23-Ind-A5 Quality Planning, Control, and Assurance · May 2014
Nivaar worked solution (AI-drafted; not reviewed by a licensed engineer)
National Exams — May 2014 — 98-Ind-A5 Quality Planning, Control and Assurance. Three-hour, closed-book exam; Casio or Sharp approved calculators only; one double-sided 8.5×11 aid sheet permitted; relevant statistical tables attached. Format: six questions, each worth 20 marks; any five constitute a complete paper, and only the first five appearing in the answer book are marked, so candidates effectively choose 5 of 6. All six are solved below for completeness.
Reference texts: Montgomery, Introduction to Statistical Quality Control (8th ed.) — control charts, process capability, acceptance sampling and design of experiments for quality improvement (the primary text for every part of this paper); MIL-STD-105E — sampling procedures and tables for inspection by attributes; ISO 9001:2015 — quality management systems and certification.
Question text not reproduced: the examination questions are © Engineers and Geoscientists BC. Open the official past paper (linked at the top of this page) to read the question, then follow the worked solution below.
Quality Function Deployment (QFD) is a structured planning tool that translates the "voice of the customer" — qualitative statements about what customers want — into specific, prioritized engineering characteristics that design and manufacturing can act on. Its central artifact, the House of Quality, is a matrix that lists customer requirements down one side, engineering characteristics across the top, a relationship matrix in the body (how strongly each engineering characteristic affects each customer requirement), a "roof" correlation matrix showing how the engineering characteristics interact with each other (positively or negatively), and a competitive-benchmarking panel comparing the firm's product to competitors on each customer requirement. Working through the House of Quality forces the design team to prioritize engineering effort on the characteristics that matter most to the customer, exposes conflicting characteristics early (e.g., lighter weight vs. higher strength), and creates a documented, traceable link from customer language to design targets that can be deployed downstream into part characteristics and process plans (a cascade of QFD matrices).
The traditional (sequential/"over-the-wall") approach to product and process design passes the design through functional silos one at a time — marketing defines requirements, design engineering designs the product, process/manufacturing engineering then figures out how to make it, and quality is inspected in at the end. Each group works largely in isolation and hands its finished output to the next; problems discovered downstream (e.g., a feature that cannot be manufactured economically, or a tolerance the process cannot hold) are expensive to fix because they require looping back through already-completed stages. Concurrent engineering instead runs product design and process design in parallel, using cross-functional teams (design, manufacturing, quality, purchasing, sometimes suppliers and customers) from the earliest concept stage, supported by tools such as QFD, DFM/DFA (design for manufacture/assembly), and early prototyping. Manufacturability, quality, and cost are considered simultaneously with function, so problems are caught and resolved while they are still cheap to change (the well-known result that the cost of a design change grows by roughly an order of magnitude at each downstream stage — concept, design, process planning, production — applies here).
Concurrent engineering is preferable in essentially every modern setting: it shortens time-to-market (parallel rather than sequential activities), reduces total development cost (fewer late-stage engineering changes and re-tooling events), and produces a more manufacturable, higher-quality design because manufacturing and quality constraints are built in from the start rather than discovered after the design is frozen. The traditional sequential approach can only be preferred in the narrow case of a very simple, low-volume, one-off product where the coordination overhead of a cross-functional team is not justified by the modest savings available — for any product built repeatedly at volume, concurrent engineering's front-loaded coordination cost is repaid many times over across the production run.
Taguchi divides product/process design into three sequential stages. System design is the creative, largely engineering-science stage: choosing the basic technology, configuration, materials, and components that will accomplish the required function, drawing on engineering knowledge and prior experience — it establishes what the product IS. Parameter design (Taguchi's central contribution) chooses the nominal levels of the design (and, where relevant, process) parameters to make the product's performance as insensitive as possible to uncontrollable "noise" factors (manufacturing variation, environmental conditions, customer usage, component-to-component variation, and product aging) — robustness is achieved without eliminating the noise sources, purely by clever choice of the controllable parameter settings, which is far cheaper than tightening tolerances or shielding against noise. Tolerance design is the final, most expensive stage, entered only if parameter design alone cannot meet the performance target: it tightens tolerances on the most sensitive parameters (identified from parameter design) and may specify higher-grade materials or components, trading increased production cost for reduced variability.
In the parameter-design stage the objectives typically considered are: (i) minimizing the mean-squared deviation of the response from its target value (equivalently, maximizing the signal-to-noise, S/N, ratio) — the primary robustness objective; (ii) reducing the transmission of variation from noise factors into the response; and (iii), where an economically preferable "adjustment factor" exists (a controllable parameter that shifts the mean without materially affecting variability), using it to move the mean onto target after variability has already been minimized. Taguchi organizes the parameter-design experiment using two orthogonal arrays run simultaneously: the inner array lists the settings of the controllable design parameters to be tested (e.g., an $L_8$ or $L_9$ array of design factors), while the outer array lists settings of the uncontrollable noise factors, replicated under each inner-array run so that every design combination is tested across the full range of noise conditions the product will actually see. Linear graphs are diagrams (dots for factors, connecting lines for interactions) used to assign factors and their interactions to the columns of a standard orthogonal array without introducing confounding, i.e., a graphical column-assignment tool. The signal-to-noise ratio is a single performance statistic computed from each inner-array run's outer-array replicates (e.g., for nominal-the-best, $S/N = 10\log_{10}(\bar y^2/s^2)$; larger-the-better and smaller-the-better have their own forms) that combines mean and variability into one number to be maximized. Finally, graphs of marginal (row) averages plot the average S/N ratio (or average response) at each level of each controllable factor, letting the analyst read off, factor by factor, which level maximizes robustness — the basis for selecting the optimal parameter-design settings.
Acceptance sampling is a lot-disposition decision tool: a random sample is drawn from an already-produced lot, inspected against a sampling plan (sample size $n$, acceptance number $Ac$), and the whole lot is accepted or rejected based on the sample result. It says nothing about the underlying process that made the lot — it is purely an after-the-fact audit of a finished batch, useful when the producer has no control over (or no data on) the process, when destructive testing makes 100% inspection impossible, or when a supplier relationship still requires an incoming-inspection checkpoint. Capability testing (process capability analysis, $C_p$/$C_{pk}$) instead characterizes the process itself — it asks whether an in-control process's natural spread, relative to the specification limits, is inherently able to produce conforming output, using data collected while the process runs rather than after a lot is finished.
Capability testing is preferable in essentially every setting where the producer controls (or can gain access to) the process, because it addresses the root cause rather than sorting the symptom: a capable, well-centered process ($C_{pk}\gtrsim 1.33$) produces conforming output continuously, eliminating the need for lot-by-lot inspection altogether, whereas acceptance sampling only screens already-produced (possibly already-defective) product and, by design, still passes some nonconforming lots (the producer's risk) and rejects some good ones (the consumer's risk is the mirror image at the LQL). Acceptance sampling remains appropriate as a residual safeguard when capability cannot be verified or maintained (e.g., receiving inspection of a supplier's product with no visibility into their process) or where destructive/expensive testing rules out both 100% inspection and continuous process monitoring.
Taguchi's methods drew sustained criticism from the statistical community on several grounds: (i) the signal-to-noise ratio conflates mean and variance into a single statistic, which can be statistically inefficient and can mask the true source of a problem (a change that "improves" S/N might do so by shifting the mean rather than reducing variance, or vice versa — analyzing mean and $\log(s^2)$ separately is often more informative); (ii) his orthogonal-array experimental designs are frequently highly fractionated and heavily confound interactions with main effects, so a significant "main effect" may actually be a disguised interaction, producing misleading conclusions; (iii) his data-analysis methods (marginal averages, ANOVA on S/N ratios) are less rigorous and less efficient than standard factorial-design analysis of the same data; and (iv) some of his statistical loss-function and quality-definition claims were viewed by academic statisticians as insufficiently justified or overstated relative to classical decision theory. The engineering PHILOSOPHY — design robustness against noise as a deliberate, quantified design objective, and the economic loss function extending "quality" beyond the specification limits — is widely accepted and highly influential; it is specifically the statistical MACHINERY (the arrays, the S/N metric, the analysis method) that remains controversial.
EVOP (Evolutionary Operation) is a technique for improving an already-running full-scale production process without disrupting production or requiring a dedicated experimental facility. Small, deliberately conservative perturbations (typically a $2^2$ or $2^3$ factorial pattern) are made to a small number of controllable process variables around their current operating point, cycled repeatedly (a "phase" of several cycles) during normal production; because each individual perturbation is small, the process stays within its normal operating/specification range throughout, but by averaging results across many repeated cycles, EVOP accumulates enough statistical power to detect small, genuine improvements in the mean and variance of the response that would be invisible in any single run. It is used to hunt for a better long-run operating point on an already-in-control process, moving one small evolutionary step at a time (hence the name) rather than through a one-shot, disruptive experimental campaign, and is applied by plant personnel as a routine part of running the process (an "EVOP committee" typically evaluates results after each phase and decides whether to shift the operating centre or run another phase).
