Electronics Guide

Variation and the Logic of Control

Every manufacturing process varies. Two solder joints formed by the same machine, from the same paste, on the same board, differ measurably. SPC begins by classifying that variation into two categories with fundamentally different management responses.

Common Cause and Special Cause Variation

Common cause variation is the background variation inherent in a process operating as designed. It arises from many small, unidentified influences acting together: minor fluctuations in ambient temperature, small differences between component reels, ordinary mechanical play in a placement head. Common cause variation is stable and predictable in aggregate. A process subject only to common causes is said to be in statistical control, which means its future output can be predicted within known limits.

Special cause variation (also called assignable cause variation) comes from identifiable events that are not part of the normal process: a squeegee blade that has chipped, a solder paste jar past its working life, a new operator using a different technique, a nozzle that has partly clogged. Special causes are sporadic, often large, and traceable to a specific source. A process affected by special causes is out of control and unpredictable.

The management implication is the heart of the method. Special causes are removed by investigation at the point of occurrence, usually by the people running the process. Common cause variation cannot be removed that way; reducing it requires changing the process itself, which is a management and engineering responsibility. Deming called the two errors of confusing them tampering: treating common cause variation as though it had an assignable cause adds variation rather than removing it, a result his funnel experiment demonstrates vividly.

Control Limits Are Not Specification Limits

The single most common conceptual error in SPC is confusing control limits with specification limits. They come from different places and answer different questions.

• Specification limits come from the design. They state what the customer or the engineering requirement will accept, and they say nothing about what the process actually does.
• Control limits are computed from the process data itself, conventionally at three standard deviations either side of the process center line. They describe what the process actually does, and say nothing about whether that is acceptable.

A process can be in perfect statistical control and produce nothing but scrap, if its natural variation falls outside the specification. It can also be wildly out of control and still pass every part, if the specification is loose. Plotting specification limits on a control chart is therefore poor practice: it invites operators to adjust the process whenever a point drifts toward a specification limit, which is precisely the tampering the method exists to prevent. Capability analysis, described later in this article, is the correct tool for comparing process behavior against specifications.

The Standards Framework

SPC practice is codified in several standards families that engineers encounter in customer requirements and audits.

• ISO 7870 series, Control charts: Part 1 gives general guidelines (ISO 7870-1:2019), Part 2 covers Shewhart control charts (ISO 7870-2:2023), Part 3 acceptance control charts (ISO 7870-3:2020), Part 4 cumulative sum charts (ISO 7870-4:2021), Part 5 specialized control charts (ISO 7870-5:2014), Part 6 EWMA charts for the process mean (ISO 7870-6:2024), Part 7 multivariate control charts (ISO 7870-7:2020), Part 8 charting techniques for short runs and small mixed batches (ISO 7870-8:2017), and Part 9 control charts for stationary processes (ISO 7870-9:2020).
• ISO 11462 series, Guidelines for implementation of statistical process control: ISO 11462-1:2001 defines the elements of an SPC system and ISO 11462-2:2010 catalogs the supporting tools and techniques; both remain current while a consolidated revision works through approval. Two technical reports support software: ISO/TR 11462-4:2022 supplies reference data sets for validating measurement analysis software, and ISO/TR 11462-5:2023 defines a quality data exchange format for SPC software.
• ISO 22514 series, Statistical methods in process management — Capability and performance: Part 1 (2014) gives general principles, Part 2 (2026) addresses time-dependent process models, Part 3 (2020) covers machine performance studies, Part 4 (2016) defines capability estimates and performance measures, and Part 7 (2021) covers the capability of measurement processes.
• AIAG reference manuals: these are the working references wherever IATF 16949 applies. The Statistical Process Control (SPC) reference manual stood at its second edition (2005) for two decades and was superseded in July 2026 by a harmonized AIAG and VDA manual, which moves the emphasis away from SPC as a standalone tool and toward process understanding, risk from the FMEA and the special characteristics, and the link to advanced product quality planning. The Measurement Systems Analysis (MSA) reference manual remains at its fourth edition (2010) and is the next core tool due for the same harmonized treatment.
• IPC-9191, General Guidelines for Implementation of Statistical Process Control (SPC): published in 1999 and still the electronics industry's own SPC guidance, aligned with the ISO 11462 approach.
• ISO 3534-2:2006 supplies the vocabulary of applied statistics used across these documents.

IATF 16949:2016 makes SPC effectively mandatory for automotive suppliers by requiring manufacturing process studies with capability or performance indices for the characteristics identified in the control plan. Similar expectations appear in aerospace and medical device quality systems.

Preparing for Control: Characteristics, Subgroups, and Measurement

Charting begins long before the first point is plotted. Poor preparation produces charts that look reassuring and mean nothing.

Selecting Characteristics to Chart

Charting everything is neither affordable nor useful. Charts consume operator attention, and attention spread thin produces charts that nobody reads. The selection normally flows from risk analysis: a process FMEA identifies the failure modes with the highest severity and occurrence, the resulting control plan designates the key characteristics, and those characteristics get charted. Good candidates share several properties:

• The characteristic is measurable on a continuous scale, or at least countable, rather than judged subjectively.
• Variation in the characteristic has a demonstrated relationship to product performance or to downstream defects.
• The process can actually be adjusted in response, so that an out-of-control signal leads somewhere.
• Measurement is fast and cheap enough to sample at a useful frequency.

Charting a downstream pass/fail result is usually inferior to charting the upstream process parameter that drives it. Solder paste volume measured at the printer gives earlier and richer warning than a solder defect count measured after reflow, because it is a leading indicator with a continuous scale.

Rational Subgrouping

Shewhart charts work by comparing variation within subgroups against variation between subgroups. Within-subgroup variation estimates the common cause noise and sets the width of the control limits; between-subgroup variation is what the chart tests for signals. The subgroup must therefore be formed so that only common causes operate inside it. This principle is called rational subgrouping, and it is the design decision that most often determines whether a chart works.

For a stencil printer, five consecutive boards printed within a few minutes form a rational subgroup: any special cause affecting the squeegee or the paste will affect the whole subgroup and show up as a shift between subgroups. Combining one board from each of four printers into a single subgroup does the opposite; the printer-to-printer differences inflate the within-subgroup estimate, the control limits widen, and the chart becomes insensitive to everything. Each printer needs its own chart, or the printer must be treated as a stratification factor.

Sampling frequency follows the process. Fast processes with quick failure modes need frequent small subgroups; slow, stable processes tolerate longer intervals. A useful rule is to sample often enough that a process change is detected before more product is produced than the organization is willing to quarantine and reinspect.

Measurement System Analysis

Every measurement contains process variation and measurement variation combined. Because variances add, the observed variance is the sum of the true part-to-part variance and the measurement system variance. If the measurement system is noisy, control limits widen, capability indices fall, real shifts hide, and false signals multiply. Measurement system analysis (MSA) quantifies this contribution before the chart is trusted.

The principal MSA study is the gauge repeatability and reproducibility (gauge R&R) study, usually run as a crossed design in which several operators each measure the same set of parts several times.

• Repeatability is the variation observed when one operator measures the same part repeatedly with the same instrument. It reflects the equipment's inherent precision.
• Reproducibility is the additional variation introduced by different operators, and in some designs by different instruments or laboratories.
• Percent study variation (%GRR) expresses the combined measurement variation as a percentage of total study variation. Following the AIAG MSA guidance, a system below 10 percent is generally acceptable, between 10 and 30 percent may be acceptable depending on the application, the criticality of the characteristic, and the cost of improvement, and above 30 percent is unacceptable.
• Number of distinct categories (ndc) estimates how many distinct groups the measurement system can resolve within the observed part variation, calculated as 1.41 times the ratio of part variation to gauge R&R variation. A value of five or more is expected.
• Bias, linearity, and stability studies address systematic offset from a reference value, how that offset changes across the measurement range, and drift over time.

Attribute inspection needs its own treatment. An attribute agreement analysis, in which several inspectors classify the same set of known-good and known-bad samples more than once, measures agreement within inspectors, between inspectors, and against the reference standard. Visual solder joint inspection frequently fails such studies, which is one reason automated optical inspection displaced it for high-volume work. Related instrument-level requirements are covered in Calibration and Metrology Systems.

Data Collection Practices

Charts are only as good as the record behind them. Each measurement should carry the time, the machine, the line, the operator or shift, the lot or reel identifier, and the fixture or nest position. That context turns an out-of-control signal into a diagnosis: a shift that tracks one nozzle, one nest, or one reel points immediately at its cause. Manual transcription introduces errors and delay, so direct capture from the inspection equipment is strongly preferred, an integration discussed in Manufacturing Execution Systems (MES).

Shewhart Variables Control Charts

Variables charts plot measurements on a continuous scale. They carry far more information per observation than attribute charts and are the first choice whenever a continuous measurement is available. Variables charts always come in pairs: one chart for location and one for spread. Interpreting the location chart before the spread chart is in control is meaningless, because an unstable spread invalidates the control limits on the location chart.

X-bar and R Charts

The average and range pair is the classic combination for subgroups of roughly two to nine observations. The R chart plots the range of each subgroup; the X-bar chart plots the subgroup average. Control limits derive from tabulated constants that depend only on the subgroup size:

• X-bar chart limits: the grand average plus and minus A₂ times the average range.
• R chart limits: D₄ times the average range for the upper limit and D₃ times the average range for the lower limit.
• The process standard deviation is estimated as the average range divided by d₂.

The R chart has no meaningful lower limit for subgroups of six or fewer, which is why D₃ is zero there. Subgroups of four or five are the traditional compromise: large enough for the subgroup average to be approximately normal even when individual measurements are not, small enough to keep sampling cost reasonable.

X-bar and S Charts

For subgroups larger than about nine, the range becomes an inefficient estimator of spread because it uses only two of the observations. The standard deviation chart replaces it, with limits set by the constants B₃ and B₄ and the process standard deviation estimated as the average subgroup standard deviation divided by c₄. Automated measurement has made large subgroups cheap, so X-bar and S charts are increasingly common where inspection equipment measures every unit.

Individuals and Moving Range Charts

Many electronics processes produce one measurement at a time: a batch chemistry titration, a daily oven profile, a single reading per production run. The individuals and moving range chart (I-MR, also written X-MR) handles these cases. The moving range is the absolute difference between consecutive observations, and the process standard deviation is estimated as the average moving range divided by d₂ for a subgroup size of two, which is 1.128. The familiar individuals chart limits at the process average plus and minus 2.66 times the average moving range follow directly from three divided by 1.128.

Individuals charts carry two cautions. They are markedly more sensitive to non-normality than X-bar charts, because no averaging is available to invoke the central limit theorem, so the distribution should be checked and transformed if it is strongly skewed. They are also slow: detecting a one-sigma shift with an individuals chart takes far longer than with a subgrouped chart, which makes the supplementary run rules described below more valuable here than anywhere else.

Establishing and Revising Control Limits

Limits are computed from a baseline study, conventionally at least twenty to twenty-five subgroups. If the baseline contains points that are out of control and the causes are found and eliminated, those points are removed and the limits recalculated. If the causes cannot be found, the points stay, because removing data merely to tighten limits produces a chart that signals constantly.

Once established, limits are held fixed and the process is plotted against them. Recalculating limits every time new data arrive lets a slow drift walk the limits along with it, and the chart never signals. Limits should be revised only when the process has been deliberately changed, when a genuine improvement has been demonstrated and its cause is understood, or on a defined periodic review when a formal reassessment is scheduled. Every revision belongs in the change control record, along with the reason.

Attribute Control Charts

Attribute charts count events rather than measuring magnitudes. They apply where the outcome is inherently a classification, such as a board that passes or fails in-circuit test, or where measurement is impractical.

• p chart: monitors the proportion of nonconforming units. It accommodates varying sample sizes, at the cost of control limits that change width from point to point.
• np chart: monitors the number of nonconforming units, and requires a constant sample size.
• c chart: monitors the count of nonconformities per inspection unit when the inspection unit is constant, such as defects per assembled board of a single type.
• u chart: monitors nonconformities per unit when the inspection unit varies, such as defects per thousand solder joints across a mixed-model line.

The p and np charts assume a binomial model; the c and u charts assume a Poisson model. Both assumptions fail when defects cluster, which is common in electronics: one misregistered stencil produces a burst of correlated defects, not independent ones. Overdispersed count data produce control limits that are far too tight and a chart that signals almost continuously. Laney's p' and u' charts correct for overdispersion and are the practical remedy when subgroup sizes are large.

Attribute charts have a structural weakness: they need large samples. The arithmetic is unforgiving. At a nonconforming rate of 500 parts per million, the common working rule that a subgroup should contain at least five nonconforming units on average puts the subgroup at ten thousand pieces, and a lower control limit above zero requires roughly eighteen thousand. When defect rates fall into the parts-per-million range, counting charts become nearly useless and the response is either to chart the underlying continuous process parameter instead, or to chart the time or the number of units between defects using a g chart or t chart. For assembly-level reporting, the electronics industry standardizes the defect metrics themselves: IPC-7912A defines end-item defects per million opportunities for printed circuit board assemblies, and IPC-9261A defines the in-process equivalent, so that opportunity counts are computed consistently between a customer and a contract manufacturer.

Interpreting Control Charts

A control chart signals when the plotted data show a pattern that a stable process would rarely produce. The rules that define "rarely" are standardized.

Western Electric and Nelson Rules

The Western Electric Statistical Quality Control Handbook (1956) divided each half of the chart into three zones, each one standard deviation of the plotted statistic wide: zone A nearest the limits, zone B in the middle, zone C nearest the center line. Its four basic rules are one point beyond zone A, two of three consecutive points in zone A or beyond on the same side, four of five consecutive points in zone B or beyond on the same side, and eight consecutive points on the same side of the center line.

Lloyd S. Nelson extended and rebalanced this set in "The Shewhart Control Chart—Tests for Special Causes," published in the Journal of Quality Technology in 1984. The eight Nelson rules were chosen so that each test carries a broadly comparable chance of a false signal, a few in a thousand per test. Nelson keeps the first three Western Electric patterns as written, states the run rule as nine points on the same side of the center line rather than eight, and adds four more: six points in a row steadily increasing or decreasing, fourteen points in a row alternating up and down, fifteen points in a row within zone C on either side, and eight points in a row on both sides of the center line with none in zone C.

The last two are frequently misunderstood because they signal on data that look unusually well behaved. Fifteen consecutive points hugging the center line usually means the control limits were computed from inflated variation, that the measurement has insufficient resolution, or that the data have been rounded or filtered. Eight points avoiding the center line entirely suggests two mixed process streams, such as two nozzles or two cavities charted together, and is a rational subgrouping failure.

Different rule sets state slightly different point counts for the same pattern; a run rule may be stated as seven, eight, or nine points depending on the source. The specific set in use should be documented in the control plan so that operators, auditors, and software agree.

False Alarms and Detection Speed

Average run length (ARL) quantifies chart performance as the average number of points plotted before a signal occurs. With three-sigma limits and the single rule of one point outside the limits, the probability of a false signal on any point is about 0.0027 under normality, giving an in-control ARL near 370: roughly one false alarm every 370 subgroups. That same chart is slow against small shifts, and how slow depends heavily on the subgroup size. A shift of one process standard deviation moves a subgroup average of four by two standard errors and is caught in about six subgroups on average; the same shift on an individuals chart, where no averaging is available, takes about forty-four points.

Adding supplementary run rules trades one for the other. Champ and Woodall showed in Technometrics in 1987 that applying the four standard Western Electric rules together reduces the in-control ARL to approximately 92, meaning roughly four times as many false alarms, in exchange for substantially faster detection of small shifts. Applying every available rule to every chart is therefore not a conservative choice: it manufactures false alarms, and false alarms train operators to ignore the chart. Rule selection should be deliberate and matched to the cost of investigation.

Reaction Plans

A signal is worthless without a defined response. Every charted characteristic needs a documented reaction plan, sometimes called an out-of-control action plan, that specifies who is notified, what the operator does immediately, how far back the affected material extends, and how the containment and disposition are recorded. The plan should distinguish between stopping the line, quarantining product back to the last known-good point, and simply recording an observation and continuing. Without it, the practical response to a signal becomes an operator initialing the chart and moving on.

Charts for Small Shifts and Difficult Processes

The Shewhart control limits test only the most recent point, and the run rules of the previous section give the chart at best a short memory. This makes it excellent at detecting large, sudden shifts and poor at detecting small, sustained ones. Several chart families address that gap.

Cumulative Sum Charts

A CUSUM chart accumulates the deviations of each observation from a target, so a small but persistent offset builds into a visible slope even though no individual point is unusual. The tabular form maintains two one-sided sums with a reference value k, conventionally half the shift to be detected expressed in standard deviations, and signals when either sum exceeds a decision interval h. The pairing of k equal to 0.5 with h equal to 5 is a standard design for detecting a one-sigma shift and yields an in-control ARL of about 465, while h equal to 4 gives an in-control ARL of about 168 with faster detection. CUSUM charts are covered by ISO 7870-4:2021 and suit slow, expensive processes such as plating baths and chemical process tanks where each observation is costly and small drifts matter.

Exponentially Weighted Moving Average Charts

An EWMA chart plots a weighted average in which the weight assigned to past observations decays geometrically, controlled by a smoothing parameter lambda between zero and one. Small lambda values, typically 0.05 to 0.25, give the statistic a long memory and high sensitivity to small shifts; lambda approaching one reduces the chart to a Shewhart individuals chart. EWMA charts perform comparably to CUSUM charts for small shifts, are easier to explain and plot, and are robust to non-normality, which accounts for their wide use; CUSUM keeps an edge where the size of the shift worth detecting is known in advance and the design can be tuned to it. ISO 7870-6:2024 covers EWMA charts for the process mean.

The EWMA statistic also has a second life as a controller rather than a monitor. Run-to-run control in semiconductor fabrication uses an EWMA of recent process results to update the recipe for the next lot, compensating for tool drift in operations such as chemical mechanical planarization and lithography overlay. Related process control practice appears in Semiconductor Wafer Fabrication.

Short-Run and Mixed-Model Charts

High-mix, low-volume electronics assembly rarely produces twenty-five subgroups of anything before changing over. ISO 7870-8:2017 addresses this case directly. The standard approach transforms each measurement into a dimensionless deviation from its own part-specific target, divided by an estimate of the common process standard deviation, so that measurements from different part numbers can be plotted on one chart. These are variously called deviation from nominal charts, standardized charts, Z charts, or DNOM charts. They work when the different products genuinely share a process variance; when they do not, the variance must be estimated separately per product before standardizing. Short-run methods are essential in the environment described in Prototype and Low-Volume Production.

Multivariate Control

Modern equipment reports dozens of correlated variables at once. Charting each separately inflates the false alarm rate: with twenty independent three-sigma charts, the chance that at least one signals falsely on a given sampling occasion is about 5 percent rather than the 0.27 percent of a single chart. Separate charts also miss failures that show only as a change in the relationship between variables. Hotelling's T-squared statistic condenses a correlated set into one distance measure from the multivariate center and charts that, and ISO 7870-7:2020 covers the construction of such charts. Its weakness is diagnostic: a T-squared signal says something changed without saying what, so it is normally paired with contribution plots that decompose the statistic by variable. For very wide data sets, such as full reflow oven and placement machine telemetry, principal component analysis reduces the dimensionality first, and the resulting scores are charted along with a residual statistic. The multivariate EWMA extends the same idea to small sustained shifts.

Autocorrelated and Time-Dependent Processes

Shewhart charts assume that consecutive observations are independent. Continuous and semicontinuous processes violate this badly: consecutive temperature readings in a reflow oven or consecutive thickness measurements in a deposition process are strongly correlated in time. Applying an ordinary chart to autocorrelated data produces control limits that are far too narrow and a chart that signals constantly. The standard remedy is to fit a time series model, then chart the residuals, which are approximately independent if the model is adequate. Sampling less frequently, so that consecutive observations are far enough apart to be effectively independent, is a simpler alternative when the process allows it. ISO 7870-9:2020 covers control charts for stationary processes, and ISO 22514-2:2026 addresses capability assessment for processes whose distribution changes over time, which is the same underlying problem viewed from the capability side.

Process Capability and Performance

Control charts answer whether a process is stable. Capability analysis answers a different question: whether a stable process is good enough for the specification. Both answers are needed, and they must be obtained in that order, because a capability index computed on an unstable process describes a distribution that does not exist.

Capability and Performance Indices

• C_p compares the specification width to six times the within-subgroup standard deviation. It measures potential capability if the process were perfectly centered and ignores centering entirely.
• C_pk takes the smaller of the distances from the process mean to each specification limit, divided by three within-subgroup standard deviations. It penalizes off-center processes, and it can never exceed C_p.
• P_p and P_pk use the overall standard deviation computed from all the data rather than the within-subgroup estimate. They therefore include variation between subgroups, such as shift-to-shift and lot-to-lot effects, and describe long-term performance.
• C_pm, the Taguchi index, measures variation about the target rather than about the process mean, penalizing a process that sits off target even when it remains within specification.

The gap between C_pk and P_pk is diagnostic. When they agree closely, the process is stable and the sources of variation are consistent over time. When C_pk is much larger than P_pk, the process is well behaved within subgroups but drifts between them, which points to setup differences, shift effects, material lots, or maintenance cycles rather than to inherent machine variation. ISO 22514-4:2016 defines these estimates and their computation formally.

Capability and Expected Defect Rates

Assuming a normal distribution and a centered process, capability indices translate directly into expected fraction nonconforming.

The familiar six sigma target of 3.4 defects per million opportunities is not the 0.002 ppm in the table. It assumes the process mean has drifted 1.5 standard deviations off center, leaving 4.5 sigma to the nearer specification limit, and counts only that tail. The 1.5 sigma shift is a conventional allowance for long-term drift rather than a measured property of any particular process, and it should be treated as such.

Requirements vary by industry and by characteristic, and the automotive convention distinguishes the initial study from ongoing production. The production part approval process judges an initial study of a stable process against a P_pk of 1.67; a result between 1.33 and 1.67 is submitted to the customer for review rather than accepted automatically, and anything below 1.33 does not meet the requirement. Ongoing production is then commonly held to a C_pk of at least 1.33, with tighter targets set by the customer for characteristics designated as special or safety related. IATF 16949 requires the control plan to carry a reaction plan for a characteristic that is not statistically capable or stable, which in practice means containment and 100 percent inspection until the process is corrected. Aerospace and implantable medical device work frequently demands more.

Non-Normal Data and Bounded Characteristics

The normal assumption behind the table above fails routinely in electronics. Many characteristics are physically bounded at zero and skewed: particle counts, flatness, warpage, leakage current, surface roughness, and pull strength among them. Applying the normal formulas to such data misstates the defect rate badly, usually in the optimistic direction.

Three approaches are standard. A distribution can be fitted directly, using a Weibull, lognormal, or gamma model, and the capability index defined from its percentiles: ISO 22514-4 uses the interval between the 0.135 percent and 99.865 percent quantiles in place of six standard deviations, which reduces to the familiar formula for a normal distribution. Alternatively, the data can be transformed, commonly with a Box-Cox or Johnson transformation, and the normal formulas applied in the transformed space. Third, the apparent non-normality can be investigated: a bimodal distribution usually means two mixed process streams rather than an exotic distribution, and separating them is more informative than fitting a curve to their sum.

Conducting a Capability Study

A defensible study follows a fixed sequence. Confirm the measurement system through MSA first, since measurement variation inflates the denominator and depresses the index. Establish statistical control with a control chart and resolve any special causes. Collect enough data, conventionally at least twenty-five subgroups or one hundred individual measurements, spanning the sources of variation the index is meant to represent: a study run within one hour on one shift with one material lot estimates short-term capability only. Verify the distributional assumption. Then compute the indices, and report them with the sample size, the time span, the subgroup structure, and the distribution used. An index quoted without that context cannot be audited or reproduced.

Applying SPC in Electronics Manufacturing

The general method takes specific forms on an electronics line, where the most valuable charts sit at the process steps that generate the most defects.

Solder Paste Printing

Stencil printing is widely identified as the largest single contributor to surface mount defects, and it is also the most thoroughly instrumented, which makes it the natural home of SPC on an assembly line. Solder paste inspection systems measure the volume, height, area, and registration of every deposit, generating enormous quantities of variables data. Volume is normally expressed as a percentage of the theoretical stencil aperture volume, and transfer efficiency defined that way is itself worth charting. The pass or fail window programmed into the inspection system is a specification limit set for the assembly, so control limits must still be computed from the printer's own output; taking the inspection window as the chart limits repeats exactly the confusion described earlier in this article. Aperture geometry constrains what is achievable: IPC-7525C recommends an area ratio, the aperture opening area divided by the area of the aperture walls, greater than 0.66 for reliable paste release, and deposits below that ratio print erratically no matter how well the process is controlled. Charting volume by stencil zone or by squeegee stroke direction, rather than pooling the whole board, isolates squeegee wear, stencil tension loss, and support tooling problems.

Component Placement

Placement machines report offset in X, Y, and rotation for each component. These are naturally multivariate and naturally stratified: the useful chart separates by placement head, nozzle, and feeder slot, because a drift confined to one nozzle is a maintenance signal while a drift affecting the whole machine is a calibration or fiducial recognition signal. Pooling all placements into one chart conceals exactly the information the chart exists to reveal. Related equipment considerations appear in Automated Assembly Equipment.

Reflow and Wave Soldering

Reflow profiles are characterized by parameters such as peak temperature, time above liquidus, soak duration, and ramp rate, each of which can be charted from periodic profiling runs on an instrumented board. For SAC305 (Sn96.5/Ag3.0/Cu0.5), the most common lead-free assembly alloy, the solidus sits at about 217 degrees Celsius and the liquidus a few degrees above it, near 220, so time above liquidus must be referred to the higher figure. The process window is bounded below by wetting requirements and above by the peak package body temperatures set out in the J-STD-020 moisture and reflow sensitivity classification. Because profiling runs are infrequent and produce one value per parameter, individuals and moving range charts are the usual choice. Oven zone temperatures and conveyor speed are logged continuously and are strongly autocorrelated, so they call for the residual or reduced-frequency methods described earlier. Wave and selective soldering add solder pot temperature, contact time, conveyor angle, and flux deposition volume to the list. See Soldering Technologies and Materials for the underlying process physics.

Test and Electrical Parameters

Functional and in-circuit test results are usually treated as pass or fail, which discards most of their information. The measured values behind those verdicts, such as supply current, oscillator frequency, gain, or offset voltage, are continuous variables that drift long before they cross a test limit. Charting them turns end-of-line test into an early warning system for component lot changes and process drift. The practice is standard in semiconductor test, where parametric data drive both yield analysis and wafer-level process feedback.

Semiconductor and Substrate Fabrication

Wafer fabrication was the earliest heavy industrial user of advanced SPC and remains the most sophisticated. Film thickness, critical dimension, overlay registration, sheet resistance, and particle counts are charted at nearly every step, with within-wafer, wafer-to-wafer, and lot-to-lot variation separated deliberately so that the chart identifies which level of the process changed. Multivariate methods and run-to-run control are routine rather than exotic here, and the equipment generates enough telemetry that automated fault detection and classification operates alongside the conventional charts.

Implementation and Deployment

The statistics are the easy part. Most SPC programs fail organizationally rather than mathematically.

Sequence of Deployment

A workable rollout starts narrow. Select a small number of characteristics with demonstrated impact, validate their measurement systems, run a baseline study, establish limits, write reaction plans, train the operators who will use the charts, and only then expand. Programs that begin by charting hundreds of characteristics across a plant generate a wall of paper that nobody reads and discredit the method for years afterward.

Automated Data Collection and Software

Manual charting survives only in low-volume and regulated niches. Modern implementations pull measurements directly from solder paste inspection, automated optical inspection, X-ray, and test systems into an SPC layer within the manufacturing execution system, which computes limits, applies the selected rules, and pushes alerts to line displays and to engineering dashboards. Automation removes transcription error and delay, but it introduces its own risks: alert routing that reaches nobody, rules applied to autocorrelated data, and limits that recalculate automatically as new data arrive, quietly following a drift. Configuration should be reviewed as deliberately as any other process parameter. The broader analytics context is covered in Manufacturing Data Analytics.

Roles and Training

Operators need to read a chart, apply the rules in use, and execute the reaction plan; they do not need to derive control limits. Technicians and engineers own the limit calculations, the rule selection, the capability studies, and the investigation of signals. Management owns the response to common cause variation, because reducing it requires resources the shop floor does not control. Where an organization pushes responsibility for common cause variation down to operators, the predictable outcome is tampering. Training approaches are discussed further in Workforce Training and Development.

Common Failure Modes

• Charting without reacting. Charts posted and initialed but never acted upon are a compliance ritual, not process control. Auditing the reaction records is a better health check than auditing the charts.
• Specification limits drawn on control charts. This invites adjustment against noise and defeats the method.
• Limits recalculated continuously. Drift becomes invisible when the limits drift with it.
• Skipping measurement system analysis. A gauge contributing 40 percent of observed variation makes every chart and every capability index unreliable.
• Irrational subgrouping. Mixing machines, cavities, or nozzles within a subgroup widens limits until the chart cannot signal.
• Capability computed on an unstable process. The index describes a distribution that does not exist, and it will not be reproducible.
• Applying every run rule everywhere. The resulting false alarm rate teaches operators to ignore signals.
• Charting outcomes instead of drivers. Downstream pass rates signal too late to prevent the defects already built.

Relationship to Other Quality Methods

SPC does not stand alone. Within the Six Sigma DMAIC framework it appears twice: in the measure phase to establish baseline stability and capability, and in the control phase to hold the gains after an improvement. Design of experiments identifies which factors drive variation; SPC then monitors the settings that the experiment established. Failure mode and effects analysis identifies what to chart, and the control plan records it. Acceptance sampling under standards such as ANSI/ASQ Z1.4 and Z1.9 inspects lots after the fact, whereas SPC controls the process that produces them; the two coexist, but a mature process control program reduces reliance on sampling inspection rather than replacing one form of inspection with another.

Machine learning methods for anomaly detection increasingly run alongside conventional charts, particularly on high-dimensional equipment telemetry where classical multivariate methods struggle. They complement rather than replace control charts: a control chart is transparent, auditable, and defensible to a customer or a regulator, and its false alarm behavior is calculable, which matters when the response to a signal is stopping a production line.

Summary

Statistical process control rests on a single idea with wide consequences: variation splits into common causes inherent in the process and special causes that can be identified and removed, and the two demand opposite responses. Control charts operationalize that distinction, using limits computed from the process itself rather than from the specification.

Effective practice requires more than plotting points. Characteristics must be chosen for their influence on the product, subgroups must be formed so that only common causes act within them, and measurement systems must be validated before their output is trusted. Shewhart charts detect large shifts quickly; CUSUM and EWMA charts detect small sustained shifts that Shewhart charts miss; short-run, multivariate, and time-series methods extend the approach to high-mix production, correlated variables, and autocorrelated data. Capability and performance indices then translate a stable process into an expected defect rate, provided stability is established first and the distributional assumptions are checked.

In electronics manufacturing the highest-value applications sit where instrumentation is already dense and defects are already concentrated: solder paste printing, component placement, reflow profiling, and parametric test. The standards framework, from ISO 7870 and ISO 22514 through the AIAG reference manuals and IPC-9191, provides the common language in which customers, suppliers, and auditors discuss it. The programs that succeed are the ones where every chart has an owner, a rule set, and a reaction plan, and where management accepts that reducing common cause variation is its own responsibility rather than the operator's.

Related Topics

The following articles cover closely connected areas of manufacturing quality and process management: