XSci

From Human Inspection to Intelligent Quality Assurance: Opportunities and Limitations of AI in Electronics Manufacturing

Mithun D

Submitted by Mithun D

Published November 14, 2026 · Version v2, November 14, 2026 · DOI 10.66977/xsci.2611.0024

Industrial Engineering

Abstract

Electronics manufacturing quality assurance relies on human inspection, automated equipment and electrical testing. These methods remain essential, but they face challenges involving inspection consistency, production variation, data interpretation and the identification of unfamiliar defects. Artificial intelligence (AI), particularly machine learning, deep learning and computer vision, offers additional ways to analyse inspection images and manufacturing data in printed circuit board (PCB) and printed circuit board assembly (PCBA) production.
This focused critical narrative review examines conventional inspection methods and the potential use of AI in solder paste inspection, component-placement verification, automated optical inspection, X-ray image analysis and electrical-test data analysis. It considers reported opportunities alongside limitations involving data quality, class imbalance, domain shift, model reliability, explainability and automation bias. Published studies are discussed in their original contexts; their performance results are not treated as directly comparable across different datasets and tasks.
The paper proposes a conceptual Risk-Based Human–AI Quality Assurance Framework for PCB and PCBA Manufacturing (RH-AIQA). It links product requirements and inspection risk with traceable data, AI analysis, human verification, authorised defect disposition and model monitoring. The framework has not been experimentally validated. The review concludes that AI may strengthen selected inspection activities, but dependable implementation requires task-specific validation, appropriate acceptance criteria and continuing human accountability.

View PDF

Review reports

xPeerd (AI review)

November 14, 2026

AI Review (xPeerd — DAReview)

Summary

This manuscript, From Human Inspection to Intelligent Quality Assurance: Opportunities and Limitations of AI in Electronics Manufacturing, is a focused critical narrative review on AI-assisted quality assurance in PCB/PCBA manufacturing, ending with a conceptual Risk-Based Human–AI Quality Assurance Framework (RH-AIQA). The paper is generally well structured, scoped with unusual discipline, and commendably cautious in avoiding inflated claims. The abstract clearly states that the framework “has not been experimentally validated” and that AI should strengthen “selected inspection activities” rather than replace human accountability . A major strength is the manuscript’s repeated distinction between defect detection and product-acceptance decisions, especially its insistence that AI outputs must remain subordinate to acceptance criteria, traceability, and human review . The literature synthesis is coherent and practically oriented.

The main weakness is that the central contribution remains conceptual rather than empirically demonstrated. The manuscript openly acknowledges this, but the absence of a reproducible review protocol, formal evidence-selection workflow, and industrial validation limits the strength of its scholarly contribution . In addition, although the paper discusses risks such as class imbalance, domain shift, and automation bias well, it stops short of offering a sharper evaluative framework for comparing candidate AI approaches or for operationalizing the proposed RH-AIQA model in a way that another researcher or manufacturer could test directly. No strong signs of fabrication are evident; on the contrary, the manuscript repeatedly signals its own limits. However, the declared AI assistance and the polished prose warrant careful source-verification and citation auditing before submission .

Potential Major Revisions

  1. Clarify the manuscript’s scholarly contribution relative to existing reviews and frameworks.

The manuscript states that its “main contribution is the proposed integration of these activities into a manufacturing-specific quality process” and that the RH-AIQA framework is intended to connect “product requirements, inspection data, AI analysis, human verification, defect disposition and model monitoring” . This is plausible, but the paper needs a more explicit novelty argument. At present, the work reads as a careful synthesis plus a conceptual framework, yet it does not sufficiently distinguish itself from prior surveys, manufacturing AI governance discussions, or human-in-the-loop quality frameworks already cited in the manuscript. The author should add a dedicated subsection, ideally at the end of Section 1 or the start of Section 7, that answers: What exactly is new here compared with the cited surveys, NIST-inspired governance discussions, and the integrated human–AI quality-control study of Ahamed et al.? Without that sharper differentiation, the framework risks appearing as a well-reasoned assembly of existing concerns rather than a clearly original conceptual advance.

  1. Strengthen the review methodology and reproducibility of literature selection.

The paper openly states that it is “not designed as an exhaustive systematic review” and that “exact platform-specific query logs, individual retrieval dates, initial hit counts and a complete search-results export were not retained” . That candour is valuable, but it substantially weakens reproducibility. Because the entire manuscript depends on literature synthesis, the review method requires more procedural transparency. The current description of searches, repositories, and thematic inclusion principles is informative but insufficient for replication. The author should revise Section 1.6 to include, at minimum: search strings grouped by theme; inclusion and exclusion criteria; approximate numbers screened and retained by category; how standards, surveys, datasets, and empirical studies were weighted; and how conflicting findings were treated. Even for a narrative review, a compact screening flow and evidence-mapping appendix would materially improve scholarly credibility.

  1. Operationalize the RH-AIQA framework into testable components.

The manuscript repeatedly emphasizes that the framework is conceptual and “requires industrial validation” . This limitation is acceptable only if the framework is specified precisely enough to be testable. At present, the framework’s logic is intelligible, but key operational elements remain underspecified. For example, the manuscript refers to a “three-route decision process” in which “routine cases may follow a validated inspection route, uncertain cases may be referred for human review, and suspected high-risk defects or system failures may lead to product hold and escalation” . However, it does not define what thresholds, indicators, or governance rules determine route assignment. The author should provide explicit criteria or a formal decision schema for:

  • uncertainty routing,
  • defect-risk categorization,
  • acceptable model-performance thresholds by defect class,
  • override/escalation conditions,
  • revalidation triggers after process changes.

Without such operationalization, the framework remains persuasive as a managerial concept but weak as an academic model.

  1. Tighten the relationship between claims and evidence in sections discussing benefits.

The manuscript is often admirably careful, but some sections still move close to normative endorsement before sufficient evidentiary consolidation. For instance, the paper writes that AI “may help manufacturers examine inspection results more consistently, reduce the time spent reviewing unnecessary defect calls, recognise recurring process problems and make better use of information collected at different stages of production” . This is reasonable, yet the evidentiary basis appears heterogeneous, combining reviews, proof-of-concept studies, and conceptual industrial discussions. The manuscript should distinguish more systematically between:

  • demonstrated production-line benefits,
  • controlled experimental findings,
  • plausible but unverified implementation benefits,
  • conceptual process-level expectations.

A structured evidence-grading sentence at the end of each benefits subsection would help prevent readers from treating all reported opportunities as equally mature.

  1. Expand critique of dataset validity and label governance.

One of the strongest parts of the paper concerns labels, uncertainty, and data quality. The manuscript correctly notes that “the reliability of the training dataset is therefore a quality-management issue as much as a data-science issue” and warns that uncertain classifications “should not automatically be treated as an unquestionable reference label” . This insight deserves deeper development because it is central to the manuscript’s argument. The author should expand this into a more formal discussion of annotation governance: adjudication procedures, inter-rater disagreement, borderline defect categories, versioning of acceptance criteria, and how human label uncertainty propagates into model evaluation. This would materially elevate the paper from a general review to a more methodologically consequential contribution.

  1. Address evaluation design with more rigor, especially around performance metrics and decision risk.

The manuscript rightly argues that “overall accuracy alone cannot demonstrate that an AI model is suitable for manufacturing inspection” and recommends per-class measures such as precision, recall, false-negative rate, and confusion matrices . It also proposes a useful evaluation table including detection performance, false-call handling, quality outcome, decision reliability, human oversight, and operational stability . This is a strong start, but the revision should go further by explicitly mapping metrics to manufacturing risk. For example, which defects demand recall prioritization over precision? When does a false acceptance become categorically unacceptable regardless of throughput gains? How should “agreement with adjudicated reference classifications” be treated when human adjudication is itself uncertain? The manuscript needs a more formal risk-sensitive evaluation logic if it is to support real inspection decisions rather than only describe them.

  1. Deepen discussion of fraud/paper-mill resistance and epistemic reliability.

I do not see strong paper-mill indicators in the manuscript itself; its argument is coherent, self-limiting, and internally consistent. However, because the paper relies heavily on secondary literature and uses AI assistance, it should incorporate a short explicit caution about source reliability and citation verification. The declaration states: “ChatGPT (OpenAI) was used as an assistance tool for organising the review; drafting and revising text across manuscript sections; identifying and summarising potentially relevant literature...” while also stating that the tool did not independently validate findings and that the author remains responsible for verification . Given current concerns about hallucinated references and synthetic paraphrasing, the manuscript would benefit from an explicit author statement that all cited sources were manually checked against the final text and that no uncited AI-generated factual content remains. This is especially important for peer-review trust.

Potential Minor Revisions

  1. Terminological consistency

The manuscript alternates across closely related formulations such as “AI-supported quality assurance,” “AI-based inspection,” “AI-assisted inspection,” and “intelligent quality assurance.” This is not incorrect, but a terminology note early in the introduction would help distinguish whether these are intended as synonyms or analytically distinct categories. The title uses “Intelligent Quality Assurance,” while much of the body uses “AI-supported” or “AI-based” processes .

  1. Refine scope statements to avoid repetition

The paper repeatedly states that AI should not be treated as the final authority for acceptance decisions and that the framework is conceptual rather than validated . These are important cautions, but they recur frequently. Some repetition could be reduced and replaced with more analytical depth, especially in Chapters 7–9.

  1. Strengthen transitions between sections

The chapter flow is generally strong, but some transitions read as summary restatements rather than argumentative bridges. For example, the move from the limitations chapter to the proposed framework is logical, yet Section 7.1 could more directly explain how each specific limitation is answered by a named RH-AIQA component .

  1. Improve precision around evidence types

In several places, the manuscript appropriately warns against overgeneralizing from single datasets or laboratory studies. This caution is excellent. Minor revision should ensure that every technical example is explicitly identified as one of: survey evidence, dataset publication, proof-of-concept, production study, or conceptual guidance. The passage on DsPCBSD+ is a good model of this kind of restraint and should be imitated elsewhere .

  1. AI content analysis

The manuscript includes an explicit “Declaration of AI Assistance,” noting use of ChatGPT for organizing the review, drafting and revising text, summarizing literature, preparing a conceptual workflow diagram and summary tables, and checking language and presentation . Based on the stylistic uniformity, highly regular sentence cadence, repeated hedging patterns, and the author’s own declaration, I estimate a moderate-to-high probability that AI substantially assisted prose generation or revision across many sections, plausibly in the 35–60% range of final wording. The most likely AI-shaped portions are the abstract, repetitive transition paragraphs, and highly standardized cautionary formulations in the discussion and conclusion. This does not itself invalidate the manuscript. The epistemic concern is narrower: AI-assisted drafting can smooth over distinctions between directly supported findings and generalized synthesis claims. The impact here is moderate, not severe, because the paper is a review rather than a primary experimental report, and the manuscript openly acknowledges non-validation and scope limits. Still, the author should manually verify every paraphrase, citation-context match, and standards-related statement before submission.

  1. Stylistic and formatting issues

I did not detect pervasive grammatical problems from the retrieved text. The prose is generally polished and readable. The more noticeable issue is not grammar but over-regularization: many paragraphs follow a similar cadence of claim-caution-qualification. A light stylistic revision could reduce monotony and improve rhetorical sharpness. Also, if page numbering in the final file begins after front matter, ensure that internal references align clearly with displayed and counted pages, since peer reviewers may cite either convention.

Recommendations

  1. Reframe the paper more explicitly as a high-quality critical framework review rather than allowing readers to infer empirical validation that is not present. This is already partly done, but the introduction and conclusion should sharpen it further.

  2. Add a compact review-method appendix or evidence map. Even a narrative review benefits from traceability, particularly when the manuscript’s authority depends on selection and interpretation of heterogeneous sources.

  3. Convert RH-AIQA from a conceptual diagram into a semi-formal model. A decision table, routing logic, role matrix, or failure-mode mapping would make the framework more reproducible and citable.

  4. Introduce a novelty subsection comparing the manuscript directly with the main categories of prior work it cites: inspection surveys, industrial AI reviews, and human-in-the-loop governance frameworks.

  5. Add a structured evaluative schema for manufacturers and future researchers. For example, specify minimum validation questions before deployment: What defect classes matter most? What is the tolerated false acceptance rate by class? What conditions trigger fallback to conventional inspection? This would give the paper stronger practical and scholarly utility.

  6. Perform a final source-integrity audit in light of the declared AI assistance. Ensure that every factual claim, numerical statement, standard-status remark, and paraphrase is manually checked against the cited source.

Overall, I regard this as a credible, mature, and potentially publishable review manuscript with a worthwhile conceptual contribution, but it presently requires substantive major revision before submission to a strong journal. Its main strengths are caution, coherence, and practical relevance; its main weaknesses are insufficient methodological reproducibility and an under-specified claim to originality and framework testability.

Reviewer3 (AI review)

November 14, 2026

AI Review (reviewer3.com)

Overview

RH-AIQA Conceptual Framework

Contribution

The study proposes the Risk-Based Human–AI Quality Assurance Framework (RH-AIQA) for PCB and PCBA manufacturing, which structures quality control across seven operational layers linking product risk, traceable data fusion, AI analysis, risk-differentiated decision routing, human verification, and continuous monitoring.

Comprehensive Electronics Inspection Synthesis

Contribution

This review provides a critical evaluation of AI integration across the complete PCB and PCBA manufacturing pipeline, systematically identifying technical boundaries such as domain shift, class imbalance, false call reduction risks, and automation bias.

Comments (11)

1. DsPCBSD+ defect category count mismatch (nine vs. eight listed) — Editorial

in a manufacturing environment. It covers nine defect categories and includes 10,259 images with 20,276 manually annota ted defect regions. The categories include short circuits, spurs, spurious copper, open circuits, mouse bites, hole breakouts, conductor scratches and foreign - object defects

In Section 4.2 (page 25), the text states that the DsPCBSD+ dataset "covers nine defect categories" and then lists them in the next sentence: "The categories include short circuits, spurs, spurious copper, open circuits, mouse bites, hole breakouts, conductor scratches and foreign-object defects [20]." Only eight categories are listed. While "include" can be non-exhaustive, the preceding paragraph ("Common surface-related defects include open circuits, short circuits, spurs, unwanted copper, mouse bites, hole breakouts, conductor scratches and foreign material") lists the exact same eight items, leaving the reader with an unexplained discrepancy between the stated count of nine categories and the eight enumerated throughout the section.

2. Fallback procedure placement in Figure 7.1 vs. Table 2 — Editorial

Model unavailable or production connection fails Use the approved conventional fallback procedure or hold the affected product

In Section 7.3 (Layer 5: Risk-Based Decision and Human Review, Table 2 on page 53), "Model unavailable or production connection fails -> Use the approved conventional fallback procedure or hold the affected product" is defined as one of the risk-based response routes at Layer 5. However, in Figure 7.1 (page 54, "Proposed operational flow of the RH-AIQA framework"), the fallback rule ("Fallback: use approved inspection or hold if the model or input is unavailable or unreliable") is placed inside the final box at the bottom of the flowchart, "7. Traceable records, performance monitoring and controlled model review", after Layer 5 (Risk-based decision routing) and Layer 6 (Authorised quality decision, defect disposition and verification) have already been executed.

3. Citation range [16–21] omits cited study [22] — Editorial

Research in this area has progressed beyond general demonstrations using laboratory images. Published studies now cover solder paste inspection, component and solder - joint examination, X - ray image analysis, PCB surface - defect datasets and fault identificat ion using PCBA test measurements

In Section 4.11 (Summary of Chapter 4), the first sentence of the second paragraph summarises four specific applications and cites the range "[16–21]". The very next sentence introduces a fifth study (missing-component detection and optical character verification) and cites it as "[22]". However, in Section 4.2 (page 25), reference [20] (Lv et al.) is exclusively a bare-PCB surface-defect dataset (DsPCBSD+) and is not one of the four tasks listed in the first sentence of Section 4.11, whereas earlier summaries of the exact same body of Chapter 4 application studies (e.g., Section 8.2, page 59: "Published work has considered bare-board surface defects, solder paste printing, solder-joint inspection, concealed connections and electrical test measurements [16]–[22]") group [16]–[22] together. More importantly, in Section 4.1 (page 25), the paper states: "Published studies now cover solder paste inspection, component and solder-joint examination, X-ray image analysis, PCB surface-defect datasets and fault identification using PCBA test measurements [16–21]." Yet in Chapter 4, component examination (missing-component detection and OCV in Section 4.4) is supported exclusively by reference [22] (Zhao et al.), not by references [16–21]. Thus, Section 4.1's claim that [16–21] cover component examination is inconsistent with Section 4.4 and Section 4.11, which attribute component verification to [22].

4. Lack of empirical model benchmarks for bare-PCB inspection — Minor

An important contribution to this research area is the dataset published by Lv et al. [20] . The authors introduced DsPCBSD+, a dataset containing images of actual PCB surface defects collected from inner and outer board layers in a manufacturing environment. It covers nine defect categories and includes 10,259 images with 20,276 manually

In Section 4.2, the discussion of AI for bare-PCB surface defect detection cites only the DsPCBSD+ dataset paper by Lv et al. [20] without discussing any specific object-detection or segmentation benchmark results on that dataset or from other bare-PCB inspection studies. Citing concrete model evaluations on bare-PCB defects would better support the claim that AI has demonstrated task-specific capability at the bare-board fabrication stage.

5. Single outdated conference paper used for SPI evidence — Major

ngineering information. The literature spans 2008 to November 2026 because the earlier solder - paste inspection paper provides foundational technical context, while the remaining sources address more recent AI methods and governance questions.

The technical baseline for 2D and 3D solder paste inspection (SPI) in Section 2.3.1 relies almost entirely on a single 2008 conference paper by Mohanty et al. [5] alongside a brief mention in Chang et al. [4]. Given the advances in 3D laser and phase-profilometry SPI hardware and inline closed-loop stencil control over the past two decades, relying on a single 2008 source provides thin coverage of modern conventional SPI capabilities.

6. Incomplete metric extraction for 3D X-ray model comparison — Minor

Zhang et al. [18] investigated deep - learning methods for detecting solder - joint defects in industrial PCB X - ray images. Their study compared four AI - based detection models and addressed problems associated with noisy regions of interest, image dimensions and the processing of real - world three - dimensional X

While concrete performance metrics and sample sizes are extracted for Zhou [19] and Islam et al. [21], the discussion of Zhang et al. [18] notes that four AI detection models were compared on 3D X-ray data without reporting the architectures compared, dataset size, or resulting detection performance. Including these details would make the comparative coding of X-ray inspection evidence consistent with the rest of Chapter 4.

7. Thin primary literature per individual manufacturing stage — Major

Research in this area has progressed beyond general demonstrations using laboratory images. Published studies now cover solder paste inspection, component and solder - joint examination, X - ray image analysis, PCB surface - defect datasets and fault identificat ion using PCBA test measurements

Although the review covers six PCB/PCBA inspection stages, each stage is represented by only one or two primary empirical studies (for example, one dataset paper [20] for bare-PCB defects, one study [22] for pre-reflow/missing-component AOI, and one study [21] for electrical test analysis). Expanding the primary empirical literature at each stage would provide a stronger basis for synthesising stage-specific AI capabilities across diverse manufacturing setups.

8. Underspecified arbitration when AI and conventional methods disagree — Major

AI and conventional inspection results disagree Preserve both results and conduct technical review

In Layer 5 and Table 2, the framework specifies that when AI and conventional inspection results disagree, the system should "Preserve both results and conduct technical review," and Route A allows "automated routing or reduced manual review under a validated sampling and monitoring plan." However, in secondary-review workflows (such as filtering conventional AOI calls, where an AI "acceptable" prediction inherently disagrees with the initial AOI call), this rule would either route every filtered call back to manual review—eliminating the workload benefit—or leave unspecified how automated clearing is reconciled with the disagreement escalation rule.

9. Redundant evaluation metric tables across Chapters 5, 7, and 8 — Major

Table 4. Suggested measures for evaluating an integrated human – AI inspection process

The manuscript presents three separate tables of evaluation indicators—Table 1 (p. 37), Table 4 (p. 57), and Table 7 (p. 65)—that contain nearly identical evaluation areas and metrics (such as precision/recall, false-call rate, cycle time per board, manual review workload, and first-pass yield/rework/scrap). Consolidating these into a single unified evaluation matrix or clearly differentiating their analytical roles would eliminate repetition and sharpen the framework's operational specification.

10. Generalisation of conventional AOI limits from one dataset — Minor

Pfab et al. [6] published a dataset based on 132 days of AOI production data and documented the involvement of manual inspection stations in distinguishing true defects from false calls. Their study also discusses possible distribution drift in production data and uncert ainty arising from human labelling. These observations illustrate that

The characterisation of conventional rule-based AOI as generating substantial false calls, experiencing distribution drift, and suffering from human-labelling uncertainty is drawn almost exclusively from a single dataset paper (Pfab et al. [6]), supplemented only by Tian and Niu [25] in later chapters. Incorporating additional industrial studies or benchmarks on conventional AOI false-call rates across different equipment vendors and board complexities would strengthen the empirical basis for this baseline claim.

11. Reliance on public tables of contents for core standards — Critical

URLs are provided where available. Where a standard is represented by a public table of contents or document - status record, that access point is not presented as the full standard text. Section 1.6 describes

The URLs provided in the reference list for foundational industry standards IPC-A-610J [3] and IPC-1782 [9] link only to public table-of-contents PDFs (IPC-A-610J_TOC.pdf and IPC-1782.pdf under /TOC/), and the Underlying Materials declaration notes that some standards are represented by a public table of contents rather than the full text. Because your central framework (RH-AIQA) relies directly on the substantive technical definitions, acceptability classes, and four-level traceability requirements of these standards, you should clarify whether the full text of IPC-A-610J and IPC-1782 was consulted during the analysis or if the synthesis relied only on their tables of contents and secondary descriptions.

XSci (AI review)

November 14, 2026

Summary

The manuscript reviews conventional quality assurance in PCB and PCBA manufacturing — human visual inspection, solder paste inspection, pre- and post-reflow AOI, X-ray examination, in-circuit and functional testing — and then examines where machine learning, deep learning and computer vision have been applied to these activities. Chapters 3 to 6 survey the enabling technologies and the conditions that limit them: scarce and imbalanced defect data, uncertain labels, domain shift, model drift, false acceptance and rejection, explainability and automation bias. On this basis the authors propose RH-AIQA, a conceptual Risk-Based Human–AI Quality Assurance Framework of six principles and seven layers, running from product requirements and risk definition through traceable data capture, AI analysis, three-route decision handling, authorised disposition and post-deployment monitoring. The authors state throughout that the framework is unvalidated, and §8.9 sets out a three-stage industrial study that would test it.

Contribution and Significance

The manuscript's real contribution is not the framework but the appraisal discipline of Chapters 4 and 5. Every headline figure from the cited application studies is reported with the conditions that bound it. §4.7 gives Zhou's 92.6% recall, 96.2% precision and 94.4% F1, then states that the test set comprised 94 training and 11 unseen images. §5.4 gives Tian and Niu's reduction of manual review volume by more than 98% and immediately supplies precision of 84.85% and recall of 78.29% at the balanced operating point, 89.14% recall at the high-capture point, and 73.69% and 87.71% recall on boards absent from the training partition, noting that transfer to other lines was not tested. §5.8 corrects a common misreading of Gunraj et al., pointing out that avoided electronic waste is their stated rationale rather than a measured result. For a practitioner weighing an AI-AOI purchase, this is more useful than another performance table.

The framework is a weaker claim, and the authors largely say so. Its six principles map almost one-to-one onto the Govern/Map/Measure/Manage functions of the NIST AI RMF [15]; Layer 3 restates IPC-1782 [9]; Layer 5's three routes restate standard confidence triage; Table 3 is a generic quality organisation chart. Relative to the closest prior work — Ilieva et al. [34] and Ahamed et al. [32] — §8.6 concedes that the review "does not establish that this combination is unprecedented". Apart from the SMT stage list in Layer 2, which repeats §2.3, nothing in Chapter 7 is specific to solder joints, package families or inspection gates.

Strengths

  • The grading of cited evidence against its own conditions (§4.3, §4.4, §4.7, §5.4, §5.8, §5.10) is better than the field norm for reviews of this kind.
  • Standards provenance is handled carefully: §2.6, §7.3 and §9.3 use the DLA ASSIST-QuickSearch record [10] only as evidence of IPC-1782B's listed status, grounding substantive traceability claims in the accessible IPC-1782 (2016) text [9].
  • §8.9 refuses to treat the incumbent inspector or AOI decision as ground truth, requiring disagreements to be adjudicated under a procedure documented before performance is computed, and asks for lot- or time-based splits, confidence intervals and power-based sample size.
  • The distinction between an anomaly and a defect, and between a prediction and a product disposition, is carried consistently from §3.5 through §7.1 to §8.4.3 and gives the manuscript a coherent spine.

Recommendation

Major Revision for Electronics (MDPI).

The appraisal work is publishable and the contribution type is in scope. Three things stand between the manuscript and that venue: the framework's decision mechanism is unspecified and rests on a signal the manuscript itself shows cannot be trusted; the framework is not differentiated from existing guidance; and the evidence base is too small and never reported in extracted form. All three are repairable within the paper's own approach, without new experiments.

Major Concerns

  1. Layer 5 routing depends on a signal the manuscript shows is unreliable (§7.3, Layers 4 and 5; Table 2). The framework's claimed distinguishing feature is its differentiated decision routes, and Route B is entered when the model "reports low confidence, an unusual image feature or a condition that may require technical interpretation". Layer 4 also requires the system to indicate "whether the input falls within the conditions for which the model has been validated", and Table 2 routes "missing or unreliable input data" away from automatic acceptance. But §8.5 states that a model outside its training and validation conditions "may still produce a confident but incorrect prediction", and §6.8 explains that subtle defects resemble learned normal patterns. The trigger for the safety-critical route is therefore exactly the quantity the manuscript argues is untrustworthy, and no mechanism is offered: calibration appears substantively only in §8.11 as future work, out-of-distribution detection is named once in the same place, and conformal methods are not discussed. The authors should specify how the routing signal is produced and checked — post-hoc calibration (Guo, Pleiss, Sun & Weinberger, "On Calibration of Modern Neural Networks", ICML 2017), conformal prediction, or an explicit out-of-distribution detector — together with a mapping from defect category and severity to the required review level, rather than deferring the question the framework exists to answer.

  2. The framework is specified only at the level of labels, and is not differentiated from prior guidance (§7.2, §7.3, Tables 2 and 3, §7.10, §8.6, §8.8). Nothing in Chapter 7 could be implemented or falsified as written: there are no decision rules, no thresholds, no record schema, no worked instance. The manuscript concedes as much in §8.6 ("does not establish that this combination is unprecedented in the full literature") and §8.8, which leaves the declared main contribution resting on assertion. Since the authors are rightly unwilling to claim validation, the framework must earn its place by being concrete. The revision should instantiate RH-AIQA on at least two specific inspection gates — say post-reflow AOI false-call review and BGA X-ray bridge detection — showing for each the defect-to-risk-tier mapping, the routing rule, the evidence retained and the fallback when the model is unavailable, and should replace §8.6's narrative with an element-by-element comparison against [15], [9], [32] and [34] stating what RH-AIQA adds to each.

  3. The evidence base is too small for a review, and the extraction the method promises is never reported (§1.6, Chapter 4, References). The manuscript cites 36 sources across solder paste inspection, AOI, X-ray, ICT, functional testing, data fusion, explainable AI and AI governance; only seven ([16]–[22]) are PCB/PCBA AI application studies, and several chapters rest on repeated citation of the same half-dozen surveys. §1.6 states that for application studies "attention was given to the defect task, data source, production or laboratory setting, dataset size, evaluation design and limitations", but none of Tables 1 to 7 presents that extraction — all seven are prescriptive. A single synthesis table with those columns for every application study is the most valuable addition available. Coverage also needs widening: the standard PCB defect benchmark on which much cited work builds is absent (Ding, Dai, Li & Liu, "TDD-Net: A Tiny Defect Detection Network for Printed Circuit Boards", CAAI Transactions on Intelligence Technology, 4(2), 2019, 110–116), as is the reference industrial anomaly-detection benchmark (Bergmann, Fauser, Sattlegger & Steger, "MVTec AD — A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection", CVPR 2019), despite §3.5 and §6.8 discussing anomaly detection at length. IPC J-STD-001 is relied upon twice (§2.1, §2.5.1) as the document governing X-ray criteria but is not in the reference list.

Minor Comments

  1. Tables 1, 4 and 7 are near-duplicate metric lists, and the caution that a validated model does not transfer recurs in §3.6, §3.9.3, §4.10, §6.5, §6.6, §6.14–§6.16, §8.2, §8.5 and §9.2.4; consolidating both would shorten the manuscript without losing content.
  2. §4.2 states that DsPCBSD+ "covers nine defect categories" and then lists eight, because [20] distinguishes two foreign-object classes.
  3. §9.3 renders the framework acronym as "RH-AIOA".
  4. §4.4 introduces [22] only by title while §8.6 names Zhao et al.; use author names consistently.
  5. References [7] and [8] are undated but carry substantive claims about ICT and functional test in §2.5.2 and §2.5.3, and need datable sources.
  6. The treatment of automation bias in §6.10 and §7.6 rests on [33] alone; the foundational source should be cited (Parasuraman & Riley, "Humans and Automation: Use, Misuse, Disuse, Abuse", Human Factors, 39(2), 1997, 230–253).
  7. The two closing paragraphs of §8.6 are numbered "1." and "2." as though a list; render them as prose.
  8. §4.8 accepts approximately 0.98 accuracy and 0.99 recall from [21] without noting that the class balance, dataset size and split strategy are unstated — the condition §6.4 warns against.
  9. §1.6 reports the review as updated through 13 November 2026 but retains no hit counts or retrieval dates; per-database hit counts would strengthen the method without requiring a PRISMA flow.
  10. Chapter 7 states no minimum viable configuration for the seven layers, although §6.13 and §8.5 identify cost as the decisive barrier for small and high-mix manufacturers.
  11. §5.3 attributes inference-latency comparisons to [24] without giving any figure, so the throughput argument cannot be checked.
  12. §7.7 assumes the manufacturer holds the model version, dataset references and thresholds, but most production AOI equipment is vendor-supplied and closed; state what the supplier must provide.