. Record every abnormality found; the count is typically an order of magnitude higher than expected and is the primary evidence for the programme.

3. Eliminate sources of contamination and hard-to-access areas, so that the restored condition is maintainable in the available time.

4. Write provisional operator standards for cleaning, lubrication and inspection with time budgets that fit the actual shift pattern. Standards that require time the operator does not have will not be followed.

5. Train operators in general inspection so that they can detect deviation, not merely follow a checklist.

6. Build the planned maintenance system on failure data: criticality assessment, failure mode analysis, and an interval structure derived from observed failure distributions rather than from vendor defaults.

7. Deploy focused improvement teams against the largest quantified losses, and quality maintenance against equipment-induced defect modes.

8. In a validated environment, integrate the maintenance regime with calibration and preventive maintenance records so that the quality-system evidence and the reliability evidence are the same dataset.

4.8.5 Worked example — ophthalmic

A cast-moulding cell reports overall equipment effectiveness of 61 percent. Decomposition: availability 81 percent (unplanned stoppages 9 percent, changeover 10 percent), performance 88 percent (minor stoppages from mould-release sticking and lens-transfer misfeeds), quality 85 percent (cosmetic rejects concentrated in the first 40 minutes after each start-up).

The quality-rate loss is diagnostically the most interesting: rejects concentrated immediately after start-up indicate a thermal or conditioning transient, not a random defect process. Initial cleaning and inspection identifies 214 abnormalities across four presses, of which 38 are classified as potential quality-affecting. Contamination-source elimination on the demould station addresses the dominant minor-stoppage mode. Operator standards are written at 12 minutes per shift per press, which is the time actually available. Planned maintenance intervals for the hydraulic and thermal control subsystems are rebuilt from two years of failure records rather than from the vendor schedule, which had specified intervals materially longer than the observed characteristic life for two components.

For the start-up quality loss, the focused-improvement team establishes a warm-up qualification protocol: the press runs to a defined thermal steady state, verified by an instrumented criterion rather than by elapsed time, before production material is released. This is a control-strategy change and enters change control, but it converts a recurring 15 percent quality loss into a bounded, verified condition. Illustrative post-intervention effectiveness: availability 92 percent, performance 94 percent, quality 96 percent, giving 83 percent overall — and, more importantly for the flow layer, a process-time distribution tight enough to make the pull system of 4.3 viable.

4.8.6 Limitations and failure modes

• Effectiveness as a scoreboard. Where the figure is reported upward without the loss decomposition, the definition drifts until the number is comfortable. Always report the six losses alongside the composite.

• Operator time not granted. Autonomous maintenance requires protected minutes per shift. Programmes that assign the standards without granting the time fail within one quarter, and the failure is attributed to operator attitude.

• Cross-machine comparison. Overall equipment effectiveness is meaningful over time on one asset and misleading across dissimilar assets; a constraint machine at 70 percent is a different problem from a non-constraint at 70 percent.

• Maintenance procedure change control. On validated equipment, maintenance and calibration regimes are quality-system controlled. Interval changes require justification and, in some cases, revalidation.

4.8.7 Primary metrics

Overall equipment effectiveness with full six-loss decomposition; mean time between failures and mean time to repair by asset; ratio of planned to unplanned maintenance hours; abnormalities found and closed per period; maintenance cost per unit produced; start-up scrap as a proportion of total scrap.

4.9 Zero Defect Manufacturing

4.9.1 Theoretical basis

Zero Defect Manufacturing is best understood not as an aspiration but as a specific control architecture. Psarommatis, May, Dreyfus and Kiritsis (2020), reviewing 280 articles published between 1987 and 2018 in the International Journal of Production Research, formalise it around four strategies — detection, repair, prediction and prevention — applied along two orientations: product-oriented, which addresses defects in the physical part, and process-oriented, which evaluates the state of the manufacturing equipment and classifies it as normal or abnormal. The four strategies are complementary rather than alternative, and the architecture question is which combination applies to which defect mode.

The conceptual advance over conventional statistical quality control is the explicit treatment of prediction and repair as first-class strategies. Classical control charting is a detection strategy operating on process parameters; Zero Defect Manufacturing adds forward inference (predict the defect before it occurs, from process state) and recovery (repair or rework the part rather than scrap it), and requires an explicit decision policy over the four. The 2024 holistic review by the same research group updates the state of the field and identifies persistent gaps in economic evaluation, in integration with existing quality systems, and in scalability.

4.9.2 Mechanism

The mechanism is compression of the interval between defect creation and defect response, taken to its limit. Conventional quality control detects at end of line, with the quantity at risk equal to the production between inspections. Detection-strategy Zero Defect Manufacturing moves inspection in-process and to 100 percent coverage. Prediction-strategy Zero Defect Manufacturing moves the response before the defect, using process-state signals to trigger intervention. The economic value scales with the quantity at risk per event, which is why the approach is most valuable in high-speed, high-volume processes — exactly the ophthalmic case, where a moulding line producing tens of thousands of lenses per shift can accumulate a large defective population between inspection points.

4.9.3 Evidence

Grade B for the framework, which rests on two systematic reviews with substantial article corpora, and grade C for individual industrial applications, which are reported as case studies without control comparison. A separate systematic review of the role of Industry 4.0 in zero-defect manufacturing establishes the technology dependencies and proposes a conceptual framework for research direction. Readiness-level assessment frameworks have been proposed to help firms determine whether the prerequisites are in place — a useful corrective to the common pattern of committing to Zero Defect Manufacturing before the measurement and data infrastructure can support it.

The candid assessment is that Zero Defect Manufacturing is a well-specified architecture with a maturing literature and a thin base of independently evaluated industrial deployments. It should be adopted as an architecture for organising existing quality investments, with the prediction strategy treated as an experiment.

4.9.4 Implementation protocol

1. Build the defect taxonomy from production data: every defect mode, its rate, its detection point, its creation point, the detection lag, and the quantity at risk per event.

2. Assign a strategy to each mode. High-rate, cheaply detectable modes go to detection. Modes with a predictive process signature and a long detection lag go to prediction. Modes where the part retains value go to repair. Modes with an identifiable design or parameter cause go to prevention — which is always the preferred terminal state.

3. Establish measurement adequacy for each detection point. Conduct attribute agreement analysis on every automated inspection station against a truth panel; quantify the false-discovery rate and the escape rate separately.

4. Instrument the process for the prediction strategy: identify the sensor set that carries the defect signature, establish sampling rates adequate to the phenomenon, and build the labelled dataset by joining process traces to inspection outcomes at the part level. Part-level traceability is the hard prerequisite and is where most programmes stall.

5. Develop and qualify predictive models offline. Evaluate on a held-out period, not on a random split, because process drift makes random splits optimistic. Report precision and recall separately at the operating threshold, and set the threshold from the asymmetric cost of escape versus false alarm.

6. Deploy in advisory mode first — the model recommends, the human decides — and measure agreement over a defined qualification period before granting the model any automated disposition authority.

7. In a validated environment, treat any model with product-disposition authority as part of the control strategy: it requires software validation, change control on model updates, and a defined procedure for model performance monitoring and retraining triggers.

4.9.5 Worked example — ophthalmic

The defect taxonomy for a soft lens line identifies eleven modes. Three examples illustrate the strategy assignment.

Edge tear (rate ~180 parts per million, detected at automated optical inspection, created at demould, detection lag ~40 minutes): assigned to prediction. The demould station carries a load-cell trace; analysis of joined process-and-inspection data shows that edge-tear events are preceded by a characteristic rise in peak demould force over the preceding 20 to 60 cycles as mould-surface condition degrades. A model on the force trace achieves, in offline evaluation, useful recall at an operating threshold whose false-alarm rate costs less than the scrap it avoids. The intervention is a scheduled mould-surface treatment triggered by the model rather than by cycle count.

Surface inclusion (rate ~40 parts per million, detected at automated optical inspection, created at monomer dosing, detection lag ~3 hours): assigned to prevention. Root-cause analysis attributes the mode to particulate ingress at a filter change; the countermeasure is a design change to the filter housing and a revised procedure, not a detection improvement. This is the correct terminal strategy and it removes the mode from the taxonomy.

Cosmetic haze (rate ~600 parts per million, detected at automated optical inspection): assigned first to detection improvement, because the attribute agreement study shows that a substantial fraction of haze calls are measurement disagreement near the acceptance boundary. Improving the inspection algorithm’s discrimination reduces the apparent rate materially before any process change is attempted — and, critically, prevents the plant from spending a year chasing process causes for measurement variation. This ordering is the most transferable lesson in the example.

4.9.6 Limitations and failure modes

• Data prerequisite underestimated. Part-level joining of process traces to inspection outcomes requires traceability that most plants do not have and cannot retrofit cheaply. Assess this before committing.

• Model drift. Predictive models degrade as the process changes. Without automated performance monitoring and a defined retraining trigger, a model silently becomes a source of error.

• Validation burden on model-based disposition. Granting a model authority over product disposition brings it inside the validated control strategy, with software validation, change control and periodic review obligations. Advisory deployment avoids this and captures much of the benefit.

• Economic evaluation gap. The literature is weak on cost-benefit method, and vendors are weaker. Build the business case on measured quantity-at-risk and measured detection lag, not on defect-rate aspiration.

4.9.7 Primary metrics

Defects per million opportunities by mode; detection lag by mode; quantity at risk per detection event; inspection false-discovery rate and escape rate; predictive model precision and recall at the operating threshold; proportion of defect modes addressed by prevention rather than detection (the maturity indicator); scrap and rework cost per thousand units.