Solder Joint Inspection AI: AOI False Call Reduction, Escape Control, RFP and FAT/SAT

Introduction: When Most AOI NG Calls Turn Out to Be Good Boards

This article sets out how to think about introducing solder joint inspection AI, written for plant managers, production engineers, quality assurance and IT staff at factories in Thailand that build printed circuit board assemblies (PCBA). When you check them, most of the NG calls made by the AOI on an SMT line turn out to be good boards. Even so, several people sit at the review station on every shift, looking at the screen board by board and pressing the “good” button.

False calls increase with every new product launch, and tuning the AOI program depends on a few specific engineers. Automotive customers demand that no escapes reach them and that every board be traceable. Then the AOI vendor proposes an “AI option,” and you are not sure what to evaluate or how to decide whether to buy it. If this sounds familiar, this article is for you.

The topic here is adding AI judgment to the inspection of solder joints on PCB assemblies. Specifically, the article focuses on AOI false call reduction, review workload, escape management, traceability and acceptance. General implementation steps for AI visual inspection, such as choosing lighting and annotation in general, are covered in our AI visual inspection implementation guide.

Here is the conclusion up front. AI for solder inspection is not “magic that makes the AOI smarter.” The goal is one of the following three, and the evaluation method changes depending on which one you aim for.

1. Keep good boards among the AOI’s NG calls from going to review (false call reduction)
2. Reduce the workload of creating and tuning AOI programs
3. Reduce variation in human judgment

Reducing false calls comes at the cost of a higher risk of escapes. For that reason, the pass criteria for the PoC and for acceptance set a “false call reduction rate” and an “upper limit on the escape rate” as a pair, and they are judged on data from a period later than the period used for training. The AI’s role stops at automatically removing boards it judges good from review; the final judgment of defects and rework stays with people and your existing criteria (the IPC classes). Also, AOI can only inspect joints it can see. Hidden joints such as BGAs are the domain of X-ray inspection (AXI), and adding AI judgment does not change the division of roles between AOI and AXI.

This article does not recommend or rank specific products. Product capabilities are presented as vendor announcements. The cost-benefit analysis later in the article is an estimate built on hypothetical figures prepared for this article.

Why Thai PCB Assembly Plants Are Revisiting Solder Inspection Now

First, a look at where Thailand’s electronics industry stands. The figures are given with their source and scope as published; we do not add them together or apply them to any particular company.

• According to a Thailand Board of Investment (BOI) press release dated August 3, 2026, cumulative applications for investment promotion in the printed circuit board (PCB) sector from 2023 to mid-2026 totaled 224 projects worth USD 9.85 billion (about THB 331 billion). These are application figures, not approved amounts.
• In the same release, the BOI Secretary General said that “more than half of the world’s top PCB manufacturers have chosen Thailand as a production base.” The same release, however, describes the shortage of skilled workers as a “major bottleneck” in relocating advanced manufacturing. BOI and NXPO are putting a talent-matching framework in place.
• According to a Nation Thailand report dated September 25, 2026, Thailand’s exports in August 2026 rose 24.3% year on year, the fastest growth in 56 months. Products that grew included computers and parts, integrated circuits and printed circuits.
• In a Nation Thailand report dated August 21, 2026, Aat Pisanwanich of Rangsit University put electronics exports in the first half of 2026 at THB 1.63 trillion (up 46% year on year). Sampan Silapanad, president of the Electronic Computer Employers Association, said employment in the electronics industry and its supply chain had grown from about 800,000 to about 1 million people.
• At its meeting on July 8, 2026, the BOI approved projects including plants for copper-clad laminate (CCL), prepreg and glass fiber cloth for PCBs. These are investments in board materials, that is, upstream of the bare board, not investments in assembly (SMT) plants.
• According to an Electronic Design article dated August 26, 2026, the BOI has launched a program to develop more than 100,000 highly skilled workers through training, bootcamps, on-the-job learning and specialized courses. The article notes that manufacturing bases are concentrated in Chachoengsao, Chonburi and Rayong.
• According to an iConnect007 article dated April 30, 2026 (on TPCA’s Thailand PCB Summit), more than 60 PCB manufacturers have entered Thailand in recent years, and supply chain companies number more than 100.

More board plants and too few people means that absorbing inspection review workload by adding headcount cannot continue. The process in which people re-check every AOI NG call is the first candidate for rethinking.

The October floods also deserve a mention. The two estimates differ in timing and scope, so they are listed separately.

• Reports on October 2 to 3, 2026 citing an estimate by the Federation of Thai Industries (FTI) put direct losses to manufacturing from floods in Bangkok and 14 provinces at about THB 1.018 billion. They said Phases 7 to 10 of Amata City Chonburi Industrial Estate were flooded to a depth of 30 to 70 cm and that at least 34 factories were directly affected.
• A Nation Thailand article dated October 8, 2026 reported that 34 factories in Zones 7, 8 and 9 of the same estate had stopped production, more than 10 of them automotive parts plants. In this article, the estimate of industrial damage was put at nearly THB 6 billion (up from an earlier estimate of THB 1.5 billion).

Neither article mentions damage to electronics or PCB plants. The only conclusion that can be drawn is from a business continuity (BCP) perspective. If AOI images, AI judgment records and trained models are kept only on servers inside the factory, they could all be lost together in a flood. Decide where inspection data will be backed up at the design stage of the AI project. As of September 26, 2026, the Industrial Estate Authority of Thailand (IEAT) had ordered 24-hour flood monitoring at all industrial estates under its jurisdiction.

The Big Picture of Solder Inspection: Roles of SPI, AOI, AXI and Visual Inspection

On a typical SMT line, solder printing is followed by SPI (solder paste inspection), then component placement and reflow, and then AOI (automated optical inspection). Boards with hidden joints such as BGAs go through AXI (automated X-ray inspection) as needed. Boards on which the AOI makes an NG call go to the review station, where a person looks at the images or the actual board and decides pass or fail. THT (through-hole) components are checked by AOI or visually after wave soldering or selective soldering.

AI judgment mainly sits at the point in this flow where it is decided “whether to send an AOI NG call to review.” Some products apply AI to SPI or AXI data, but this article focuses on the review process after AOI.

On the division of roles between AOI and AXI, inspection equipment vendor GÖPEL electronic explains that what the human eye cannot see cannot be inspected optically by AOI either. According to the company, AXI can also detect hidden joints such as those under BGAs and QFNs, and for BGAs it inspects for head-in-pillow, voids and similar defects. It describes a division in which only selected hidden joints need to be inspected by X-ray, while AOI handles the remaining joints.

On the standards side, it is also important to get the scope right. IPC-A-610J (Acceptability of Electronic Assemblies, March 2024 revision) is described, in the scope shown in IPC’s table of contents sample, as a collection of visual acceptability requirements for electronic assemblies. The same section states that it does not include criteria for cross-section evaluation or X-ray inspection, and that X-ray guidelines are in J-STD-001. Mapping AOI judgment logic to clauses of IPC-A-610 is natural, but trying to explain AXI judgments with IPC-A-610 falls outside its scope.

Inspection cycle time can also change when AI is introduced. Whether the processing time of AI judgment affects line flow is measured at the acceptance stage, using the same approach covered in our article on high-speed inline AI visual inspection.

AOI “False Calls” and “Escapes”: What Is the Problem?

Start with the definitions. An HP article published in Electronic Design in 1998 defines a false call as “something judged defective that is not actually defective” and an escape as “a defect that was not called.” The article is old, but the definitions still hold.

A paper published on arXiv in 2025 by researchers at Siemens AG and the University of Helsinki uses a production SMT line as its example and frames this problem as a business metric. Its premise is that non-defective boards move on to the next process and boards judged defective go to a manual inspection station (MIS). The paper defines two metrics.

• Volume reduction: how much the number of good boards sent to the MIS was reduced. Defined as TN/(TN+FP); in this article we call it the “reduction rate of good boards sent to manual review”
• Slip rate: the share of defects that were missed. Defined as FN/(TP+FN); in this article we call it the “escape rate”

The paper treats these two as key business metrics and states that escapes are far worse. A false call costs review time, but an escape means a defect flows to downstream processes or the customer.

Why do false calls occur? A 2020 paper by Bosch and TU Braunschweig explains that shadows, reflections from adjacent components and similar factors produce both escapes and pseudo defects (defects that are not real) in inspection results. Loosening thresholds on the AOI side reduces false calls, but it also makes it easier for real defects to pass.

This creates a structural pressure. A shop floor that wants to cut review workload is tempted to tune toward fewer false calls. Without a way to measure whether that tuning is increasing escapes, only the false call reduction gets reported, and the rise in escapes is first discovered through customer complaints. Adding AI does not remove this pressure. Boards that the AI judges good do not go through review, so the AI’s escapes flow straight to downstream processes. The general approach to designing over-detection and escapes together is covered in our article on false reject control in AI visual inspection.

What AI Solder Inspection Can Do: Three Uses

There are broadly three ways to use AI in solder inspection. The metrics that improve and the prerequisites differ depending on which one you aim for.

Vendors have also announced AI functions, each from its own angle. All of the following are vendor announcements, and their conditions are stated as given.

• Mycronic DeepReview (press release dated April 3, 2024): AI-based automatic defect classification for the MYPro I series of 3D AOI. The company says it can remove 50 to 90% of the time operators spend reviewing false calls, but this is conditional: it depends on the share of eligible components. It further says that when customers regularly retrain on their own database of inspection images, potential false calls can be reduced by 50 to 100% for eligible components. These figures apply only to eligible components, under an operation in which the model is regularly retrained on the customer’s own images. A company executive also said that human judgment cannot be replaced.
• Mycronic Escape Tracker (introduction page for a company webinar): Presented as a function based on the idea of optimizing AOI program parameters and tolerances so that known false calls do not occur, while still catching every real defect.
• Koh Young (exhibition announcement for NEPCON ASIA 2024): The company says that Smart Review reduces false calls and operator intervention by automatically evaluating OCV and OCR (inspection of component markings) readings. This statement concerns automatic evaluation of marking inspection, and no false call reduction rate for solder joints in general is given. KAP is presented as a company function that uses AI technology, and KPO as one that uses deep learning technology.
• OMRON VT-S10 (article reporting an announcement dated October 8, 2021): The company says multi-directional, multi-color lighting and MPS-based 3D measurement remove the effects of shadows and secondary reflections. “Zero false calls and escapes” is an expression of the company’s goal or philosophy, not an achieved result. The VT-S1080 page on the company’s European site promotes AI support for defect judgment and AI technology that reduces programming and visual verification work.
• Saki 3Di-LS2-CASE (announcement in September 2021): The company says its AI engine, which learns from large numbers of good and defective images, eliminates the need for data tuning in solder inspection. Its statement about significantly reducing false calls concerns judgment of non-wetting.
• Viscom vAI (press release and product brochure dated October 15, 2025): The company says it can create and optimize inspection programs for 3D AOI and 3D AXI almost fully automatically. The product brochure states that an AOI inspection plan can be created in under 15 minutes. On training, the company says the model is pre-trained on millions of images from diverse production environments, with customer data added on top.
• MIRTEC INTELLI-PRO (article reporting an announcement dated September 2, 2021): For its auto-teaching function using deep learning, the company claims a 90% reduction in teaching work time compared with manual teaching and 50% compared with auto-teaching without deep learning. This is a claim about work time, not a reduction rate for false calls or escapes. The company also says that using deep learning for foreign object detection (FOD) sharply lowers false calls.
• TRI (exhibition announcement for IPC APEX 2025): The company says its optical inspection runs on AI algorithms and metrology functions, and it presents AI-related tools such as AI Station and AI Smart Programming. No figure for the false call reduction rate is given.

Even where figures are published, the target components, defect modes and retraining assumptions all differ. Vendor figures cannot be lined up and compared against one another. The only option is to measure on your own line, using your own images.

The Pitfall Research Reveals: Passes in Testing, Fails in Production

The Siemens AG and University of Helsinki paper mentioned above (arXiv, June 17, 2025) is the most useful case for thinking about PoC design.

The data are pin-level solder measurements taken by the AOI on a production SMT electronics manufacturing line. They are tabular, with 77 columns, ordered in time sequence. Defects are very rare, and the class imbalance is extreme. The data are publicly available.

The researchers set the target as “a reduction rate of good boards sent to manual review of 40% or more, and an escape rate of 1% or less.” This is an example of target setting in this study, not an industry standard value.

The results were as follows. When two models that met the target on the test data were applied to five evaluation intervals covering later periods in the time sequence, the escape rate came out at 5.3 to 9.4% for one model and 10.2 to 15% for the other. In every interval, it far exceeded the 1% target. The authors conclude that the models would fail in production from the start.

The paper also points out that standard accuracy is misleading. On data with very few defects, even a classifier that simply answers “good” for everything shows high accuracy. An explanation that “the AI’s accuracy is high” does not reveal how many escapes there are.

Translated into a factory PoC, the paper’s recommendations look like this.

• Set success criteria in business metrics first: use the reduction rate of good boards sent to manual review and the escape rate, not accuracy, and document the pass line before the PoC
• Split the data by time: evaluate on data from a period later than the training period. If the data are shuffled randomly before splitting, similar images from the same period end up in both training and evaluation, and the results look better than they are
• Check across multiple periods: passing in one evaluation period does not mean it will hold in another. Include periods in which component lots or seasons change

In addition, a 2025 paper (arXiv) by Ulster University and others, covering AOI in semiconductor manufacturing, presents an approach that adjusts tolerances based on data while maintaining a state in which all defects are caught (100% recall). It is not a study of PCB assemblies, but its sequence of “fix the upper limit on escapes first, then reduce false calls” is also useful for solder inspection.

Deciding the Scope: Which Components, Joints and Defect Modes to Start With

There is no need to introduce solder inspection AI across the whole factory at once. Narrowing the scope by component type and defect mode makes the PoC easier to evaluate.

• Chip components: They are numerous and tend to generate many AOI NG calls. A 2020 paper by Bosch and TU Braunschweig used production images provided by Bosch to study a CNN that sorts chip components (C0402 to C1210, R0402, R0603) into five classes: good, missing, foreign object, misalignment and insufficient solder. The aim was to reduce the risk of defects escaping, as a second-stage judgment. On the C0402 test set, recognition performance of up to 97.6% was reported. This is a result for this component group and these data. The paper compares several training datasets of up to 300,000 images.
• QFP and SOP leads: There are lead-specific defect modes such as bridging and lifted leads. Mycronic lists bridging, coplanarity and others as targets of DeepReview (vendor announcement).
• THT (through-hole components): These are joints made by wave soldering or selective soldering. A 2025 paper by researchers at FAU (Friedrich-Alexander-Universität Erlangen-Nürnberg) studies a method that uses GAN-generated synthetic data and active learning to address the scarcity of defect data in classifying THT solder joints. Difficulty in collecting defect samples is a challenge shared by many factories.
• Specific defects such as solder splash: A 2023 paper in Scientific Reports studies detecting solder splash with object detection models on a specific power module board. Comparing seven models, it reports a detection time of 90 milliseconds and a mean average precision of 96.6% for a YOLOv8n model with 1.9 million parameters. This is a result for a specific application commissioned by industry and does not carry over directly to other boards.
• Hidden joints such as BGAs and QFNs: As noted above, these cannot be seen by AOI and are the domain of AXI. A 2020 study at FAU examined a CNN that classifies voids and head-in-pillow in 2D X-ray images and states that it significantly reduced false calls compared with a commercial X-ray system. This, however, is a separate matter from AI judgment on AOI.

Good candidates for the first target are components and defect modes that meet all three conditions: “many NG calls,” “human judgment tends to vary” and “the defect mode is easy to define.” For defect modes with few defect samples, the PoC may not collect enough cases to measure the escape rate.

Some causes of solder defects lie before inspection. Defects such as those from reflowing components that have absorbed moisture will not occur any less often however smart the inspection becomes. Upstream control is covered in our article on MSD management systems.

Training Data and Labels: Your Own Images, Golden Boards and Judgment Criteria

AI judgment depends on the quality of the images and labels used for training. Here are the points specific to solder inspection.

Golden boards: A blog post by VSE explains that the first step in AOI programming is to build a known golden board as the board for first article inspection, and that all component types and values must be checked manually for correctness. The same applies to AI training: you need a board and images that serve as the reference for ground truth. The same blog post also mentions lowering image recognition sensitivity to reduce false calls or raising it only for critical areas, as well as programming that accepts alternate part numbers so that legitimate substitute components are not treated as defects.

Pre-training and your own data: Vendor explanations differ. Mycronic says the model can be regularly retrained on the customer’s own inspection images and adapted to the customer’s inspection criteria. Viscom says it trains by integrating customer data into a model pre-trained on millions of images from diverse production environments. With either approach, how far retraining on your own images is assumed, and who does that work, are points to confirm in the RFP.

Judgment criteria and IPC classes: Before labeling, decide “whose criteria, and for which class, will determine pass or fail.” IPC J-STD-001J (March 2024 revision) divides products into Classes 1, 2 and 3 and states that it is the responsibility of the user (the customer placing the order) to specify the class in which an assembly is built. If the user does not specify and document the class, the manufacturer may do so. The classes are said to reflect differences in producibility, complexity, functional performance requirements and the frequency of verification (inspection and testing).

The same joint may pass or fail depending on the class. In a factory where the class differs by customer, if labels are mixed into a single model, the model learns contradictory criteria. Record on each label “which class’s criteria it was assigned under.” If reviewers’ judgments vary, training on those judgments as they are teaches the variation too. How to align judgment criteria and accept annotation work is covered in our article on annotation.

Setting Pass Criteria: False Call Reduction and Escape Rate as a Pair

Build the PoC evaluation table with four columns: metric, definition, data period and judgment. Numerical targets are for each company to decide. The way the “judgment” column is written in the table below is one example, not an industry standard.

For reference, the study above set “a reduction rate of 40% or more and an escape rate of 1% or less” as its target. This is an example of target setting in a study. Your own targets are set by customer requirements, product class and whether downstream inspection exists.

To measure the escape rate, boards that the AI judges good must also be checked against human judgment. For that reason, the PoC includes a “parallel run” period in which the AI’s judgments are only recorded and people continue to review every board as before. Collect the boards on which human and AI judgments disagree and confirm which one was actually right.

One more caution is that the number of escapes is small. On a line where defects themselves are rare, a short period cannot measure the escape rate with sufficient precision. Determine the length of the parallel run by working back from the defect occurrence frequency.

Traceability to Board Serial Numbers and MES Integration

Adding AI judgment increases what must be recorded for each board: AOI images and measurements, the AI judgment and its confidence, the model version, the final human judgment and rework records. Unless these are linked to the board serial number (or panel ID and lot), you cannot later trace “which model version judged this board good.”

Related standards are introduced here by scope only.

• IPC-1782B (Standard for Manufacturing and Supply Chain Traceability of Electronic Products; listed on a sales site as issued September 1, 2023): A standard that sets minimum requirements for manufacturing and supply chain traceability based on perceived risk. Its scope states that it applies to products, processes, components, equipment and so on used in the manufacture of printed board assemblies.
• IPC-2591 (IPC-CFX): IPC’s official page lists Version 1.7 (August 2023) as the latest version. It is a standard that applies to communication between all automated, semi-automated and manual processes in the manufacture of printed board assemblies. GÖPEL electronic says that AOI, AXI and SPI can be integrated with MES through IPC-CFX.

We could not confirm any specific number of years required by standards for retaining images or judgment records. Set the retention period based on customer requirements and internal rules, and design the storage location based on image volume.

How fine-grained the linkage should be is not something inspection alone decides. Production management across an electronics plant is covered in our article on production management systems for electronics and EMS, the units to trace in our article on electronics traceability, and how to assign and read serial numbers in our article on serial number management systems. For designing the evidence trail that meets automotive customer requirements, our article on IATF 16949 traceability is also useful.

When a Defect Is Found: Rework and Upstream Feedback

Boards that the AI and people judge defective go to rework. A reference standard for rework is IPC-7711/7721D (Rework, Modification and Repair of Electronic Assemblies, December 2023 revision). This standard also uses the Class 1, 2 and 3 categories and states that the product class should be specified by the user and, if not, by the manufacturer.

For THT, there are also examples of passing inspection results directly to rework equipment. A Circuits Assembly article dated October 11, 2017 describes a method that inspects after wave soldering, finds problems with through-hole joints, passes data from the AOI to a selective soldering machine and reworks only the affected joints. If AI judgment results are recorded by board serial number and joint location, they can also feed this kind of localized rework. The flow of rework instructions, quarantine and re-inspection is covered in our article on rework control systems.

Feeding inspection data back upstream also matters. Without relying on any specific vendor’s product function, linking SPI measurements and AOI defect modes by the same board serial number lets the process side trace “which printing deviation led to which defect.” The approach to detecting process changes early is covered in our article on SPC anomaly detection AI, and the procedure for tracing back from a defect to the process that caused it in our article on in-process defect tracing.

If an escape is found in a downstream process or at the customer, trace that board’s AOI images, AI judgment and model version, and identify the range of boards judged good by the same model version. If the records are not linked to board serial numbers at this point, you cannot narrow the range and will end up sorting broadly.

After Go-Live: Retraining, Change Control and Drift Monitoring

AI judgment is not necessarily at its best on the day it is installed. As the study above showed, performance changes as the period changes. Here are the points to manage after go-live.

• Re-verification at changes: Re-verify AI judgment when a new product is introduced, when a component manufacturer or lot changes, when the solder material changes, or when the AOI program changes. Even a slight change in a component’s appearance can change the pattern of false calls and escapes. For where to catch changes and what to confirm on the first article, the framework in our article on 4M change management systems can be used.
• Model version control: Increment the model version with every retraining and record which version ran on which line from when. Make it possible to roll back to the previous version if a problem appears with a new one.
• Sampling audits for escapes: Keep an operation in which a set number of boards the AI judged good are sampled and checked by a person. Even after the parallel run ends, this is the only way to monitor the escape rate. Designing degradation detection and rollback is covered in our article on factory AI model monitoring.
• Golden board and equipment checks: If the condition of the AOI camera or lighting changes, the images change and so does the AI judgment. Periodically run the golden board and boards containing known defects, and confirm that the results have not changed. For how to keep check records, the approach in our article on calibration management systems is a useful reference.

Your own images for retraining accumulate every day in operation. Saving images that carry the final human judgment, together with class information, provides material for the next retraining.

Requirements to Write in the RFP

When comparing AOI vendors’ AI options or AI judgment software, writing the following requirements into the RFP aligns the assumptions behind the quotes.

Prices for AOI and AI options vary widely with configuration, number of units and scope of support. We could not confirm any published prices, so this article does not give price guidelines. When comparing quotes, asking for costs to be split into separate buckets as in the table above makes annual running costs visible. Choosing inspection equipment vendors in general is covered in our inspection equipment vendor selection guide.

FAT/SAT and Acceptance Criteria

Plan acceptance testing in two parts: the factory acceptance test before shipment (FAT) and the site acceptance test (SAT). In both, numerical targets are for your company to decide. The items below are examples of where to set targets.

FAT (verification before shipment or before installation)

• Reproduction on historical image sets: include images from periods later than the training period in the evaluation. Split into multiple periods by time and look at the results
• Known defect set: prepare a set of images of real defects and confirm that the AI sends all of them to the defect side (or to hold). This is to confirm there are no escapes
• Output of judgment records: confirm that board serial number, joint location, AI judgment and model version are output as records
• Handling of holds: confirm the behavior of passing images the AI cannot judge to a person

SAT (parallel run on the actual line)

• The AI’s judgments are only recorded, people review every board, and the two are reconciled
• Items to measure: reduction rate of good boards sent to manual review, escape rate, AI processing time, impact on line cycle time, data loss (number of cases where images or judgments were not recorded)
• Duration: set a length, based on defect occurrence frequency, long enough to collect enough defects to judge the escape rate
• Switchover after passing: remove boards the AI judges good from review only after the SAT pass has been confirmed. Continue sampling audits after the switchover

In the acceptance report, record which model version, which period of data, which metrics and which criteria were used for the judgment. This is so that after future retraining, results can be compared under the same conditions.

Cost and Benefit Model Estimate (Hypothetical Figures)

All figures from here on are hypothetical figures prepared for this article. We could not confirm official prices for AOI or AI options, so costs are also assumed. Please use this as a framework when running your own estimate.

Assumed factory: A Japanese-owned EMS in Thailand (serving automotive and industrial equipment). It has 4 SMT lines, each with 1 existing AOI, for a total of 4 units, 2 of which are older models. Two shifts, 250 operating days a year.

Current costs (baseline, annual)

Configuration A: Add AI Re-judgment to the 4 Existing AOIs

Add AI re-judgment (automatic defect classification) to the 4 existing AOIs. This includes review terminals, a data platform and linkage to board serial numbers; the AOI units themselves are not replaced. Assume an initial cost of 3,200,000 THB and annual running costs (license, retraining work, maintenance) of 400,000 THB.

The net benefit is 1,120,000 – 400,000 = 720,000 THB/year, and the payback period is 3,200,000 ÷ 720,000 ≈ about 4.4 years.

Configuration B: Configuration A Plus Replacing the 2 Older AOIs with New 3D AOIs

This is evaluated as an addition after Configuration A has been installed. The 2 older AOIs are replaced with new 3D AOIs that have AI functions. Assuming 2,500,000 THB per unit, the additional initial cost is 2 × 2,500,000 = 5,000,000 THB, and the additional annual running cost is 200,000 THB.

• A further 10% reduction in review workload, against the baseline of 1,800,000 THB: 180,000 THB
• A further 20% reduction in escape-related response, against the baseline of 1,500,000 THB: 300,000 THB
• Total additional benefit: 180,000 + 300,000 = 480,000 THB/year

The net benefit of the addition is 480,000 – 200,000 = 280,000 THB/year, and the payback period for the addition is 5,000,000 ÷ 280,000 ≈ about 17.9 years.

Sensitivity: If the Review Workload Baseline Were Twice as High (Configuration A)

If the review workload baseline were twice as high, at 3,600,000 THB/year, the 40% review reduction in Configuration A would be 1,440,000 THB. The total benefit would be 1,440,000 + 300,000 + 100,000 = 1,840,000 THB/year, the net benefit 1,840,000 – 400,000 = 1,440,000 THB/year, and the payback period 3,200,000 ÷ 1,440,000 ≈ about 2.2 years.

What the Estimate Shows

• The core of the benefit is adding AI re-judgment to the data from existing AOIs (Configuration A). It reduces review workload without replacing the AOI units
• On false call reduction benefits alone, the payback for equipment replacement (Configuration B) is long, at about 17.9 years. Replacement is decided based on defect modes the older models cannot see (such as height and lifted leads), customer requirements and equipment life. Starting AI with data from existing AOIs is the realistic approach
• The more people a factory assigns to review, the larger the benefit. In the sensitivity case, the payback period shrinks to about 2.2 years

There is an important assumption. The benefit of a “20% reduction in escape-related response” assumes that the AI does not increase escapes. Do not count this benefit until the escape rate has been confirmed in the parallel run. If escapes increase, the cost in this bucket will rise rather than fall. Also, the additional benefits of Configuration B are set as a percentage of the baseline, not of the amount remaining after Configuration A. Opportunity losses from line stoppages, new orders and reductions in audit workload are not included in this estimate.

Frequently Asked Questions

Will AI solder inspection eliminate AOI false calls?

It is safer not to assume it will. Even in vendor announcements, false call reduction figures are given with conditions such as target components, defect modes and retraining on your own images. OMRON’s “zero false calls and escapes” is also an expression of the company’s goal. Design on the premise of an operation in which images the AI cannot judge are passed to a person.

How much AOI false call reduction can we expect?

You will not know until you measure it on your own line. For example, Mycronic says potential false calls can be reduced by 50 to 100% for eligible components when the model is regularly retrained on your own images, but the effect varies with the share of eligible components. Because conditions differ between vendors, figures cannot be compared side by side either. In the PoC, measure the reduction rate of good boards sent to manual review using data from a period after the training period.

Can we stop visually checking boards the AI judges good?

Continue visual checks until the parallel run confirms that the escape rate is below your upper limit. Research has reported a case in which models that met the target in testing deteriorated to escape rates of 5.3 to 15% in later periods. Even after the switchover, we recommend continuing sampling audits of boards the AI judged good.

How much training data does PCB inspection AI need?

There is no single answer. It depends on component type, defect mode and whether a pre-trained model is available. In one research example, a study of chip components using Bosch production images used up to 300,000 training images. What tends to be the problem in factories is not the number of good-board images but the scarcity of defect images. Check first how many defect images can be collected, and confirm with the vendor how shortfalls are handled (such as synthetic data).

Can AI inspect BGA solder joints too?

Not with AI judgment on AOI. Joints under a BGA cannot be seen optically, so they are the domain of X-ray inspection (AXI). There are studies and products that apply AI to X-ray images, but that concerns AXI judgment; adding AI to AOI does not make it able to inspect BGA joints. Keep the division of roles between AOI and AXI unchanged, and consider whether to use AI for the judgments of each.

Should we replace our AOIs with 3D machines?

If false call reduction is the only reason, there may be no need to rush. In this article’s hypothetical estimate, the payback for the addition of replacing 2 older units with new 3D AOIs was about 17.9 years. Base the replacement decision on defect modes the older models cannot see, customer requirements and equipment life. AI can start with data from existing AOIs.

How many weeks does an automated solder inspection PoC take?

The duration depends on your production volume and defect occurrence frequency. The false call reduction rate can be measured even over a short period, but the escape rate cannot be judged until enough defects have been collected. Decide the number of defects needed first, then work back from that to the length of the parallel run. Also remember to separate the training period from the evaluation period.

Summary

• Solder joint inspection AI is not “magic that makes the AOI smarter.” The evaluation method changes depending on whether you aim for false call reduction, less programming workload or less variation in judgment
• Reducing false calls comes at the cost of a higher risk of escapes. Judge the PoC and acceptance on the reduction rate of good boards sent to manual review and an upper limit on the escape rate, as a pair
• Evaluate on data from periods after the training period, split into multiple periods. In research, models that passed in testing showed sharply more escapes in later periods
• Vendor figures differ in target components, defect modes and retraining assumptions. Read them together with their conditions and measure on your own images
• Record on each label which IPC class’s criteria, decided by whom, it was assigned under. IPC-A-610J is a collection of visual acceptability criteria and does not include X-ray or cross-section criteria
• Link AOI images, AI judgments, model versions, final human judgments and rework to board serial numbers. Set the retention period based on customer requirements and internal rules
• Hidden joints such as BGAs are the domain of AXI. Adding AI judgment does not change the division of roles between AOI and AXI
• In the estimate with hypothetical figures, the core of the benefit was adding AI re-judgment to existing AOIs (Configuration A). Count escape-related benefits only after the escape rate has been confirmed in the parallel run

You are welcome to consult TOMAS TECH even at the stage of taking stock of which line and which defect mode to start with, and what data can be obtained from your existing AOIs. Please contact us through this form.

References

• https://www.electronics.org/news-release/ipc-releases-j-revisions-two-leading-standards-electronics-assembly
• https://www.electronics.org/TOC/IPC-J-STD-001J_TOC.pdf
• https://www.electronics.org/TOC/IPC-A-610J_TOC.pdf
• https://www.electronics.org/TOC/IPC-7711D-7721D_TOC.pdf
• https://www.dinmedia.de/en/standard/ipc-j-std-001j/380272355
• https://www.technickenormy.cz/en/ipc-1782-revision-b/
• https://www.electronics.org/ipc-2591-connected-factory-exchange-cfx
• https://storage.mfn.se/667898b5-414f-4ce9-83c5-86a795e97ad6/press-release-deepreview-20240403.pdf
• https://mycronic.com/de/product-areas/pcb-assembly/webinars-and-events/webinars/fighting-aoi-false-calls
• https://kohyoung.com/en/koh-young-technology-showcasing-its-inspection-solutions-at-nepcon-asia-inshenzhen-on-november-6-8-2024/
• https://www.circuitsassembly.com/ca/editorial/menu-new-indust-prod/36396-omron-automation-americas-launches-vt-s10-3d-aoi.html
• https://inspection.omron.eu/en/products/vt-s1080
• https://circuitsassembly.com/ca/editorial/menu-new-indust-prod/36194-saki-develops-3di-ls2-case-3-d-aoi.html
• https://www.emsnow.com/?p=25838
• https://www.viscom.com/en/company/press/press-releases/detail/viscom-introduces-ai-powered-fast-program-generation-vai-provision-at-productronica-2025/
• https://www.viscom.com/fileadmin/content/downloads_brochures/en/Viscom_Brochure_vAI_en.pdf
• https://iconnect007.com/index.php/article/128868/mirtec-launches-ai-based-smart-factory-automation-solution-intelli-pro/128871
• https://electronicspecifier.com/events-news/tri-to-showcase-test-and-inspection-tech-at-ipc-apex-2025
• https://www.goepel.com/en/inspection-solutions/axi
• https://www.goepel.com/en/news/read/inspection-solutions-with-ipc-cfx-standard
• https://circuitsassembly.com/ca/current/28202-selective-meets-the-wave.html
• https://www.electronicdesign.com/home/article/21200236/how-to-select-an-aoi-system
• https://www.vse.com/blog/2021/07/07/identifying-pcb-fault-detection-using-image-processing/
• https://arxiv.org/html/2506.14521v1
• https://export.arxiv.org/abs/2505.03576
• https://cris.fau.de/publications/247698623/
• https://cris.fau.de/publications/338842958
• https://jsss.copernicus.org/articles/9/167/2020/
• https://pmc.ncbi.nlm.nih.gov/articles/PMC10673910/
• https://www.boi.go.th/upload/content/PR122_2569EN.pdf
• https://www.nst.com.my/amp/business/economy/2026/08/1503841/thailand-attracts-us268bil-chip-electronics-investment-bids-supply
• https://www.nationthailand.com/business/economy/40071468
• https://www.nationthailand.com/business/40070085
• https://www.nationthailand.com/business/40071775
• https://www.nationthailand.com/news/40071514
• https://www.nationthailand.com/business/automobile/40071986
• https://www.thailand-business-news.com/?p=332413
• https://macaubusiness.com/thailand-approves-1-99-billion-in-new-investment-led-by-ai-and-advanced-electronics/
• https://www.electronicdesign.com/resources/industry-insights/article/55401031/electronic-design-thailand-is-upskilling-its-workforce-to-meet-the-demands-of-an-expanding-tech-economy-part-1
• https://iconnect007.com/article/149833/tpca-and-industry-leaders-convene-in-bangkok-for-thailand-pcb-summit/149830/ein