Analog Gauge Reading AI Implementation: Automating Patrol Inspections in Thai Factories

Analog gauge reading AI implementation, which reads gauge values from camera images, is the option that comes up when maintenance section managers and utility staff (boilers, compressors, cooling water, water receiving tanks) at Japanese-owned factories in Thailand bring us this problem: “Every shift, someone walks the rounds, reads the pressure gauges, thermometers and level gauges by eye, writes the values on paper, and later transcribes them into Excel. On night shift, readings get missed or transcribed wrongly, and we notice abnormalities late.”

Here is the conclusion up front. Whether analog gauge reading AI pays back is not decided by “how much patrol labor cost goes down.” It is decided by “how many minutes apart you can now look at the gauges you cannot afford to have fail, and how much earlier you notice an abnormality.” That is why narrowing the scope to the points you cannot afford to have stop pays back faster than putting a camera on every gauge in the plant (the model calculation later in this article also shows that cutting patrols alone does not pay back).

There are five things to decide in an implementation: (1) which gauges to target, (2) the method, (3) how to create the correct reading (ground truth), (4) abnormality judgment and notification, and (5) how inspection records and statutory requirements are handled. Deciding them in this order keeps product selection, the RFP and the acceptance tests from drifting.

Note that all gauge counts, patrol times, labor costs, costs, downtime losses and payback periods for the factory in this article are this article’s own estimates and assumed (placeholder) values based on the model factory described below. They are neither industry averages nor survey figures. No vendor prices are used in the calculation either. Read it as a “calculation template” to be replaced with your own actual figures and quotes.

Why Analog Gauge Reading AI Implementation Is on the Table for Thai Factories Now

Patrols and transcription always leave “time when no one is looking”

In many factories, patrol inspection of analog gauges is still done by people. A maintenance technician walks the round with an inspection sheet, reads the pressure gauge needle, checks the level gauge position, writes on paper, and transcribes into Excel after returning to the office. We covered digitizing the inspection sheet itself in inspection checklist digitalization, but even after digitization, the fact that “you do not know the value unless someone goes to look” does not change.

The weakness of this approach is less the effort than the “time when no one is looking.” With one patrol per shift, a gauge goes unseen by anyone for several hours until the next round. If cooling water pressure is slowly dropping, or air pressure falls below its setting, no one knows until it is noticed on the next round. During night shifts with fewer people, the patrol itself can get pushed back. On top of that, if a digit is miscopied or a row is shifted when transcribing from paper to Excel, a value that was read correctly breaks down at the recording stage.

What analog gauge reading AI can change is exactly this “time when no one is looking.” A camera photographs the gauge at fixed intervals, AI reads the value, and a notification goes out when a threshold is exceeded. Depending on the method, this means the gauges you cannot afford to have fail can be checked every few minutes. That, rather than reducing the number of human patrols, is the core of the benefit.

Meter image recognition announcements continue in 2026

The meter image recognition field saw notable moves in 2026 as well.

On April 14, 2026, Google DeepMind released “Gemini Robotics-ER 1.6,” a model for robots, and announced that in an instrument reading evaluation, the success rate was 23% for the previous-generation ER 1.5, 86% for ER 1.6, and 93% for ER 1.6 with agentic vision (a combination of visual reasoning and code execution). In the same evaluation, Gemini 3.0 Flash scored 67%. However, these are values from DeepMind’s own evaluation, and we could not confirm the size of the evaluation data or the types of instruments. A general-purpose AI’s figures will not necessarily carry over as-is to the gauges in your own factory. On April 30, 2026, Automation World reported that Boston Dynamics is using this model on its quadruped robot Spot to have it read gauges and sight glasses in facilities.

In Japan, too, Mitsubishi Electric Digital Innovation presented kizkia-Meter, which automatically reads gauges with cameras and AI, at the “Smart Factory EXPO” held at Tokyo Big Sight on January 21–23, 2026. On July 30, 2026, IXS (Kawasaki City) announced the launch of “GENBA-Meter Read,” which periodically photographs analog gauges with cameras installed on site and converts the indicated values into data with AI. According to the announcement, it raises an alert when a threshold is exceeded and centrally manages images and inspection data from multiple sites in a web app.

What these show is that “reading gauges from images” is itself no longer an unusual technology. For buyers, the question has shifted to which method to use, which gauges to read, how often, and how to catch misreadings.

The labor and labor-cost context

As far as we could confirm, the latest revision of Thailand’s minimum wage took effect on July 1, 2025, with 17 tiers ranging from THB 337 to 400 per day. Bangkok went from THB 372 to THB 400. It is not uniform nationwide; the amount differs by province and industry. Over the medium to long term, a commentary by RSIS (S. Rajaratnam School of International Studies, Singapore) cites a projection by Thailand’s National Economic and Social Development Council (NESDC) warning that labor demand will reach 44.71 million by 2037, while Thailand’s working-age population is only 40.70 million (the commentary does not state the reference date of the working-age population figure). These are national figures, not limited to manufacturing.

However, as the calculation later shows, introducing analog gauge reading AI on the grounds of rising labor costs alone did not pay back in this article’s model calculation. Even if maintenance staff patrol time goes down, the cost of cameras, communications and licenses tends to exceed it. Labor shortages can be a motive for “wanting to reduce patrols,” but the pillar of the investment decision should be the losses avoided by noticing abnormalities early.

What Is Analog Gauge Reading AI: How Meter Image Recognition Reads Values from Images

The processing flow of analog meter reading

Analog meter reading is not done by a single AI producing an answer in one step, but by chaining several processing stages. The paper “Under pressure: learning-based analog gauge reading in the wild,” presented at ICRA 2024 by a research team at ETH Zurich (Swiss Federal Institute of Technology in Zurich), shows this flow clearly.

1. Find the gauge in the image and crop it
2. Detect the scale tick marks (notches)
3. Fit an ellipse to the dial shape (because a circle looks like an ellipse when photographed at an angle)
4. Segment the needle region and determine the needle direction as a line
5. Read the scale numbers and unit with OCR (optical character recognition)
6. Determine the correspondence between scale angles and values, and convert the intersection of the needle and the ellipse into a value

In the paper, when OCR correctly read two or more scale numbers, the mean error relative to full scale (the entire scale range) was 1.15% for images taken head-on and 0.77% for images taken at an angle. The average on a comparison dataset created by other researchers (5 types of gauges) was 2.01%, but this value excludes one type with a non-linear scale and is calculated only from images that could be read (reading failed entirely for one type, and more than 70% of readings failed for another). All of these are values under the paper’s conditions, not product accuracy.

Error sources: OCR, oblique angles, reflections and shadows

In the same paper’s failure analysis, the most common failures occurred at the OCR stage. The OCR failure rates by condition were reported as 40% head-on, 57.5% oblique, 42.5% rotated, and 50% oblique and rotated. Factors the paper cites as making reading difficult include distortion from oblique shooting, reflections on the glass, insufficient contrast such as a white needle on a white background, needle shadows, inability to read decimal points or minus signs, and unusually shaped gauges not in the training data.

However, this paper uses the demanding setting of “reading without being told the scale range in advance.” Many products register the minimum value, maximum value and scale range for each gauge up front, in which case OCR failures have less effect on the results. The paper’s OCR failure rates should not be applied as-is to product performance. On the other hand, the tendency for oblique angles, reflections and shadows to cause errors can be used as conditions to verify for products in acceptance testing as well.

There is also research on training gauge reading with synthetic data. The WACV 2024 paper “Learning to Read Analog Gauges from Synthetic Data” reports that on validation data of 4,813 real images, it improved the mean error over existing methods by 4.55 degrees (angular error), a 52% relative improvement. Note that the unit of error is “degrees.” Since each study defines error differently, the figures cannot be compared simply by lining them up.

What analog gauge reading AI can and cannot do

Most important is that even if the AI reads correctly, the value is wrong if the instrument is out of calibration. Analog gauge reading AI is a substitute for “human eyes,” not for “instrument calibration.”

Choosing an Automated Gauge Reading Method: Battery Camera / Wired Edge / Smartphone / Patrol Robot / Replace with Transmitter

There are broadly five methods for automated gauge reading. The following are examples of options on the market, not recommendations or comparative evaluations. All figures are conditional values such as vendor-published values, and the same values will not necessarily be achieved at your own site.

The difference between battery cameras and wired edge is “frequency”

The strength of battery-powered cameras is that no power or wiring work is needed. In exchange, the more shots you take, the faster the battery drains. The “about 3 years” published by LiLz assumes 3 shots per day, and a July 2025 news report described the company’s wireless camera LC-20 as lasting “up to 3 years.” If 3 times a day is enough, it suits the use of automating patrol records, but it is not suited to “noticing abnormalities within minutes.”

Wired edge requires power and wiring work but can read at intervals of seconds to minutes. kizkia-Meter is described as a cloud-free on-premises configuration that can connect to existing monitoring systems via PLC or HTTP API. Edge AI configurations that process reading results inside the factory are covered in detail in edge AI factory deployment.

As a published pricing example, Hitachi Systems announced a “camera-based automatic meter reading service” on March 30, 2018, with an initial fee from JPY 250,000 and a monthly fee from JPY 500 per meter (excluding tax). This was a 2018 announcement, and the current price could not be confirmed. It is a minimum price in yen for the Japanese domestic market, not a price in Thailand. When thinking about cost levels, always obtain a quote for your own company.

Smartphone capture and patrol robots are methods that “keep the patrol”

Smartphone capture is a method in which people patrol and take photos. Since AI does the reading and ledger entry, transcription errors disappear, but the viewing frequency stays the same as current patrols. On March 27, 2024, hakaru.ai announced an update to its AI for round analog gauges, making it usable immediately with a standard model without setting sample images or reference points.

With patrol robots, a robot patrols and takes photos instead of a person. ugo mini can adjust its camera height between 55 and 175 cm; a service in which it autonomously patrols facilities and photographs gauges, with hakaru.ai reading the values, started in January 2025. ANYmal is described as able to shoot from a distance with 20x optical zoom. However, robot patrols also run on a “number of rounds,” so the time when no one is looking does not disappear.

Replacing each gauge with a transmitter is another route to remote pressure gauge monitoring

Instead of reading with a camera, there is also the option of replacing the gauge with an instrument that has a transmitter. Emerson’s wireless pressure gauge is a battery-powered instrument supporting WirelessHART; its manual states that it contains one D-size lithium thionyl chloride battery, that the update interval can be configured, and that repeatedly switching it ON/OFF shortens battery life. WirelessHART became an international standard as IEC 62591 in March 2010.

The replacement approach eliminates the problem of errors when reading values from images (the accuracy and calibration of the instrument itself are still required), but in exchange, piping work or a line stoppage may be needed, as well as infrastructure such as a gateway. A realistic approach is to combine them: replacement for critical points that are few in number and where a shutdown can be scheduled, and cameras for points that are numerous and where you want to avoid construction work. For the approach of converting old equipment to IoT with sensors, see also legacy equipment IoT retrofit and IoT retrofit for aging equipment.

Which Gauges to Target: Start with “Points You Cannot Afford to Have Stop,” Not All of Them

Putting cameras on every gauge tends to delay payback

Depending on the factory, the number of gauges to be read can range from dozens to hundreds. Putting a camera on all of them greatly reduces patrols, but the costs of cameras, installation, communications and licenses increase in proportion to the number of points. Meanwhile, many gauges are “fine if seen several times a day,” and noticing an abnormality a few minutes earlier does not change the loss. If you put cameras on all points, the cost scales with the number of points while most of the benefit comes from only some of them.

Three criteria for choosing target gauges

Choose target gauges by the following three criteria.

1. Points with large losses when abnormal: points such as cooling water pressure, compressed air pressure, and pressure and water level around boilers, where noticing an abnormality late leads to line stoppages, quality defects or equipment damage
2. Points that are hard to reach: high places, rooftops, inside pits, areas requiring entry procedures, and so on, where the patrol burden is heavy and missed readings are likely
3. Points where you want higher frequency: points that change quickly, where one patrol per shift is not enough; points that tend to change during night hours when fewer people are around

Conversely, defer points whose values move only a few times a year, points that already have transmitters and can be seen on a monitoring screen, and points that are hard for AI to read, such as non-linear scales. For flow meters and power meters that already have transmitters, taking the signal directly is more reliable than using a camera. Those methods are summarized in flow meter monitoring and power meter data collection.

Count from past downtime records

The most reliable basis for choosing targets is past downtime records. Review records of line stoppages and equipment trouble over the past several years, and check one by one “which gauge, if watched earlier, would have prevented the stoppage.” If there are records such as a line stopping due to a drop in cooling water pressure, or equipment stopping on an alarm due to a drop in air pressure, that gauge is a top-priority candidate. If you have no system for organizing failure history by equipment, one option is to consider it together with introducing an equipment maintenance management system.

Decide abnormality judgment and notification downstream of reading

When evaluating analog gauge reading AI, attention tends to go to reading accuracy. But what creates the benefit is the “judgment and notification” that comes after reading. There are three things to decide.

• Thresholds: upper and lower limits, how many consecutive exceedances count as an abnormality, how to handle needle fluctuation (pulsation)
• To whom: the person in charge for each shift, the maintenance on-call person, the equipment owner, etc. Who receives it during night shift hours
• Within how many minutes: the time until the notification arrives and the time until the recipient reaches the site

Even if readings happen every few minutes, it is meaningless if the notification is an email read the next morning. If you want to catch signs from value trends rather than just watching fixed thresholds, the AI anomaly detection approach can also be used.

Analog Gauge Reading AI Costs and ROI: A Model Calculation

From here on is a calculation for a model factory. All figures below are this article’s own estimates and assumed (placeholder) values, not industry averages or survey figures. No vendor prices are used.

Common assumptions (Model Factory M)

Assume a Japanese-owned parts factory in central Thailand.

Patrols are not reduced to zero. Leaks, sounds, smells and vibration need to be checked by people, and gauges outside the AI’s scope are also read by people. Therefore, even with cameras, patrol time only gets shorter; it does not disappear.

The number of stoppages that can be avoided varies completely from factory to factory. The 2 per year here is a placeholder for calculation purposes, not a figure with any basis. Also, the benefit of avoided downtime arises “only when a stoppage is actually avoided as a result of noticing early.” If you notice but cannot respond in time, the benefit is 0.

Comparing three cases

Annual net is “annual benefit − annual cost,” and payback period is “initial cost ÷ annual net.” All amounts are in THB.

The cost breakdown is as follows.

• Initial cost: for 80 points, 80×25,000 + 300,000 = 2,300,000; for 30 points, 30×25,000 + 300,000 = 1,050,000
• Annual cost: for 80 points, 80×3,000 + 100,000 = 340,000; for 30 points, 30×3,000 + 100,000 = 190,000
• Downtime avoided: 4 hours × THB 60,000 × 2 per year = 480,000 (the same amount for B and C)

Case A puts cameras on 80 of the 120 points and shortens each patrol round from 2.0 hours to 0.7 hours. The reduction is 1.3 hours × 3 rounds × 365 days = 1,423.5 hours, worth 1,423.5×120 = 170,820. This does not reach the annual cost of 340,000; 170,820 − 340,000 = −169,180, meaning a net outflow of THB 169,180 every year. Patrol reduction alone does not pay back.

Case B uses the same 80 points and also includes the benefit of avoided downtime. Annual benefit is 170,820 + 480,000 = 650,820, annual net is 650,820 − 340,000 = 310,820, and payback is 2,300,000 ÷ 310,820 = 7.39…, or about 7.4 years.

Case C puts cameras only on the 30 points you cannot afford to have stop. The remaining 90 points are read by people, so each patrol round only drops from 2.0 hours to 1.5 hours. The reduction is 0.5 hours × 3 rounds × 365 days = 547.5 hours, worth 547.5×120 = 65,700. However, the same 480,000 of avoided downtime as in B is generated, so annual benefit is 65,700 + 480,000 = 545,700, annual net is 545,700 − 190,000 = 355,700, and payback is 1,050,000 ÷ 355,700 = 2.95…, or about 3.0 years.

The assumption here is that all abnormalities that lead to stoppages occur within the 30 selected points. Choosing targets from past downtime records is how you verify this assumption against your own track record. If abnormalities leading to stoppages occur outside the 30 points, Case C’s avoided downtime will be smaller by that amount.

Sensitivity: if fewer stoppages can be avoided

In Case C, vary the number of stoppages that can be avoided.

• Once a year: avoided downtime is 4×60,000×1 = 240,000. Annual benefit is 65,700 + 240,000 = 305,700, annual net is 305,700 − 190,000 = 115,700, and payback is 1,050,000 ÷ 115,700 = 9.07…, or about 9.1 years
• Zero times: annual benefit is only the patrol reduction of 65,700, and 65,700 − 190,000 = −124,300. It does not pay back

Even with the same 30 points, if the number of avoidable stoppages falls from 2 to 1 per year, payback extends from about 3.0 years to about 9.1 years. At zero, the patrol reduction does not even cover the annual cost.

Key takeaway: payback is decided by neither reading accuracy nor patrol reduction

What this calculation shows is that payback is decided not by “reading accuracy” or “patrol reduction,” but by “at which points, and how many times a year, abnormalities occur that would not have caused a stoppage had they been noticed early.” Across Cases A to C, the patrol reduction amount is at most 170,820 and never exceeds the annual cost. What makes the difference is the 480,000 of avoided downtime, and Case C, which goes after it with fewer points, pays back fastest.

That is exactly why, before starting a PoC (proof of concept), the first step is to count “stoppages that could have been avoided by noticing early” from past downtime records. If there is not even one a year, analog gauge reading AI becomes a tool for automating records, and whether it is worth the cost must be judged separately. For how to end a PoC and the criteria for a full rollout decision, see also AI PoC exit criteria.

Cautions to avoid double counting

• Patrol reduction is maintenance staff time; avoided downtime is line loss. They are different things, so it is fine to add them together
• Avoided downtime counts “only when it is avoided as a result of noticing early.” If you notice but cannot respond in time, it is 0. Unless it is decided who acts within how many minutes at the receiving end of the notification, this benefit cannot be expected
• The transcription reduction created by automated reading is included in the patrol time (the 2.0 hours per round include transcription). It is not added separately

When calculating for your own company, replace at least these five with actual figures: “number of target points,” “number of patrols and time per round,” “maintenance staff labor cost,” “loss per hour of line stoppage,” and “number of stoppages that could have been avoided by noticing early.”

12 Items to Include in an RFP for Analog Gauge Reading AI

When obtaining proposals from multiple vendors, include the following 12 items in the RFP to align the proposals’ assumptions.

Items 4, 5 and 7 in particular are often missing from proposals. If you place an order without acceptance criteria, you cannot verify anything at acceptance beyond “it runs.” A system that outputs plausible but wrong values when it cannot read is more dangerous than one that honestly returns “unreadable.” If you will use it for inspection records, also confirm that the original images and readings can be retained in a form that cannot be rewritten later.

What to Verify in FAT/SAT for Analog Gauge Reading AI

Build the evaluation set first

Build an evaluation set before acceptance testing. For example, photograph 10 representative gauges (including round pressure gauges, thermometers, level gauges and 7-segment displays) for 2 weeks, and treat the values read by people at the same times as the “ground truth.”

In FAT (testing in the vendor’s environment), provide the evaluation set images to the vendor and test in the vendor’s environment. In SAT (testing with actual cameras at your own site), photograph the same gauges with the actual cameras installed on site and run the same procedure. By running both with an evaluation set of the same composition and the same procedure, you can see in numbers any gap where it worked in the vendor’s environment but failed on site.

The target values below are placeholder examples. They are not standard values. The factory decides its own acceptance criteria.

Metrics to check

1. Reading error: compute the difference from human readings as a % of full scale. An example target is “within half of the smallest scale division used for control.” Look not only at the average but also at the conditions of the images with large differences
2. Number of failed readings: count the rate of unreadable cases. Confirm whether it correctly returned “unreadable” in those cases and did not silently output a wrong value
3. Tests under varied conditions: reflections from the afternoon sun or lighting, nighttime, fogged or oil-stained glass, needle fluctuation (pulsation), oblique shooting, different scale units (kPa, bar and psi shown together), non-linear scales
4. Time until notification: the time from exceeding a threshold until the notification reaches the person in charge
5. Reconfiguration burden: the time required for reconfiguration when the camera shifts or after a gauge is replaced, and who does it
6. Power and communication recovery: for battery-powered units, battery level notifications and replacement cycle; for wired units, recovery after power outages or network interruptions
7. Output and records: output to CSV, API and PLC; whether records are stored in a tamper-proof form as inspection records

Always include tests under varied conditions

Testing only with well-conditioned images will not reveal the problems that appear once deployed on site. In the paper mentioned earlier, oblique shooting, reflections, insufficient contrast and shadows were also cited as error sources. Deliberately include poor-condition images in the evaluation set: late afternoon when the low western sun hits the dial, times with only nighttime lighting, glass fogged in the rainy season, dials stained with oil. For pressure gauges whose needle fluctuates due to pulsation, the correct value cannot be determined from a single image, so verify how multiple images are handled. For gauges with dual kPa and psi scales, confirm which scale is being read.

After go-live, compare against human readings once a month

Even after go-live, have a person read at the same time at least once a month and compare against the AI’s values. Readings can quietly degrade due to camera shifts, lighting replacement, dirty glass, gauge replacement and so on. The records of this comparison also serve directly as the basis for “whether the AI’s values may be used as inspection records.”

Thailand- and ASEAN-Specific Issues

1. Statutory inspections are not replaced by AI

In Thailand, statutory safety inspections have been required for boilers and similar equipment. According to JETRO’s unofficial English translation of Ministry of Industry Notification No.18 (B.E.2528, 1985), factories using boilers or similar equipment must conduct a safety inspection at least once a year, the inspection is to be performed by a mechanical engineer under the Engineering Profession Act or a special boiler inspector, and the results must be submitted in the form prescribed by the Department of Industrial Works (DIW) within 30 days of the inspection date. The same notification specifies that pressure gauges must be at least 100mm in diameter, with a scale range of 1.5 to 2 times the maximum working pressure. The legally binding version is the original Thai text.

This notification dates from 1985, and related regulations have been issued since. For example, the Ministry of Labour’s ministerial regulation on safety of machinery, cranes and boilers (B.E.2564, 2021) was published in the Royal Gazette on August 6, 2021. This article has not been able to confirm the contents of the latest regulations. Please confirm current inspection requirements individually with the competent authorities or experts.

In any case, records from analog gauge reading AI do not replace statutory safety inspections. AI is a tool for recording daily values and noticing abnormalities early. How to digitize statutory inspection records is covered in statutory inspection digitalization.

The scale range requirement for pressure gauges also relates to AI reading. By this article’s own calculation, if the scale starts at 0 and the scale range is 1.5 to 2 times the maximum working pressure, then even at maximum working pressure the needle stays at about 50–67% of the full scale (1÷2 = 50%, 1÷1.5 ≒ 67%). As a result, even with the same error relative to full scale, the error relative to the values used for control is larger than it appears. It is safer to set acceptance criteria not only relative to full scale but also by how far readings differ from human readings near the values actually used for control.

2. Calibration and metrological traceability

Even if the AI reads correctly, the value is wrong if the instrument itself is out of calibration. Analog gauge reading AI only reads the display; it does not calibrate the instrument. At the top of Thailand’s metrological traceability is NIMT (National Institute of Metrology (Thailand)), the national metrology institute, established on June 1, 1998, which provides calibration services in fields such as pressure and temperature. Separately from automating readings, continue your instrument calibration plan as before.

3. PDPA: dealing with people captured in images

Thailand’s Personal Data Protection Act (PDPA, B.E.2562) came fully into force on June 1, 2022. Note that even with a camera intended to photograph gauges, if workers appear in the field of view, it may constitute processing of personal data. The following are recommendations; please confirm the legal conclusions individually.

• Narrow the field of view to the gauge dial, excluding aisles and work areas
• If people appearing in the image cannot be avoided, mask that area
• Set a retention period for original images and do not keep them longer than necessary
• Post signage in Thai where cameras are installed

The PDPC (Thailand’s Personal Data Protection Committee) held a public hearing on draft guidelines on April 1–2, 2026, and one of the topics was the use of CCTV and access control systems. However, these are at the draft stage, and their content assumes housing estates and condominiums; they do not directly address factory inspection cameras. How to organize factory IoT data in general from a PDPA perspective is summarized in factory IoT and PDPA.

4. The labor-cost context

The minimum wage, under the revision effective July 1, 2025, is THB 337 to 400 per day in 17 tiers. In this article’s model calculation, patrol reduction alone did not pay back the investment. Think of labor costs as a reason for “wanting to reduce patrols,” but not as something likely to be the pillar of payback.

5. Communication and environment

Thai factories face conditions such as outdoor utility equipment, high temperature and humidity, heavy rainy-season downpours and dust. For outdoor battery-powered cameras, check whether LTE signal reaches; for indoor powered cameras, check wiring routes and dust countermeasures. Fogged or dirty glass is likely in the rainy season and in dusty processes, so include these in the acceptance test conditions as well.

6. Local maintenance

Camera cleaning, correcting misalignment, battery replacement and reconfiguration after gauge replacement will inevitably occur after go-live. Decide before ordering who in Thailand will handle these and how quickly. Remote support from Japan alone cannot wipe the glass on site.

7. Vietnam sites

In Vietnam, the Personal Data Protection Law (Law No. 91/2025/QH15) was passed by the National Assembly on June 26, 2025 and took effect on January 1, 2026. It also contains provisions on audio and video recording in public places. When extending the same system as in Thailand to a Vietnam site, separately confirm how people captured in images are handled.

90-Day Plan for Analog Gauge Reading AI Implementation

The key is to count downtime records in the Days 0–30 stage. If you start by installing cameras and checking accuracy, you stop at “the AI could read the gauges,” and you never obtain the number of avoided stoppages needed for the investment decision. Proceed in this order: count first, narrow the targets, then trial methods on those points.

Frequently Asked Questions (FAQ)

How accurate can AI-based analog meter reading be?

As a research example, the ETH Zurich paper (ICRA 2024) reports mean errors relative to full scale of 1.15% head-on and 0.77% oblique when OCR correctly read the scale numbers. However, these are values under the paper’s conditions, and the most common failures occurred at the OCR stage. Many products register the scale range in advance, so the conditions differ. How accurate it is on your own gauges should be verified in acceptance testing with an evaluation set using human readings as ground truth, including conditions such as oblique angles and reflections.

Which is better: replacing gauges with transmitters or reading them with cameras?

Decide based on whether piping work or a line stoppage is possible, the number of points, and the required reading frequency. Replacing with transmitters eliminates the problem of errors when reading from images (calibration of the instrument itself is still required), but in exchange requires construction work and infrastructure such as gateways. Cameras keep construction work small, but reading errors and the effort of reconfiguration remain. A realistic approach is to combine them: replacement for critical points that are few in number, cameras for points that are numerous and where you want to avoid construction work.

Can patrol inspections be eliminated entirely after introducing AI?

No. Leaks, abnormal noise, smells, vibration and the like need to be checked by people with their five senses, and gauges outside the AI’s scope are also read by people. What patrol inspection automation changes is that patrol time gets shorter and the gauges you cannot afford to have fail can also be seen between patrols.

What are typical costs and payback periods for automated gauge reading?

In this article’s model calculation (placeholder values), putting cameras on 80 points and only reducing patrols did not pay back; adding avoided downtime gave about 7.4 years, and narrowing to the 30 points you cannot afford to have stop gave about 3.0 years. If the number of avoidable stoppages falls to once a year, it is about 9.1 years, and at zero it does not pay back. Both costs and benefits vary greatly between factories, so replace the figures with your own downtime records and quotes when calculating.

Can AI records substitute for statutory boiler inspections in Thailand?

No. Records from analog gauge reading AI are records of daily values and do not replace statutory safety inspections. The AI does not calibrate instruments either. Please confirm current inspection requirements individually with the competent authorities or experts.

What should be checked in the RFP and acceptance tests (FAT/SAT) for analog gauge reading AI?

In the RFP, align 12 items: the list and types of target gauges, reading cycle, method with power and communication, how ground truth is created and acceptance criteria, handling when unreadable, thresholds and notification, storage of inspection records and tamper prevention, PLC/MES/Excel integration, people captured in images and PDPA, reconfiguration, Thai-language screens and notification language, and local maintenance and battery replacement. In FAT/SAT, use an evaluation set of the same composition with human readings as ground truth to verify reading error, the number of failed readings and the behavior in those cases, tests under varied conditions, and time until notification.

Summary

• Whether analog gauge reading AI pays back is decided not by patrol labor costs but by how many minutes apart you can see the gauges you cannot afford to have fail and how much earlier you notice abnormalities
• The order of decisions is five: target gauges, method, how to create ground truth, abnormality judgment and notification, and handling of inspection records and statutory requirements
• Methods are battery-powered cameras, wired edge, smartphone capture, patrol robots and replacement with transmitters. The core differences are reading frequency and the scale of construction work
• Vendor and research figures are published values, maximum values, minimum values or values under evaluation conditions, and the same values will not necessarily be achieved at your own site
• In the model calculation (placeholder values), patrol reduction alone does not pay back; 80 points with avoided downtime gives about 7.4 years, and narrowing to 30 points gives about 3.0 years. If avoided downtime is once a year, about 9.1 years
• Align assumptions with the 12-item RFP, and in FAT/SAT run the same evaluation set using human readings as ground truth
• Statutory inspections and calibration are not replaced by AI. Proceed on Thailand-specific issues such as PDPA, communication environment and local maintenance while confirming individually with the competent authorities and experts

You are welcome to start from choosing target gauges, taking stock of past downtime records, or creating ground-truth data from human readings. If you are considering analog gauge reading AI implementation at a factory in Thailand, we can help you sort things out from the very first stage of your evaluation, so please feel free to reach out via our contact page.

References

• LiLz Gauge product page (LiLz Inc.): https://lilz.jp/products/lilz-gauge
• LiLz beta release of lamp-type AI (AIsmiley): https://aismiley.co.jp/ai_news/lilz-releases-ramp-type-ai/
• LiLz funding round and LC-20 (THE BRIDGE, July 9, 2025): https://thebridge.jp/2025/07/lilz-raises-430-million-yen-to-expand-its-sensory-inspection-solution-overseas
• kizkia-Meter product listing page (AIsmiley): https://aismiley.co.jp/product/mdis_kizkia-meter/
• Mitsubishi Electric Digital Innovation exhibit at Smart Factory EXPO (January 21–23, 2026): https://www.mitsubishielectric.co.jp/medigital/event/2026/0121/
• Hitachi Systems camera-based automatic meter reading service (March 30, 2018): https://www.hitachi-systems.com/news/2018/20180330.html
• hakaru.ai product version launch (GMO Cloud, January 24, 2019): https://group.gmo/pdf/news/gmo_news_6285.pdf
• hakaru.ai update to AI for round analog gauges (Cloud Watch, March 28, 2024): https://cloud.watch.impress.co.jp/docs/news/1579601.html
• ugo mini × hakaru.ai integration (MONOist, January 8, 2025): https://monoist.itmedia.co.jp/mn/articles/2501/08/news094.html
• ugo mini × hakaru.ai integration announcement (livedoor News, December 13, 2024): https://news.livedoor.com/article/detail/27754892/
• IXS GENBA-Meter Read launch (LOGISTICS TODAY, July 30, 2026): https://www.logi-today.com/983158
• ANYbotics ANYmal product page: https://www.anybotics.com/anymal-autonomous-legged-robot/
• Gemini Robotics-ER 1.6 (Google DeepMind, April 14, 2026): https://deepmind.google/blog/gemini-robotics-er-1-6
• Boston Dynamics and Google integrate Gemini into Spot (Automation World, April 30, 2026): https://www.automationworld.com/factory/plant-maintenance/news/55372413/boston-dynamics-and-google-bring-gemini-ai-to-spot-robot-for-smarter-facility-inspections
• Under pressure: learning-based analog gauge reading in the wild (arXiv, ICRA 2024): https://arxiv.org/abs/2404.08785
• PDF of the same paper: https://arxiv.org/pdf/2404.08785v1
• Learning to Read Analog Gauges from Synthetic Data (arXiv, WACV 2024): https://arxiv.org/abs/2308.14583
• Rosemount Wireless Pressure Gauge reference manual (Emerson, revised May 2025): https://emerson.com/documents/automation/manual-rosemount-wireless-pressure-gauge-wirelesshart-protocol-en-80198.pdf
• Standardization of ISA100 and WirelessHART (Processing Magazine, June 30, 2010): https://www.processingmagazine.com/process-control-automation/instrumentation/article/55345996/status-update-isa100-wirelesshart
• Thai Ministry of Industry Notification No.18 B.E.2528, unofficial English translation (JETRO): https://www.jetro.go.jp/ext_images/thailand/e_activity/pdf/moinoti2.pdf
• Thai DIW Regulation B.E.2528, unofficial English translation (JETRO): https://www.jetro.go.jp/ext_images/thailand/e_activity/pdf/diwreg1.pdf
• Thai Ministry of Labour ministerial regulation on safety of machinery, cranes and boilers B.E.2564 (Enviliance ASIA): https://enviliance.com/regions/southeast-asia/th/report_4310
• Postponement of full enforcement of Thailand’s PDPA (LawPlus Ltd., May 2021): https://www.lawplusltd.com/2021/05/pdpa-is-postponed-again-for-one-more-year/
• Public hearing on draft PDPA guidelines (Tilleke & Gibbins, April 9, 2026): https://www.tilleke.com/insights/thailands-public-consultation-on-proposed-pdpa-guidelines-key-updates
• NIMT (National Institute of Metrology (Thailand)): https://en.nimt.or.th/?p=215
• Thailand minimum wage revision (Forvis Mazars Thailand, July 2025): https://www.forvismazars.com/th/en/insights/doing-business-in-thailand/payroll/forvis-mazars-payroll-flash-news-july-2025
• Thailand minimum wage revision (Tilleke & Gibbins, July 4, 2025): https://www.tilleke.com/insights/thailand-raises-minimum-wage-for-bangkok-and-certain-businesses-nationwide/26/
• Thailand’s aging society and labor force (RSIS Commentary, September 3, 2024): https://rsis.edu.sg/wp-content/uploads/2024/09/CO24128.pdf
• Vietnam Personal Data Protection Law, Law No. 91/2025/QH15 (Tilleke & Gibbins, August 27, 2025): https://tilleke.com/insights/vietnams-new-personal-data-protection-law-a-closer-look