Quality Control Problems in Practice

• Quality control problems are tasks that determine if manufacturing products, software outputs, or datasets are reliable for safe use.
• They integrate methodologies like statistical process control and machine learning to overcome measurement error and heterogeneous data challenges.
• Applications span from predictive inspections in manufacturing to real-time monitoring in software systems and data services.

Searching arXiv for the cited papers to ground the article in current records. Quality control problems are the tasks of determining whether products, processes, datasets, or algorithmic outputs are sufficiently reliable for downstream use. In manufacturing, quality control is an essential operation ensuring that products meet the necessary standards of quality, safety, and reliability, yet traditional methods such as visual inspections, measurements, and statistical techniques are often time-consuming, costly, and reactive (Rajasekaran et al., 12 Mar 2026). In software systems, continuous quality assessments are used to identify quality problems early, when their removal is still inexpensive (Deissenboeck et al., 2016). In a formal algorithmic sense, “Quality Control Problems” are specified by a positive, real-valued quality function $\rho$ and a distribution $D$ such that, with high probability, a sample drawn from $D$ is high quality, meaning its $\rho$-value is near $1$; the goal is to accept inputs $x \sim D$ and reject potentially adversarially generated inputs $x$ with $\rho(x)$ far from $1$ (Marcussen et al., 22 Aug 2025). Across manufacturing, software, medical imaging, crowdsourcing, scientific data services, multimedia delivery, water networks, and random graphs, the recurrent issues are reactivity, heterogeneity, measurement error, limited labels, large data volumes, and the need to aggregate evidence into operational decisions.

1. Formal scope and conceptual structure

A central distinction in the recent literature is between quality control as a practical monitoring activity and quality control as a formal computational objective. In continuous software quality control, the process is described as a cyclical loop with three key elements: definition of quality goals, assessment using tools and techniques, and implementation of improvements. Quality goals may be expressed as quality models, standards, KPIs, or software metrics, and the loop is repeated throughout the lifecycle to enable early detection and inexpensive resolution of quality problems (Deissenboeck et al., 2016).

The formal theory introduced for “Quality Control Problems” is deliberately weaker than both testing whether $\rho(x) \approx 1$ and testing whether $D$0. A quality control algorithm must satisfy completeness—accepting most inputs drawn from $D$1—and soundness—rejecting any input with $D$2—but it does not need to accept every instance with $D$3 near $D$4 (Marcussen et al., 22 Aug 2025). This weaker objective is operationally significant because it permits algorithms that certify “typical enough for safe use” without solving a full recognition or estimation problem.

The random-graph case study makes the distinction concrete. For $D$5 and

$D$6

testing whether $D$7 from one sample is impossible, and testing whether $D$8 requires $D$9 samples. By contrast, the quality control problem with respect to $D$0 can be tested with $D$1 queries and time when $D$2, and more generally, for a motif $D$3 of maximum degree $D$4, the corresponding problem can be solved with $D$5 queries and running time (Marcussen et al., 22 Aug 2025). This establishes quality control as a distinct layer between worst-case verification and average-case trust.

2. Information structure, observability, and provenance

Quality control problems are defined as much by their observables as by their decision rules. In the foundry case study on castings for powertrain components, the analysis integrated four primary datasets: Manufacturing Execution System product data on approximately 572,000 products with 9 attributes; quality deviations and defects data with 26 attributes; process data from core-making machines A and B with 11 attributes; and maintenance logs covering breakdown frequency, failure types, timestamps, and downtime. The preparation pipeline included data translation, merging MES with defects and process logs, linking to core-level components, pivoting machine operations into features, merging in maintenance effects, and preprocessing through imputation, categorical encoding, normalization, and outlier removal (Rajasekaran et al., 12 Mar 2026). The resulting picture is that a quality control problem is often an integration problem before it is a classification problem.

In open-ended crowdsourcing, the observability problem is more structural. The proposed two-tiered framework first partitions the quality model into task, worker, answer, and system, then refines each aspect through quality dimensions, evaluation metrics, and design decisions. The framework explicitly treats answer embedding, reliability estimation, aggregation, worker expertise, incentives, decomposition, workflow design, and system-level coordination as components of the quality model rather than ancillary metadata (Chai et al., 2024). This reframes quality control from an after-the-fact judgment on answers to an end-to-end design of information flow.

Provenance is similarly central in theoretical data services. TheoSSA distinguishes precomputed and custom spectral energy distributions, and TMAW-produced SEDs are flagged with metadata such as DataID.Creator=TMAW to indicate provenance and calculation method. The service differentiates quick standard models from full, detailed atomic-data models and emphasizes warnings about included elements, atomic detail, and suitable wavelength regions (Rauch et al., 2011). In large-scale MRI processing, the same concern appears in a different form: a team-oriented QC pipeline standardized visualization through PNG files, standardized judgment through “Yes”, “No”, and “Maybe”, and standardized aggregation through uniform CSV outputs across 16 processing pipelines (Kim et al., 2024). In both cases, provenance and representation are not peripheral; they are part of the quality object itself.

3. Monitoring formulations and analytical methodologies

The methodological repertoire for quality control problems spans SPC, statistical modeling, machine learning, active learning, and control theory. In multi-product manufacturing, one approach is to standardize measurements product-wise and then apply joint control charts. The study on a multi-product environment uses per-product normalization analogous to short run SPC, then applies Individual Range and EWMA charts to the standardized data, with

$D$6

and follows any out-of-control signal with process-information analysis via partition trees for root-cause identification (Muehlenstaedt, 2018). Here the quality control problem is not merely signal detection; it is detection coupled to diagnosis.

Measurement error fundamentally alters such monitoring schemes. For multivariate CUSUM-CoDa charts, the effect of a linear covariate error model was assessed through Average Run Length, using a Markov chain method across four variance-covariance scenarios. The reported result is that the ARL increases with an increase in the value of the error variance-covariance matrix, while the ARL decreases with an increase in subgroup size $D$7 or the constant powering $D$8 (Imran et al., 2022). For binary non-conformance monitoring, the corrected EWMA $D$9-control chart explicitly inverts the misclassification model. Its corrected proportion is

$\rho$0

and the corrected chart uses asymmetric control limits to remain valid under misclassification and small sample size (Chen et al., 2022).

When the monitored object is functional and multimodal, the monitoring statistic itself changes. The Functional Mixture Regression Control Chart models in-control heterogeneity through a mixture of functional linear models estimated by EM after projection onto principal component scores, and monitors new profiles using a likelihood-ratio-based statistic often written as

$\rho$1

The number of clusters and covariance parameterization are selected via BIC, and a studentized version is available for robustness (Capezza et al., 2024). This formulation treats multiple nominal regimes as part of the in-control population rather than as nuisance variation.

Machine-learning-based quality control further extends the methodology toward predictive and label-efficient regimes. In the casting case, Random Forest and Gradient Boosting were trained under CRISP-DM, with undersampling to address class imbalance and evaluation by accuracy, precision, recall, and F1-score, with contextual emphasis on recall (Rajasekaran et al., 12 Mar 2026). In seismic denoising QC, active learning selects examples by local classification error,

$\rho$2

estimated by a Nadaraya-Watson kernel regressor; selection may be deterministic or survey-style via inclusion probabilities proportional to estimated local risk (Chambefort et al., 2022). In video coding, PID-based quality control formulates frame-level QP adjustment to jointly minimize control error and quality fluctuation, leading to control errors consistently below $\rho$3, roughly $\rho$4 lower fluctuation than the prior method, and an added processing time of only $\rho$5 per frame (Song et al., 2017). The common pattern is that quality control increasingly acts before failure is finalized.

4. Representative domain realizations

The term “quality control problem” covers distinct technical objects: defective cores, code decay, frame-level distortion trajectories, synthetic spectra, crowdsourced answers, segmentation masks, resistance curves, and networked water quality states. Representative results illustrate how different these realizations are.

Domain	QC setup	Reported result
Foundry core-making	RF and GB predict five main defect classes from process, product, quality, and maintenance data	Five defect types accounted for 88% of failures; accuracy ranged from 41% to 74%; sand temperature was the most critical predictor (Rajasekaran et al., 12 Mar 2026)
Resistance spot welding	FMRCC monitors dynamic resistance curves with scalar covariates under multimodal in-control structure	On 37 known OC welds, FMRCC detected 32, giving $\rho$6 with 95% CI $\rho$7 (Capezza et al., 2024)
Cardiac MRI segmentation	Reverse Classification Accuracy predicts per-case segmentation quality without ground truth for the test case	99% accuracy on 400 scans, 95% classification accuracy on 4,800 scans, and 99.8% accuracy for CNN segmentations in large-scale validation (Robinson et al., 2019)
Large MRI pipeline review	Team-based visual QC with PNG generation, Flask app, and standardized CSV aggregation	QC of 200 segmentations took roughly 20 minutes in FSLeyes versus about 50 seconds in the QC app, a speedup of over 20 (Kim et al., 2024)
Virtual Observatory synthetic spectra	Benchmarking TheoSSA/TMAW SEDs against detailed TMAP models and AA Dor optical analysis	Ionizing fluxes were accurate to better than 10%; optical parameter estimation was targeted within 20%; AA Dor deviations were about 1% in $\rho$8 and 7% in $\rho$9 (Rauch et al., 2011)
Video coding	PID-based QP control for certain and constant quality	Control errors were below $1$0 and quality fluctuation dropped by about 75% (Song et al., 2017)

These examples show that quality control may target early defect rejection, per-case segmentation certification, benchmarked theoretical-data reliability, low-latency visual review, or feedback stabilization of a dynamic encoder. A plausible implication is that the unifying feature is not the artifact being inspected but the requirement to make a trustworthy go/no-go or adjustment decision under incomplete information.

5. Recurrent failure modes, ambiguities, and misconceptions

One persistent misconception is that quality control is equivalent to final inspection. Several studies instead frame traditional inspection as reactive. In manufacturing, visual inspections, measurements, SPC, and FMEA are described as time-consuming, expensive, and reactive, whereas predictive models support preemptive rejection or process correction before final casting (Rajasekaran et al., 12 Mar 2026). In large MRI datasets, outlier detection is explicitly said not to capture every instance of algorithm failure, which is why the proposed pipeline insists on visual inspection of every output, albeit in a scalable form (Kim et al., 2024). The literature therefore treats post-hoc inspection as insufficient rather than definitive.

A second misconception is that quality can always be reduced to a single consensus statistic. Open-ended crowdsourcing shows why this fails: answer spaces are large to infinite, acceptable answers are non-unique, and exact correctness may be undefined, so majority voting and Boolean-style probabilistic aggregation break down (Chai et al., 2024). A related issue appears in multimode profile monitoring, where a single functional linear model is inadequate when different operating conditions induce multiple in-control patterns (Capezza et al., 2024). In both cases, the “true” quality state is structured, heterogeneous, and sometimes non-unique.

A third misconception is that once a model reports a strong global metric, deployment is straightforward. Continuous software quality control emphasizes that automated tools can generate overwhelming quantities of data, false positives, and non-actionable results unless they are customized, aggregated, and role-specific (Deissenboeck et al., 2016). Automated brain MRI QC illustrates the same point statistically: training on the heterogeneous CATI dataset improved ROC-AUC on unseen data, and omitting preprocessing improved unseen-data ROC-AUC by about 0.10 for a model trained on a CATI subset, yet generalization remained questionable because site-wise and study-wise predicted probability distributions differed and optimal thresholds were unstable. For ABIDE-trained models the site/study-specific optimal thresholds ranged from 0.25 to 0.8, whereas CATI-trained models yielded more stable thresholds from 0.2 to 0.6 (Reguig et al., 2022). High aggregate discrimination does not eliminate deployment-specific calibration problems.

Measurement error is another recurrent source of distortion. In multivariate CUSUM-CoDa monitoring, measurement errors significantly affect chart performance through ARL changes (Imran et al., 2022). In binary non-conformance monitoring, ignoring misclassification biases the observed proportion, yields biased control limits, and reduces detection power; the corrected EWMA $1$1-chart is introduced precisely to eliminate this effect (Chen et al., 2022). Scientific data services face a parallel but non-statistical variant: users of virtual observatory spectra may not have the background to judge SED reliability, and without benchmark tests and metadata they cannot know the uncertainty associated with $1$2, $1$3, abundances, or spectral regions (Rauch et al., 2011). Across these cases, unmodeled uncertainty is itself a quality defect.

6. Research directions and broader significance

Several trajectories recur in recent work. In industrial predictive QC, recommended next steps include more advanced models such as LSTM for time dependencies, hybrid approaches, expanded data collection, and development of retraining pipelines (Rajasekaran et al., 12 Mar 2026). In open-ended crowdsourcing, the survey emphasizes generalized quality control frameworks, domain-specific design, intelligent workflow design, synergies with LLMs, and explicit mechanisms for quality control involving LLMs (Chai et al., 2024). These proposals indicate that quality control is moving toward continuously updated, workflow-level systems rather than static inspection modules.

Another direction is tighter coupling between quality objectives and operational control. In drinking water networks, quality-aware hydraulic control embeds water-quality controllability metrics into pump scheduling and studies the trade-off between pump control energy cost and water-quality performance across network sizes and scenarios (Elsherif et al., 2024). In multimedia delivery, PI controllers jointly regulate transmission and encoding rates to target video quality fairness and similar delivery delays among streams (Changuel et al., 2014). This suggests that, in networked infrastructures, quality control increasingly means shaping the dynamics that produce quality rather than merely screening the outputs.

Scalability and metrology remain decisive. For CFRP mirror manufacturing, a wide range of metrology techniques and quality control methods is needed across spatial frequency ranges, and automation together with improved optical testing for convex surfaces is identified as necessary for scalable, repeatable production (Soto et al., 2020). At the theoretical end, sublinear-time quality control shows that certification of typicality can be provably superpolynomially more efficient than full property testing in random graphs (Marcussen et al., 22 Aug 2025). The field therefore spans both practical streamlining of visual review and formal reductions in query complexity.

Taken together, the literature treats quality control problems as a family of tasks that combine specification, sensing, aggregation, uncertainty modeling, and intervention. The practical forms differ sharply by domain, but the underlying question is stable: what evidence is sufficient to trust an output, a process state, or an input instance well enough to act on it?