AI Abnormal Noise Inspection: RFP and Acceptance Test Guide

“Our pre-shipment noise inspection relies on veteran inspectors listening through headphones and deciding pass or fail. Judgments differ from person to person, and there is a limit to how long we can make people listen on a noisy line all day. Every time a customer files a noise complaint, we write another 8D report.” We increasingly hear this kind of concern from quality managers at Japanese-owned motor and electric unit plants in Thailand. That is where AI abnormal noise inspection comes into the discussion.

Here is the conclusion up front. What decides the success of an AI abnormal noise inspection project is not how clever the AI model is. It is “sound that can be recorded under the same conditions every time” and “ground-truth data on which human judgments agree.” Before AI replaces the human ear, the shortest path is to bring it in first as a tool for measuring how much human judgments vary. The decisions you need to make come down to three: (1) fix how sound is captured first, (2) have people align on the correct answers, and (3) decide pass or fail through acceptance testing.

Note that all production volumes, headcounts, counts, monetary amounts and payback periods for the factory in this article are original estimates and assumed values (placeholder values) created for this article, based on the model factory described later. They are neither industry averages nor survey figures. No primary vendor pricing or performance data is used either. Please read them as a “calculation template” to be replaced with your own actual figures and quotations.

Why AI Abnormal Noise Inspection Is Now on the Table for Thai Factories

More options claim to automate listening tests

In pre-shipment inspection of products that incorporate motors, gears, pumps or fans, many processes still rely on people listening and deciding pass or fail. Over the past few years, there has been a steady stream of product and research announcements aimed at automating this “auditory sensory inspection.” Below are examples of options on the market that we were able to confirm. This is neither a recommendation nor a price comparison. All performance figures shown are vendor-side announcements, and they do not mean the same values will be achieved on your own line.

There is activity on the research side as well. In August 2026, a study was posted to arXiv that performs anomaly detection on EOL vibration data from EV e-transaxle assemblies, split by operating stage. It reports that narrowing an anomaly down to a specific operating condition helps root cause analysis. However, this is a preprint that has not yet been peer-reviewed, and the abstract contains no specific accuracy figures.

The key point here is that these figures and descriptions cannot be compared on the same yardstick. “As short as 3 seconds” is the minimum judgment time, and “above 95%” is a detection rate whose conditions are not disclosed. None of them answers “for this part number of ours, how many units will be missed and how many will be falsely rejected?” That is exactly why you need to build your own yardstick for comparison. That yardstick is the evaluation set and acceptance testing explained later.

Thai auto production shows signs of recovery, and requirements are getting stricter

According to the Federation of Thai Industries (FTI), Thailand’s automobile production in August 2026 was 124,646 units, up 10.93% year on year. Domestic sales were up 25.59% year on year, and exports were down 2.04% year on year. July production was also 117,383 units, up 6.12% year on year, reversing June’s 7.55% year-on-year decline. On the other hand, the FTI forecasts full-year 2026 production to fall 3.33%, citing lower exports due to the conflict in the Middle East.

When volumes come back, staffing of inspectors can fail to keep up. In addition, with the shift to electrification, requirements for the operating sound of components are becoming stricter. A technical paper by AISIN explains the background: vehicle electrification has made cabins quieter, raising requirements for component operating sounds.

The very practice of “making people listen” on a noisy line is being questioned

Another reason is inspector health. Thailand has a ministerial regulation and a notification governing noise exposure, with a standard that the 8-hour average exposure must be 85 dB(A) or less (details later). The practice of having people wear headphones and keep listening for long hours in the middle of a noisy line is worth reviewing from both a quality standpoint and an occupational safety and health standpoint.

What Is AI Abnormal Noise Inspection? A Mechanism That Learns Normal Sounds to Find “Sounds That Differ from Usual”

The premise: you cannot collect sounds from defective products

The conceptual foundation of AI abnormal noise inspection lies in Task 2 of “DCASE,” an acoustic research competition. In 2020, Task 2, “Unsupervised Detection of Anomalous Sounds for Machine Condition Monitoring,” provided only normal sounds as training data. The reason is that actual abnormal sounds “rarely occur and are highly diverse.” That year saw 117 submissions from 40 teams.

Factory shipment inspection is the same. There are many types of defects, such as bearing damage, gear dents, foreign object entrapment and brush squeal, but you can hardly obtain defective units for each type. So the AI mainly learns a large number of normal product sounds and quantifies how far a sound deviates from “the usual sound” (the anomaly score). If the anomaly score exceeds a threshold, the unit is judged NG. The fact that the AI is not taught each defective sound one by one is different from image-based visual inspection, where images of defective products are collected for training. The image-side approach is covered in “A Guide to Implementing AI Visual Inspection.”

Research datasets and what they contain

Research in this field has used public datasets.

• MIMII Dataset: A dataset of industrial machine sounds published by researchers at Hitachi, Ltd. It covers four machine types (valves, pumps, fans and slide rails), with 5000 to 10000 seconds of normal sound and about 1000 seconds of abnormal sound per model. Background noise recorded in real factories is mixed in, with three SNR conditions of -6 dB, 0 dB and 6 dB. In other words, “distinguishing abnormal sounds amid factory noise” is itself a research problem.
• ToyADMOS: A dataset built by deliberately damaging miniature machines to collect abnormal sounds, motivated by the problem that the difficulty of collecting abnormal sound data has been a barrier to building large-scale datasets. It contains over 180 hours of normal operating sound and over 4,000 abnormal sound samples.
• ToyADMOS2: A dataset for evaluating anomalous sound detection under “domain shift” conditions caused by differences in operating speed, microphone placement and machine model. It contains over 27,000 normal sound samples and over 8,000 abnormal ones.

Research focuses on “when conditions change” and “machines seen for the first time”

DCASE Task 2 has added conditions closer to factory reality year after year. In 2023, the “first-shot” problem setting was introduced for the first time. In 2024, systems were required to work both when attribute information such as operating conditions is available and when it is intentionally hidden for some machine types.

The 2025 Task 2 explicitly assumes that, because anomalies in factories are rare and diverse, models are trained using normal sounds only. For each machine type, the development data consists of only 990 normal clips under the original conditions (source domain) and 10 normal clips after the conditions change (target domain). Domain shift is described as arising from differences in operating speed, noise type, machine load, microphone placement and so on. “First-shot” refers to a setting in which a completely new machine type may be encountered, so parameters cannot be tuned by looking at test data. The organizers include researchers from Hitachi, Ltd., NTT and others.

Translated to a factory, this means the AI’s judgments can break down as soon as “the part number changes,” “the microphone is replaced” or “the neighboring equipment starts running.” And you will not have the capacity to re-collect large amounts of normal sound each time. The research challenges are, as they stand, the operational challenges after deployment.

Research metrics are not factory pass/fail criteria

DCASE evaluation metrics are AUC and pAUC (AUC over the range where the false positive rate is 0.1 or less), and since 2023 the official score has been their harmonic mean. These are research competition metrics for comparing methods with one another. What a factory wants to know is “if we set the threshold here, how many known defective units will be missed and how many good units will be rejected,” and a high AUC does not answer that. When placing an order, decide pass or fail not by AUC but by the number of misses and the false rejection rate counted on your own evaluation set.

What AI can and cannot do

Limits of Auditory Sensory Inspection, and “Measuring Judgment Agreement” Before AI

Problems with inspection that relies on skilled ears

AISIN’s technical paper “Development of a New Inspection Method for Product Operating Sound” states that auditory sensory inspection by skilled inspectors is widely used for operating sound inspection of automotive parts, and that quantitative evaluation and automation are needed because of man-hours, the burden on inspectors and the problem of passing on skills. The paper classifies abnormal sounds into 7 types and reports a method that detects “abnormal sounds whose operating-sound frequency changes,” which are difficult to detect with conventional methods, using signal processing called cepstrum analysis and time-frequency ridge analysis. Note that it does not describe detection using machine-learning AI. Even so, the stance of first deciding “which sound to capture and how to define it” is directly applicable to AI implementation.

For methods of human sensory evaluation, there is JIS Z 9080:2004, “Sensory analysis — Methodology.” It specifies methods such as the paired comparison test, the triangle test and the ranking method, and the 2004 revision aligned it internationally on the basis of ISO standards. However, this is a standard on “methods of evaluation” written mainly with sensory evaluation of foods and similar products in mind; it does not define pass/fail criteria for abnormal noise inspection.

Criteria on which humans do not agree cannot be reproduced by AI either

This is the point this article most wants to convey. When training or evaluating AI, people assign the ground-truth labels: “this unit is OK, this unit is NG.” But if those human judgments differ between inspectors, or even differ for the same person from day to day, the AI cannot tell which to follow. Measuring AI accuracy while the ground truth is inconsistent produces a meaningless number.

Also, we could not confirm any reliable figure that can be cited as an industry benchmark for the miss rate of human auditory inspection or the agreement rate between inspectors. That is exactly why you have no choice but to measure your own figures yourself.

Measuring current inspection with attribute agreement analysis (kappa)

A method you can use is “attribute agreement analysis,” which is used in quality management. For judgments such as OK/NG (attribute data), it evaluates the following three things:

1. Whether the same inspector gives the same answer when judging the same unit several times (within-appraiser consistency)
2. Whether answers agree between inspectors (between-appraiser agreement)
3. Whether answers match the standard (the correct answer) (agreement with the standard)

The degree of agreement is expressed by the kappa coefficient. Kappa takes values from -1 to +1, where 1 is perfect agreement and 0 is agreement no better than chance. Minitab’s statistical software help explains that AIAG (a US automotive industry organization) states that “kappa values of 0.75 or greater indicate good agreement, and larger values such as 0.90 are preferred.” The same help also cautions that consistent judgments are not necessarily correct judgments.

IATF 16949 requires measurement system analysis (MSA) for the inspection, measurement and test systems listed in the control plan. This article’s view is that if abnormal noise inspection by the human ear is an inspection in the control plan, it can also be a subject of attribute MSA. In other words, preparation for AI implementation starts with the MSA of auditory inspection that should have been done in the first place. Specifically, the steps are as follows:

• Prepare a sample that mixes normal units, boundary samples (units at the borderline between pass and fail) and known defective units
• Have 3 inspectors listen 2 to 3 times each in a changed order, and record OK/NG
• Calculate within-appraiser, between-appraiser and against-standard agreement with kappa; have everyone re-listen to the units with mismatches, and redefine the criteria with words and samples

This work is what finally assembles the ground-truth data for “what the AI learns and what it is evaluated against.” The procedure for creating ground-truth labels in image inspection is explained in “Annotation Workflow for Visual Inspection,” and the same thinking applies to sound.

Sound Capture Is 90% of the Work: Inspection Booth, Fixtures, Microphones and Operating Conditions

Much of the accuracy of AI abnormal noise inspection is determined by “how you record” before you choose a model. Because the AI looks for deviations from normal sounds, it will judge something as “different from usual” in the same way even when the cause of the deviation is not the product but ambient noise or a difference in how the unit is placed.

Inspection booth: keep factory noise out

Factory noise, such as the press next door, air blowing, conveyors and forklift warning sounds, can be much louder than a product’s small abnormal sounds. The basic approach is to place the product in a sound-insulated, vibration-isolated inspection booth so that outside sound does not enter the recording. Also check the opening and closing of the door and vibration transmitted through the booth floor. Some vendor products take the approach of suppressing the influence of ambient noise with a contact-type microphone. Whichever approach you take, always verify in acceptance testing “how the judgment moves when noise conditions are changed.”

Fixtures: place the unit the same way every time

The sound changes depending on where, in which orientation and with how much force the product is held. Use fixtures to standardize position and clamping force so that the same state is recorded even when operators change. Since fixture wear and rubber degradation can also change the sound, decide the inspection interval as well.

Microphones and sensors: fix position and type, and calibrate

Fix the type of microphone (a microphone that picks up airborne sound, a contact type, or a vibration sensor), its distance from the product and its orientation. Assume that replacing a microphone will change the sound characteristics, and decide the verification procedure after replacement. For how to choose sensors, “How to Select Anomaly Detection Sensors (Vibration, Sound, Current)” is also a useful reference.

Operating conditions: decide speed, load and direction of rotation

Even for the same motor, the sound changes greatly with rotational speed, load, and forward or reverse rotation. Decide the operating pattern for inspection (for example, ramp-up, constant speed, deceleration, reverse) and run the same pattern every time. Looking at each stage separately shows under which condition the anomaly appeared, which also helps with cause analysis. This idea overlaps with the e-transaxle study introduced earlier. For products such as speed reducers, where noise from gear meshing is the issue, the structural aspects are covered in “A Guide to Selecting Speed Reducers.”

Write recording conditions down as a “specification”

The conditions described so far will not last as verbal agreements. Write the booth specification, fixture drawings, microphone model and position, operating pattern and sampling conditions into both the work standards and the purchase specification. Of the 12 RFP items discussed later, 3 relate to this “sound capture.”

Designing the Judgment: AI First-Pass Judgment and Human Re-Judgment, and How to Think About Misses and False Rejections

Count misses and false rejections separately

There are two kinds of errors in abnormal noise inspection: “misses,” where a defective unit is passed as OK, and “false rejections,” where a good unit is judged NG. Misses lead to customer complaints and field failures, while false rejections add the work of re-inspection and teardown checks. Lowering the AI’s anomaly score threshold reduces misses but increases false rejections. Raising the threshold does the opposite. A single number like “accuracy X%” does not show this trade-off, so always count the two separately.

Think of the AI’s role in two stages

Deployment can be broadly divided into two forms.

• Configuration A (double judgment): Keep human auditory inspection as it is and add AI alongside it to prevent misses. Even if a person passes a unit as OK, it is stopped if the AI says NG
• Configuration B (first-pass judgment): The AI performs a first-pass judgment on every unit, and people re-judge only the units the AI marked NG. Because people no longer listen to every unit, Configuration B on its own carries a higher risk of misses than Configuration A. That is why passing the evaluation set is a prerequisite

Configuration B can greatly reduce human man-hours, but AI misses translate directly into outflow. So moving to Configuration B comes only after passing the evaluation set and confirming a track record by running Configuration A for a while. Removing people abruptly before passing may increase misses.

Always keep a human re-judgment route

In either configuration, provide a “re-judgment station” and screen where people can re-listen to units the AI marked NG and decide while looking at the waveform and spectrum. When a person overrides the AI’s judgment, record the reason and use it for retraining and threshold review. Without these records, the shop floor will stop trusting the AI’s judgments and will eventually ignore them.

Keep waveforms and judgments linked to serial numbers

When a customer files a noise complaint, being able to retrieve that product’s sound and judgment at the time of shipment greatly changes the speed of the investigation. Store the waveform, anomaly score, judgment and re-judgment result for each serial number, and connect them to MES and quality data systems. For storage design, also see “How to Choose a Quality Data Management System.” Its position as the final gate before shipment is covered in “Preventing Defect Outflow and the Shipping Gate.”

Cost and ROI of AI Abnormal Noise Inspection: A Model Estimate

From here on is an estimate for a model factory. All figures below are original estimates and assumed values (placeholder values) created for this article; they are neither industry averages nor survey figures. No primary vendor pricing or performance data is used.

Common assumptions (Model Factory N)

We assume a Japanese-owned plant making small automotive motors in eastern Thailand (Chonburi Province).

This starting point is common to both Configuration A and Configuration B.

Configuration A: Add AI alongside, keep human auditory inspection

Annual net benefit is 900,000 − 150,000 = 750,000 THB/year. The simple payback period is 2,000,000 ÷ 750,000 = 2.666…, or about 2.7 years.

Configuration B: Configuration A + AI first-pass judgment, with people re-judging only units the AI marked NG

Annual net benefit is 2,334,000 − 200,000 = 2,134,000 THB/year. The simple payback period is 2,600,000 ÷ 2,134,000 = 1.218…, or about 1.2 years.

Note that Configuration B’s benefits replace Configuration A’s benefits; they are not added on top of A. The 75% complaint reduction includes Configuration A’s 50%. Adding Configuration A’s 900,000 and Configuration B’s 1,350,000 would be double counting. Also, because people continue to listen to every unit in Configuration A, the labor cost and false rejection benefits are treated as 0 there.

Sensitivity: what if the redeployment does not happen?

Of Configuration B’s benefits, the 864,000 in labor cost is realized only if “4 people could be redeployed to other processes.” If there is nowhere to redeploy them and the labor cost benefit is set to zero, the result is as follows:

• Annual net benefit: 1,350,000 + 120,000 − 200,000 = 1,270,000 THB/year
• Simple payback period: 2,600,000 ÷ 1,270,000 = 2.047… → about 2.0 years

Even without the labor cost benefit, Configuration B’s payback period (about 2.0 years) is shorter than Configuration A’s (about 2.7 years).

Takeaway 1: The add-on pays back quickly, but check what the benefit is made of

Compare the 5-year cumulative totals.

• Configuration A: 750,000 × 5 − 2,000,000 = 3,750,000 − 2,000,000 = 1,750,000 THB
• Configuration B: 2,134,000 × 5 − 2,600,000 = 10,670,000 − 2,600,000 = 8,070,000 THB

Looking only at the Configuration B add-on, the additional initial investment is 600,000 THB, and the additional annual net benefit is 2,134,000 − 750,000 = 1,384,000 THB/year. The payback period for the add-on is 600,000 ÷ 1,384,000 = 0.433…, or about 0.43 years (about 5 months). The difference in the 5-year cumulative total is 1,384,000 × 5 − 600,000 = 6,320,000 THB, which matches 8,070,000 − 1,750,000.

What the add-on 600,000 buys is the re-judgment station, storage of waveforms and judgments, and evaluation of judgment agreement between 3 inspectors and the AI. In other words, it is an investment in “a system that aligns judgments, records them and lets people re-judge.”

However, looking at the breakdown of the additional annual net benefit of 1,384,000 THB, it consists of the extra complaint reduction of 1,350,000 − 900,000 = 450,000, inspector redeployment of 864,000 and halving of false rejections of 120,000, minus the increase in operating cost of 200,000 − 150,000 = 50,000 (450,000 + 864,000 + 120,000 − 50,000 = 1,384,000). A little over 60% of the add-on’s benefit comes from redeployment. If redeployment does not happen, the additional annual net benefit is 1,270,000 − 750,000 = 520,000 THB/year, and the add-on payback is 600,000 ÷ 520,000 = 1.15…, or about 1.2 years.

There is one more caution. In Configuration B, people no longer listen to every unit, so looking at the form of operation alone, there is no reason for misses to be fewer than in Configuration A. This estimate sets the complaint reduction at 75% on the assumption that the following accumulate: redefinition of judgment criteria using the results of judgment agreement measurement, retraining based on re-judgment records, and cause tracing through stored waveforms and judgments per serial number. Removing people from 100% inspection does not in itself reduce misses. Moreover, Configuration B presupposes an operation in which people re-judge the AI’s NG judgments, and passing the evaluation set. If people are removed before passing, this estimate does not hold.

Takeaway 2: What if you skip the inspection booth (Configuration A′)?

Wanting to lower the initial cost, people sometimes propose skipping the inspection booth. Removing the booth (400,000 per line) from Configuration A makes the initial cost 2,000,000 − 400,000 × 2 = 2,000,000 − 800,000 = 1,200,000 THB.

Here we assume a situation in which noise from the neighboring press, air jets and conveyors enters the recordings, and the AI issues frequent NG judgments. The assumption is that the shop floor starts ignoring the AI’s judgments and the complaint reduction benefit becomes 0. The annual balance is then 0 − 150,000 = −150,000 THB/year, and the investment cannot be paid back.

As a result of cutting 800,000, the annual net benefit falls from 750,000 to −150,000. Moreover, this estimate does not include the re-inspection cost of units stopped by erroneous AI NG judgments. The actual balance would be worse than this. Incidentally, keeping false rejections at 1.0% in Configuration A is also because the re-inspection cost from AI NG judgments is placed outside the model. Unless sound can be recorded under the same conditions, the result is the same whichever AI you choose.

Summary of the estimate

When calculating for your own plant, replace at least these four with actual figures: “production volume,” “number of inspectors and labor cost,” “number of noise complaints and cost per complaint,” and “false rejection rate and re-inspection cost.” For the cost per complaint, aggregating sorting labor, 8D response man-hours, expedited shipping costs and so on separately makes the discussion easier to align.

12 Items to Include in an RFP for AI Abnormal Noise Inspection

When asking vendors for quotations, simply saying “we want to implement AI abnormal noise inspection” will produce proposals with differing scopes that cannot be compared. Write the following 12 items as specifications to align proposals under the same conditions. The overall process for ordering inspection equipment is also covered in “Automating Inspection Equipment in Thailand: Ordering and Acceptance.”

Items 5 and 9 in particular are prone to later disputes. Check that the arrangement does not end up as: your company collected the normal sounds, yet the trained model belongs to the vendor, and a fee is charged each time a part number is added. Also, whether judgment is completed on the edge PC or processed on a higher-level server relates to item 10. For how to place this on the shop floor, also see “Deploying Edge AI in Factories.”

What to Check in FAT/SAT for AI Abnormal Noise Inspection

Build the evaluation set first

Before acceptance testing, prepare a set of products to use for evaluation. The quantities are only target examples, not standard values.

• 300 normal units
• 30 boundary samples (units at the borderline between pass and fail)
• 30 known abnormal-noise defective units (bearing damage, gear dents, foreign object entrapment, brush squeal, etc.; actual units or returned units)

Note that there are limits to what an evaluation set of this size can tell you. With 300 normal units, 1 unit corresponds to about 0.33%, so “false rejection rate of 0.5% or less” effectively means “no more than 1 NG unit.” Even with zero misses on 30 defective units, the range that can be statistically guaranteed is narrow. Treat the evaluation set as the minimum bar for pass/fail, and keep verifying with actual performance after go-live.

Run this set in the same way in both FAT (acceptance test at the vendor’s factory) and SAT (acceptance test on your own line). A system that passes FAT can break down in SAT, where your own line’s noise, power supply and operator handling are added.

Metrics to look at

Handling cases with few abnormal-noise defective units

Even if training can start mainly with normal sounds, acceptance testing requires known defective units, because “not missing” cannot be verified without actually running defective units. Supplement the evaluation set by, for example, keeping returned units and past complaint units, or building evaluation products that incorporate intentionally damaged parts. Record in the evaluation records that artificially created defects may sound different from defects that actually occur in the field. Approaches to compensating for a lack of defect data are also covered in “Using Synthetic Data for Manufacturing AI.”

Use the same set to watch for degradation after go-live

Even after passing acceptance, judgments gradually drift due to fixture wear, microphone degradation, the addition of part numbers and so on. After go-live, run the evaluation set periodically and check whether the detection rate and false rejection rate have changed since acceptance. For criteria to decide when to move from trial operation to production, also refer to “Exit Criteria for AI PoCs.”

Issues Specific to Thailand and ASEAN

1. Noise and inspector health

The Thai Ministry of Labour’s “Ministerial Regulation Prescribing Standards for Administration, Management and Operation of Occupational Safety, Health and Environment Concerning Heat, Light and Noise B.E.2559 (2016)” prohibits exposure to impulse noise with a peak sound pressure level above 140 dB and to continuous steady noise above 115 dB(A), and requires the average over the daily working time to be managed at or below the standard in the Director-General’s notification. It also stipulates that workplaces with an 8-hour average exposure of 85 dB(A) or more must take hearing protection measures in accordance with the notification. The Department of Labour Protection and Welfare’s notification B.E.2561 (2018) then sets the average noise level over an 8-hour working period at 85 dB(A) or less.

The practice of “making people listen” for long hours on a noisy line is itself a reason for review. The inspection booth is meaningful not only for recording quality but also for the inspectors’ working environment. The English rendering of the regulations here is a summary. Please confirm individually with the competent authorities or experts how they apply to your own workplace.

2. The language of judgment criteria

Japanese onomatopoeia such as “kiin,” “goro-goro” and “shaa” mean nothing to Thai inspectors. Prepare explanations of boundary samples and work standards in Thai, and define abnormal sounds by waveform, frequency band and physical samples. AI implementation can also be an opportunity to advance this work of “defining with samples and waveforms rather than words.”

3. Auto production trends and customer audits

As mentioned earlier, Thailand’s automobile production turned to year-on-year growth in July 2026 and grew again in August, but the FTI’s full-year forecast is a 3.33% decline. The harder volumes are to predict, the more difficult it becomes to respond by increasing or decreasing inspectors. As requirements for operating sound rise with electrification, customer audits are likely to ask more often about the criteria for abnormal noise judgment and how variation in those judgments is managed. Records of attribute agreement analysis can also serve as material for that explanation.

4. BOI special incentives for automation and robotics

Thailand Board of Investment (BOI) Announcement No. 4/2569 (published in the Royal Gazette on March 31, 2026) provides, to promote advanced automotive manufacturing technology, exemption from import duties on machinery and a 3-year corporate income tax exemption capped at 50% of the investment amount (excluding land and working capital) for investments in automation and robotic systems. This is expanded to 100% if 30% or more of the total machinery value relates to or supports the domestic automation machinery industry. The minimum investment is 1 million baht, and the application deadline is the end of 2027.

However, eligibility is limited to Category 3.6 (general automotive manufacturing) and 3.8 (PHEV and HEV manufacturing). It does not apply to plants for home appliances or general machinery. Whether abnormal noise inspection equipment qualifies for this incentive needs to be confirmed individually with the BOI or experts.

5. PDPA and the draft AI law

The microphones in an inspection booth may pick up inspectors’ voices and conversations. Thailand’s Personal Data Protection Committee (PDPC) has begun a public consultation on draft sector-specific guidelines under the Personal Data Protection Act (PDPA, 2019), but we have not been able to confirm any provisions specifically addressing audio recording in manufacturing or employee monitoring. Please confirm individually with experts how to handle the scope of recording, notification, retention period, access restrictions and so on.

In addition, Thailand’s Electronic Transactions Development Agency (ETDA) published a draft AI law in July 2026 and held a public consultation. It is a risk-based regulation partly modeled on the EU AI Act, and it has been said that enactment and enforcement could take 6 to 24 months. As of October 2026, it is at the draft bill stage and has not been enacted. Please check with the competent authorities or experts on future developments.

6. Maintenance structure

Decide who within Thailand will check microphones, inspect booth doors and vibration-isolation materials for degradation, and retrain when part numbers are added. If the vendor’s engineers are only in Japan or Europe, there will be waiting time every time a part number is changed over. It is important to draw the line at the RFP stage between how much your own Thai engineers will handle and where you will rely on the vendor.

A 90-Day Plan for Implementing AI Abnormal Noise Inspection

The most important thing in the first 30 days is to look at the variation in human judgment as numbers before touching AI. If agreement is low here, aligning criteria and samples comes before AI. Conversely, if the criteria align here, you will already have in hand ground-truth data usable for AI training and evaluation. The difference between equipment-side anomaly detection (monitoring equipment in operation) and shipment inspection is explained in “How to Run an Equipment Anomaly Detection PoC,” and the approach to automating 100% inspection itself in “Automating 100% Inspection.”

Frequently Asked Questions (FAQ)

Can abnormal noise inspection be fully automated with AI?

We do not recommend removing people completely right away. First, run the AI alongside human auditory inspection (Configuration A in this article) and build a track record as a double judgment to prevent misses. Even after passing the evaluation set and moving to an operation in which people re-judge units the AI marked NG (Configuration B), keep the re-judgment route. Human judgment is also needed to respond to new defect modes and part number changeovers.

Can AI be built even with almost no defective noise samples?

Training can start mainly with normal sounds. In the DCASE research competition as well, tasks are set on the premise of training with normal sounds only. However, acceptance testing requires known defective units. “Not missing” cannot be verified without actually running defective units. Keep returned units and past complaint units, and build an evaluation set.

Can AI abnormal noise inspection be used on a noisy factory line?

If factory noise enters the recordings, the AI picks it up as a “sound different from usual,” and misjudgments increase. The basic approach is to use a sound-insulated, vibration-isolated inspection booth and fixtures so that recordings are made under the same conditions every time. Some products use a contact-type microphone to suppress the influence of ambient noise, but whichever method you use, carry out tests with changed noise conditions at acceptance, such as opening the door or running neighboring equipment.

What are typical implementation costs and payback periods for AI abnormal noise inspection?

In this article’s model estimate, Configuration A came out at an initial 2,000,000 THB with about 2.7 years, and Configuration B at an initial 2,600,000 THB with about 1.2 years (about 2.0 years if inspector redeployment does not happen). However, these are original placeholder values for this article, not industry averages or quotations. Replace production volume, number of inspectors and labor cost, number and cost of noise complaints, and false rejection rate with your own actual figures and calculate.

What should be checked in the RFP and acceptance testing (FAT/SAT) for AI abnormal noise inspection?

In the RFP, write the 12 items: target part numbers and defect modes, operating conditions and takt time, scope of booth and fixtures, microphone calibration and replacement procedures, ownership of training data, storage of waveforms and judgments, human re-judgment route, evaluation set and acceptance criteria, retraining procedure and cost, security, Thai-language support, and local maintenance. In FAT/SAT, run the same evaluation set, count misses and false rejections separately, and check repeatability, agreement with humans and behavior when noise conditions are changed.

What should we watch out for regarding regulations and the BOI when implementing at a Thai plant?

Issues include the noise exposure standard (8-hour average of 85 dB(A)), the BOI special incentives for automation and robotics (limited to the automotive sector), handling under the Personal Data Protection Act (PDPA) when inspectors’ voices are captured in recordings, and the AI law, which is still at the draft bill stage. In every case, the explanations in this article are summaries or general discussion. Please confirm individually with the competent authorities or experts how they apply to your company.

Summary

• What decides the success of AI abnormal noise inspection is not how clever the AI model is, but “sound that can be recorded under the same conditions” and “ground-truth data on which human judgments agree”
• AI abnormal noise inspection mainly learns normal sounds to find “sounds that differ from usual.” Research competition metrics (AUC/pAUC) are a different thing from factory pass/fail criteria
• Before AI, measure the judgment agreement of the current auditory inspection with attribute agreement analysis (kappa). Criteria on which humans do not agree cannot be reproduced by AI either
• Lock down the inspection booth, fixtures, microphones and operating conditions as specifications. In the model estimate, skipping the booth dropped the annual net benefit from 750,000 THB to −150,000 THB, and the investment could no longer be paid back
• Start with Configuration A (double judgment), and move to Configuration B (AI first-pass judgment + human re-judgment) only after passing the evaluation set. In the model estimate, the 600,000 THB add-on was calculated to pay back in about 5 months. However, much of that comes from inspector redeployment, and if redeployment does not happen, the add-on payback becomes about 1.2 years (all placeholder values)
• Align proposals with the 12-item RFP, and in FAT/SAT count misses and false rejections separately on the same evaluation set
• Proceed on Thailand-specific issues, such as noise regulations, Thai-language judgment criteria, the BOI, the PDPA and draft AI law, and local maintenance, while confirming individually with the competent authorities and experts

It is perfectly fine if you are still at the stage of wanting to measure the judgment agreement of your current auditory inspection, or of wanting to start by organizing defect samples and returned units. If you are considering AI abnormal noise inspection at a plant in Thailand, we can help you sort things out from the very first step, so please feel free to reach out via our contact page.

References

• NTT DATA CCS “Monone” article (EE Times Japan, July 27, 2026): https://eetimes.itmedia.co.jp/ee/articles/2607/27/news063.html
• Hmcomm “FAST-D” official site: https://fast-d.hmcom.co.jp/
• Hmcomm company history (EDINET DB): https://edinetdb.jp/company/E40058/history
• Hitachi, Ltd. news release (October 22, 2020): https://www.hitachi.com/ja-jp/press/articles/2020/10/1022/
• HBK integrates torque and NVH testing (Automotive Powertrain Technology International, April 21, 2026): https://www.automotivepowertraintechnologyinternational.com/news/testing/hbk-integrates-torque-and-nvh-testing-to-enhance-end-of-line-automotive-quality-control.html
• HEAD acoustics press release (November 6, 2024): https://www.head-acoustics.com/news-events/press-releases/press-details/efficient-end-of-line-testing-with-head-acoustics-quality-assurance-at-the-highest-level/
• FPT US patent for acoustic anomaly detection technology (Viet Nam News / Bizhub, November 10, 2025): https://bizhub.vietnamnews.vn/fpt-secures-us-patent-for-breakthrough-neural-network-technology-in-acoustic-anomaly-detection-post390252.html
• AISIN “Development of a New Inspection Method for Product Operating Sound” (AISIN TECHNICAL REVIEW VOL.28, in Japanese): https://www.aisin.com/jp/technology/technicalreview/28/pdf/10.pdf
• Stagewise Anomaly Detection for E-Transaxle Quality Monitoring (arXiv, preprint not yet peer-reviewed, August 27, 2026): https://arxiv.org/abs/2609.22172
• DCASE 2020 Task 2: https://dcase.community/challenge2020/task-unsupervised-detection-of-anomalous-sounds
• DCASE 2023 Task 2: https://dcase.community/challenge2023/task-first-shot-unsupervised-anomalous-sound-detection-for-machine-condition-monitoring
• DCASE 2024 Task 2: https://dcase.community/challenge2024/task-first-shot-unsupervised-anomalous-sound-detection-for-machine-condition-monitoring
• DCASE 2025 Task 2: https://dcase.community/challenge2025/task-first-shot-unsupervised-anomalous-sound-detection-for-machine-condition-monitoring
• MIMII Dataset (Zenodo): https://zenodo.org/records/3384388
• ToyADMOS (arXiv): https://arxiv.org/abs/1908.03299
• ToyADMOS2 (arXiv): https://arxiv.org/abs/2106.02369
• JIS Z 9080:2004 Sensory analysis — Methodology (Japanese Standards Association preview, in Japanese): https://webdesk.jsa.or.jp/preview/pre_jis_z_09080_000_000_2004_j_ed10_ch.pdf
• Minitab, Interpret the key results for Attribute Agreement Analysis: https://support.minitab.com/en-us/minitab/help-and-how-to/quality-and-process-improvement/measurement-system-analysis/how-to/attribute-agreement-analysis/attribute-agreement-analysis/interpret-the-results/key-results/
• Copy of the Royal Gazette text of Thai Ministerial Regulation B.E.2559 (heat, light and noise; in Thai): https://ams.medsci.nu.ac.th/wp-content/uploads/2022/07/LAW2022/S14-%20กฎกระทรวงกำหนดมาตรฐานในการบริหาร%20จัดการ%20แสงสว่าง%20เสียง%202559.pdf
• Academic article on Thai noise standards (Journal of SciTech-ASEAN): https://li01.tci-thaijo.org/index.php/STJS/article/view/260269
• Overview of Thai occupational safety and health regulations (Enviliance ASIA): https://enviliance.com/regions/southeast-asia/th/th-osh
• Thailand car production, August 2026 (Capital.com): https://capital.com/en-gb/news/thailand-s-august-car-production-rises-10-9-on-year
• Thailand car production, July 2026 (New Straits Times): https://www.nst.com.my/amp/business/economy/2026/08/1518403/thailand-car-production-rises-61pct-year-july
• New BOI incentives for automotive and HEV/PHEV manufacturing (Tilleke & Gibbins): https://www.tilleke.com/insights/thailand-unveils-new-incentives-for-automotive-and-hev-phev-manufacturing/8/
• Thailand’s proposed AI law (Rajah & Tann Asia): https://www.rajahtannasia.com/viewpoints/thailands-proposed-ai-law-moving-towards-a-risk-based-regulatory-framework/
• Thailand’s draft AI law (Tilleke & Gibbins): https://www.tilleke.com/insights/thailand-releases-new-draft-artificial-intelligence-act/
• Public consultation on sector-specific PDPA guidelines (HLC): https://hlc.com/en/publications/thailand-pdpc-consults-on-sectorspecific-pdpa-guidelines