Tool Life Prediction AI: Wear and Breakage Detection, RFP and FAT/SAT for Thai Plants

Many Japanese machining plants in Thailand that are interested in tool life prediction AI currently decide tool changes either by piece counts (fixed tool life) or by operator intuition. Set the fixed life short, and you throw away tools that could still cut, which drives up tooling costs. Set it long, and you keep cutting with worn tools, which leads to dimensional or surface roughness defects, or to tool breakage that stops the machine. Many shop floors will recognize the feeling of struggling with both changing too early and changing too late.

This article makes one argument. The success of tool life prediction AI depends not on model accuracy, but on whether you first define the tool change rule (what counts as end of life, and who changes the tool, when, and on what basis), and then whether the shop floor can keep collecting the signals and ground-truth labels (records of measured actual wear) that the rule requires. The high accuracy figures seen in research papers come from experiments with fixed cutting conditions; they cannot be carried over as-is to high-mix production lines. What acceptance testing should confirm is whether tool change decisions come out according to the rule you defined, within the tolerated ranges for false alarms and missed detections.

Below, so that production engineering managers, quality managers and plant managers can assemble the material they need for internal decisions, we go through how to set the life criterion, the signals used for tool wear detection and tool breakage detection, how to read the figures from research datasets, the product and approach options, integration with existing systems, a model calculation of costs and benefits, the RFP items, the FAT/SAT checkpoints, and issues specific to Thailand and ASEAN. Note that the tool quantities, unit prices, incident counts, amounts and payback periods for the model plant in this article are all hypothetical figures set for this article. They are not industry averages, survey values or manufacturer prices.

Why tool life prediction is now being asked about in Thai machining plants

In the second half of 2026, the numbers surrounding Thai machining plants point in different directions.

According to a report distributed by Reuters, Thailand’s automobile production in August 2026, as announced by the Federation of Thai Industries (FTI), was 124,646 units, up 10.93% year on year. In the same month, domestic sales rose 25.59% year on year, while exports fell 2.04%. On the other hand, the FTI forecasts that full-year 2026 automobile production will be down 3.33% from the previous year. The reason given is a decline in exports due to fighting in the Middle East. Even if a single month shows growth, the full-year outlook gives no grounds for optimism. For machining plants making automotive parts and dies, this is a situation in which it matters more than ever to keep tooling and defect costs under control despite fluctuating orders, and to meet delivery dates with limited machines and people.

Tools have always had an element of unpredictability. A 2023 paper in Data in Brief introducing a public dataset cites earlier research stating that tool life can vary by more than 100% even under the same cutting conditions, and that, as a result, in practice all tools are in many cases replaced uniformly after a set time, which wastes the remaining life of many tools. This statement comes from a paper dealing with milling tools, but it captures well the frustration of shop floors that change tools on a fixed life. A fixed life has to be set to match the “bad side” of that variation.

The use of AI itself is also spreading. According to figures reported as coming from a Microsoft report, Thailand’s AI adoption rate rose from 9.1% in early 2025 to 12.4% in the first quarter of 2026, and its growth rate of 36.4% year on year was the second highest in the world. However, these figures are based on Thailand’s overall population and workforce, not limited to manufacturing or machining. The same report states that 87.6% of the population still does not use AI in daily life or work. Expectations around the word AI may be rising, but it is reasonable to read this as meaning that the stage of putting it to practical use on factory floors is still ahead.

At trade shows, too, the combination of AI and machine tools has become a prominent theme. For IMTS 2026, held in Chicago from September 14 to 19, 2026, the organizer announced an “Industrial AI Arena” (32 exhibitors of AI manufacturing solutions plus Sandia National Laboratories) and an “Industrial AI Conference” covering predictive maintenance, quality and data preparation. At the same venue, Caron Engineering says it demonstrated real-time CNC tool monitoring and adaptive control (TMAC) in a fully automated, live machining cell using a Tsugami SS20. In Japan, the 33rd Japan International Machine Tool Fair (JIMTOF 2026) is scheduled to be held at Tokyo Big Sight from October 26 to 31, 2026.

Against this background, the question “Can AI predict tool life?” is reaching plants in Thailand as well. The answer is: “Yes, under certain conditions, but there are things to decide first.” The following sections look at those things to decide first, in order.

What is tool life prediction AI: fixed life management vs tool condition monitoring vs life prediction

The term “tool life prediction AI” tends to be used for three things that are actually different in nature. Discussions about implementation often talk past each other, largely because this distinction is not made.

1. Fixed life management (count-based cutting tool life management)

This is the method of counting the number of uses or the usage time with the CNC’s tool life management function or a tool management sheet, and replacing the tool when it reaches a set value. This is where most plants are today. The mechanism is simple and the data collection is clear. Its weakness is that it cannot cope with the variation in tool life mentioned in the previous section. Set the life long and you get breakage and defects; set it short and tooling costs rise.

2. Tool condition monitoring

This is a mechanism that watches signals during machining and detects that “an abnormality is happening right now.” The typical example is tool breakage detection, which captures changes such as in cutting load and stops the machine. FANUC’s cutting load monitoring function, introduced later, is, according to the manufacturer’s description, a function whose main purpose is detecting tool breakage. Most tool condition monitoring systems first take on this role of “quickly finding what has happened.”

3. Life prediction (wear estimation and change timing proposals)

This is a mechanism that estimates the wear state of a tool from signals and proposes things like “how many more parts or minutes until it is likely to reach end of life” or “whether it should be changed at the next setup break.” It is easiest to understand as tool wear detection extended to the estimation of a continuous quantity. Much of the research addresses this area, but to use it on the shop floor you need a criterion for what counts as end of life, and records of actual wear to confirm whether the estimates are correct.

The three are not in opposition. A realistic way to layer them is to keep the fixed-life count as a safe-side upper limit, have monitoring stop breakage within that limit, and have prediction move the change timing earlier or later. In many cases, tool change timing optimization does not mean “removing the count limit,” but “moving the change timing, based on data, within that limit.” Whether you can write down in one sentence at the start which of these you are implementing becomes the starting point of the project.

How to set the tool life criterion (flank wear, standards, quality criteria)

To predict life, you first have to decide what counts as end of life. If you bring in AI while this remains vague, it is not settled what the model is trying to predict.

What the standards cover is “tool life testing”

Two ISO standards are frequently mentioned in connection with tool life.

• ISO 3685:1993 “Tool-life testing with single-point turning tools”: a standard specifying recommended procedures for tool life testing when turning steel and cast iron with single-point turning tools made of high-speed steel, cemented carbide and ceramics. It covers the work material, tool, cutting fluid, cutting conditions, equipment, assessment of tool deterioration and tool life, test procedure, and the recording and presentation of results. Its periodic review was completed in 2022, and it is a current standard.
• ISO 8688-1:1989 “Tool life testing in milling — Part 1: Face milling”: recommended procedures for tool life testing in face milling of steel and cast iron with cemented carbide tools. Its scope states that it can be applied in production environments as well as in laboratories. It was confirmed in December 2022 and is current. For end milling there is Part 2 (End milling), which papers cite as “ISO 8688-2.”

As their names indicate, these are standards for procedures to test tool life. They do not directly specify when a tool should be changed on a production line.

In turning, many studies take VB 0.3 mm as end of life

Some turning research papers cite ISO 3685 and set the tool life criterion as a width of flank wear. For example, a study on carbide grooving inserts takes as tool life the time until the average width VBB of the uniformly worn flank wear land reaches 0.3 mm, in accordance with ISO 3685:1993. Other papers also define tool life as the time at which VB reaches 0.3 mm, or replace the insert once flank wear exceeds 0.3 mm. In other words, it can be said that studies citing ISO 3685 often take an average flank wear VBB of 0.3 mm as end of life. This is a criterion used in research, not the change criterion of a production floor itself.

In milling, the wear type and threshold are decided per process

Milling is a different matter. Studies citing ISO 8688-2 distinguish wear types such as VB1 (uniform), VB2 (non-uniform) and VB3 (localized), and which indicator and threshold are adopted differs from paper to paper. A study on high-speed milling of titanium alloy regards VB3 exceeding 0.5 mm as worn. A study on milling Inconel 718 selects VB3, based on the approach of analyzing all wear phenomena and adopting the most influential one as the life criterion. The Data in Brief study that created the public dataset ran its experiments to end of life at a VB of around 150 µm (0.15 mm). In milling, even across studies the range runs from 0.15 mm to 0.5 mm, and no uniform “X mm” can be stated.

Shop-floor life criteria are worked backward from quality

On a production floor, the issue is less the wear width itself than whether the parts cut by that tool meet the drawing requirements. A realistic approach is therefore to set the life criterion by combining several perspectives, as below.

The important point here is to decide, as one row in this table, which criterion the AI’s predicted remaining life is measured against. Whether you predict the wear amount or the time when a quality indicator approaches its limit changes the ground-truth labels you need to collect. If you predict the wear amount, you need to keep measuring and recording the flank wear of replaced tools; if you predict quality, you need to keep records linking the surface roughness and dimensions of parts to the tool that cut them.

Write out the change rule at the same time. Who decides the change (the operator, the team leader, an automatic stop), when the tool is changed (immediately on alarm, at the next setup break, at shift change), and on what basis (predicted value, count, quality measurement). If this rule exists first, it becomes clear which field of the rule the AI output is supposed to fill. If you display “remaining life X minutes” without such a rule, the shop floor will end up continuing to change tools by count.

Incidentally, the relationship between tool life and cutting speed is classically expressed in the form V·T^n = C, presented by F. W. Taylor in a 1907 paper (V is cutting speed, T is tool life, and n and C are values determined experimentally for each tool material, work material and set of conditions). The formula reminds us of the obvious fact that changing the conditions changes the life, but n and C are values determined by experiment for each set of conditions. In a high-mix plant, the AI design must also assume that the life assumptions change every time the conditions change.

Figure: How tool wear progresses (initial wear, steady-state wear, rapid wear) and an example life criterion line. A conceptual diagram showing the relative positions of changing too early, changing too late, and prediction-based changing (changing just before the criterion line). The value of the criterion line is an illustrative example to be set per process, not a specific standard value.

Signals for tool wear detection and tool breakage detection (comparison table)

Once the life criterion and change rule are decided, choose the signals needed to run that rule. Here, based on the organization in a review paper on sensor technologies for machining (Hassan et al.) published in Sensors in 2026, we compare the strengths and weaknesses of the main signals. The review covers machining in general and is not based on measurements at any specific plant.

Signals alone are not enough: CNC context and labels

Collecting signal waveforms alone does not make them usable for tool life prediction. At a minimum, the following three types of data must be recorded aligned on the same time axis.

• Signals: load, vibration, AE, cutting force, sound and so on from the table above
• CNC context: which tool number was cutting, in which program, in which block, at what spindle speed and feed. Part counts and tool life counter values
• Labels: measured actual wear at the time of change, records of chipping and breakage, and quality measurement results of parts cut by that tool

Without the CNC context, you cannot tell whether a rise in load was due to tool wear or due to a switch to a different tool or a different process. How to acquire status data from machine tools is covered in detail in “Machine Tool Monitoring: Choosing MTConnect or umati, RFP and FAT/SAT for Thai Plants.” For the idea of using the motor current of spindle and feed axes for diagnosis, “Motor Current Signature Analysis Implementation: RFP, FAT and SAT” is also a useful reference.

The order for choosing signals

More signals are not necessarily better. Giving priority to what the shop floor can sustain, the following order is a way to proceed with less rework.

1. First, test whether data already available from the controller is sufficient. Spindle load, current and torque can in some cases be obtained without adding sensors. If breakage detection is the main purpose, this is often enough.
2. If not, add vibration, which is easy to install. However, because results change greatly with mounting position, make it a rule to decide the position, fix it and not change it.
3. Consider AE or dynamometers only for processes that need earlier signs, such as incipient chipping or small-diameter tools. Narrow the scope, accepting that data volume, durability issues, cost and the burden of retraining will all increase.

Figure: Example architecture for signal acquisition in tool life prediction. Signals such as spindle load and current, vibration, AE, cutting force and sound are aligned with CNC context (tool number, program, conditions, counters) and ground-truth labels (actual wear measurement, breakage records, quality measurement), and pass through edge judgment and machine learning to feed breakage alarms, change timing proposals and records in the MES.

How to build the AI model, and how to read research dataset figures

Research datasets are “experiments with fixed conditions”

One of the most frequently used datasets in tool life prediction papers is the PHM Society 2010 Data Challenge (milling tool wear). According to the organizer’s page, machining was done on a high-speed CNC milling machine with a 6 mm diameter carbide ball-nose end mill under a single set of conditions: spindle speed 10,400 rpm, feed 1,555 mm/min, radial depth of cut 0.125 mm and axial depth of cut 0.2 mm. The signals are 7 channels, cutting force (X/Y/Z), vibration (X/Y/Z) and AE-RMS, each sampled at 50 kHz. There are records for 6 tools, c1 to c6, with c1, c4 and c6 for training and c2, c3 and c5 for testing. The training records come with a file giving the wear amount per cut (in units of 10⁻³ mm). According to the paper discussed next, each record has one file per cut, for 315 cuts.

One example of research using this dataset is a paper on the PSD-CVT model published in Scientific Reports in 2024. The paper uses the 3 training tools (C1, C4, C6), takes the average wear of the 3 flutes as the label, and splits the data into 80% training, 10% validation and 10% test. The reported errors (MAE and RMSE, in units of 10⁻³ mm) were 0.683 and 0.939 for C1, 0.734 and 0.979 for C4, and 2.107 and 2.954 for C6. The abstract states that the accuracy was higher than CNN alone or ViT alone.

These figures need to be read together with the scope of the evaluation.

• The evaluation split training and test data within the records of the same tool. Because this kind of split can mix earlier and later cuts of the same tool between the training and test sides, it does not show performance when applied to “a different tool never seen before” or “different conditions.”
• There is only one set of cutting conditions. The premise differs from a production line where conditions change by part number.
• The authors themselves state in their conclusion that improving performance in real production environments and studying practical application are needed going forward.

Another frequently mentioned dataset is NASA’s Milling Wear dataset. Provided by the BEST Lab at the University of California, Berkeley, and published via the NASA Prognostics Center of Excellence, it records milling experiments at different speeds, feeds and depths of cut, along with the wear VB of the milling inserts (last updated May 29, 2025). These experiments vary the conditions, but they are still laboratory data.

The “gap between research and industry” that researchers themselves point out

The gap between laboratory data and production is something the researchers who build datasets are aware of themselves.

The MU-TCM face milling dataset from Mondragon University (Scientific Data, 2025) was acquired in the laboratory under varying cutting conditions, capturing the machine’s internal signals along with external cutting force, vibration and AE. The abstract states that it is expected to support the development of industrial tool condition monitoring using only internal CNC signals and to help narrow the gap between research and industry, and the main text notes that datasets acquired in the laboratory are limited in the range of cutting conditions and the number of signals.

The Data in Brief dataset (2023) mentioned earlier was created by wearing 9 carbide end mills of the same type (4 flutes, TiN-TiAlN coated, from different lots) from new to end of life (VB around 150 µm) by shoulder milling cast iron on 3 five-axis machining centers of similar size. A dynamometer was recorded at 25 kHz, and spindle and feed axis forces and torques and position deviations from the controller at 500 Hz, for a total of 6,418 files. The paper states that such public datasets are very scarce. Even with the same tool and the same conditions, the data changes if the machine is different. It is one of the few examples of public data that can address this.

Steps for building a model on the shop floor

Rather than taking research figures as expected values, build the model on the shop floor with the following steps.

1. Narrow the scope. Start with expensive tools, processes with high breakage risk, and part numbers with high defect costs. If you target all tools at once, label collection cannot keep up.
2. Decide how to collect ground-truth labels. Decide who measures the flank wear of replaced tools, with what, at which position, and where it is recorded. Because measurement takes labor, one option is to separate tools measured every time from tools measured by sampling.
3. Keep data separated by part number, tool and conditions. If you build a single model from data with mixed conditions, there is a risk of mistaking differences in conditions for wear.
4. Evaluate on “a different tool.” Test on tools other than those used for training, keeping in mind how this differs from the way research datasets are evaluated.
5. It is fine to start with rules and thresholds. For uses that capture sudden changes, such as breakage detection, upper and lower limits on load captured during a reference cut can be easier to explain than machine learning. It is enough to add prediction models in stages once labels have accumulated.

Even after the model is in place, signal trends change with tool brand changes, condition changes, machine repairs and so on. Mechanisms for watching the degradation of models in operation and deciding on updates or rollbacks are covered in “Factory AI Model Monitoring: Detect Drift and Roll Back.”

New research-stage methods are also appearing. In May 2026, the US Oak Ridge National Laboratory (ORNL) announced a method that images the surface of a tool’s cutting edge with a flexible tactile sensor and uses an algorithm to detect wear on the machine. It is a research-stage result published in the journal Wear, not a commercial product. If such technologies become practical, they may lighten the burden of collecting ground-truth labels, but for now they are something to watch as a trend.

Product and approach options (4 types of tool condition monitoring system)

The options for a tool condition monitoring system can be broadly divided into four. All product functions mentioned below are manufacturers’ descriptions from company announcements or trade press, that is, vendor claims. None of them have published prices.

CNC built-in functions

• FANUC “Cutting Load Monitor Function” (announced June 2022): monitors cutting load using servo and spindle data. The manufacturer explains that it measures the cutting load during a reference cut to automatically create tool breakage detection conditions, and monitors the cutting load during production machining to detect tool breakage. The action on breakage can be customized with PMC ladder or custom macros, and the monitored cutting load can be output to a log for use in investigating the cause of breakage, it says. The supported CNC models and sampling period are not stated on the announcement page. Also, the main focus of the description is breakage detection; it does not say it predicts wear amount.
• Okuma “OSP-AI Machining Diagnosis” (jointly developed by Okuma and NEC, reported in November 2018): it was reported that AI built into the OSP numerical controller detects abnormalities in drilling in real time and visualizes the wear state. It aims to stop machining and retract before the tool breaks, and to reduce tooling costs by optimizing change timing, and is described as using NEC’s “RAPID machine learning” technology for the AI. It targets drilling on Okuma machines (OSP control).

Dedicated monitoring devices

• Marposs (ARTIS) Genior Modular: a machine and process monitoring system that monitors tool wear, tool breakage and missing tools, and performs adaptive control of machining. According to a 2017 manufacturer announcement article in the trade press, it can monitor up to 10 machining processes simultaneously, captures signals during setup to automatically set upper and lower limit alarms, and during production dynamically adapts limit values taking into account influences such as temperature and material structure. Marposs acquired ARTIS in 2008; according to the announcement at the time, ARTIS was founded in 1983, based in Bispingen, Germany, and provided in-process tool, process and machine condition monitoring and adaptive control, mainly for metal cutting.
• Caron Engineering “TMAC”: as mentioned earlier, it demonstrated real-time tool monitoring and adaptive control in a live machining cell at IMTS 2026. The company promotes scrap reduction, shorter cycle times, extended tool life and improved quality.

Retrofit IoT plus AI

• MachineMetrics “Tool Anomaly Detection” (announced March 2025): the company explains that it detects abnormal machining behavior such as changes in load, torque and spindle speed to predict tool failure. It is described as a “sensorless” approach that acquires high-frequency data from the machine controller. The company’s case study page also contains quantitative results, but since the evaluation conditions and calculation methods are not shown, this article does not use those figures.
• Siemens’s Sinumerik Edge apps (announced 2019) include an app that continuously monitors workpiece quality during machining using high-frequency data and notifies deviations from the reference, but the announcement focuses on workpiece quality and machine condition and does not explicitly mention measuring tool wear. If considering it as a tool monitoring candidate, you need to check individually whether it can be used for that purpose.

In-house development

This is the approach of assembling sensors, edge devices and a data platform in-house or with a system integrator, and building the models yourselves. It allows customization to the target process, but you take on label collection, model updates and maintenance yourselves.

Comparison table

As a rough guide, choose based on the purpose and the machine configuration. If stopping breakage is the top priority and the target machines all use the same manufacturer’s controllers, first check whether the CNC built-in function is enough. If you require change timing proposals based on wear, or horizontal rollout across machines from multiple manufacturers, compare dedicated devices and retrofit systems. In-house development is an option when the target process is clear and you can maintain an internal structure for continuously collecting labels. Whichever you choose, the plant side must decide the “life criterion and change rule” and “how labels are collected.”

Data integration and existing systems (machine tool monitoring, MES and tool management)

Tool life prediction is not self-contained. Even if it issues change proposals, shop-floor behavior will not change unless they are connected to who looks at them and what gets recorded. When considering integration, check the following four connection points one by one.

1. Tool life counters on the CNC side. In many plants, CNC tool life management (counters of number of uses or time) is already running. Even after introducing a prediction mechanism, the basic approach is to keep the counter as the safe-side upper limit. In that case, unless you decide whether the “number of uses” counted by the prediction mechanism matches the CNC-side counter, and whether both are reset at the same timing when a tool is changed, the two numbers will diverge and confuse the shop floor.

2. Machine tool monitoring. CNC context such as tool number, program and part count overlaps with the data collected by machine monitoring systems. If you already have, or plan to introduce, machine monitoring, taking the context from the same collection path avoids time misalignment and duplicate data holdings.

3. MES and quality records. Keep records of changes, breakages, wear measurements at change, and quality measurements of parts cut by that tool, linked to part number and lot. These become the ground-truth labels for the next model improvement and also material for investigating causes when defects occur.

4. Tool management (tool inventory, regrinding and issuing). Keeping information per tool, such as when each tool was issued, how many times it was reground and which lot it came from, provides clues when investigating life variation later. The Data in Brief dataset introduced earlier also explicitly notes that the tools used came from different lots.

If you adopt an architecture where judgment is made at the edge and recording and analysis are done at the upper level, the thinking on which processing to place where is also covered in “Edge AI Factory Deployment 2026: RFP, 90-Day PoC and Acceptance.” For breakage detection in particular, the time from judgment to machine stop is important. Routing the stop command via an upper-level server exposes it to delays and communication outages, so the general thinking is to complete the stop judgment close to the machine and send only the records to the upper level.

Model calculation of costs and benefits (hypothetical figures)

What follows is a model calculation to see how much the benefits depend on assumptions. All figures below are hypothetical, set for this article. They are not industry averages or survey values, nor manufacturer prices. How much tool life extends and how much breakage falls are not “values AI can achieve,” but values that should be measured in a PoC and substituted.

Model plant assumptions

The model plant covers only one tool group: the carbide end mills and drills used on 8 machining centers. It is not the whole plant. The currency is the Thai baht (THB). There is only one world being compared against: “the current state, where tools are changed on a fixed life and there are 24 breakages per year.”

Let us check the calculation step by step. The current tooling cost is 4,000 tools × 1,500 THB = 6,000,000 THB. Assuming the number of tools used is reduced by 10% (400 tools), the tooling cost reduction is 400 tools × 1,500 THB = 600,000 THB. Assuming 12 of the 24 breakages, half of them, are eliminated, the reduction in breakage losses is 12 × 25,000 THB = 300,000 THB. The total benefit is 900,000 THB; subtracting the annual operating cost of 300,000 THB gives a net benefit of 600,000 THB; and dividing the initial cost of 1,800,000 THB by the net benefit gives a simple payback period of 3.0 years.

Note that here the “reduction rate in tools used” is set directly. If the life of each tool were extended 1.1 times, the number of tools needed would be 1÷1.1, about 0.909 times, meaning a reduction of about 9.1%, not exactly 10%. For simplicity, the calculation sets 10% as the reduction rate in the number of tools.

How the payback period moves when assumptions change

This is the sensitivity when only the reduction rate in tools used is varied to 5%, 10% and 15%, with the reduction in breakage losses ((2) 300,000 THB) held fixed.

In the 5% case, 200 tools × 1,500 THB = 300,000 THB; together with (2) that is 600,000 THB; subtracting the operating cost leaves 300,000 THB; and payback is 1,800,000÷300,000 = 6.0 years. In the 15% case, 600 tools × 1,500 THB = 900,000 THB, the total is 1,200,000 THB, the net benefit is 900,000 THB, and payback is 1,800,000÷900,000 = 2.0 years. Simply moving the reduction rate within the range of 5% to 15% moves the payback period across a threefold range, from 6.0 years to 2.0 years. Even a drop of just 5 points from the mid case of 10% doubles the payback period from 3.0 years to 6.0 years. This range shows why the actual values should be measured in a PoC before making an investment decision.

Figure: Sensitivity of the model calculation. Changes in annual benefit, annual net benefit and simple payback period when the reduction rate in tools used is varied to 5%, 10% and 15% (all figures hypothetical).

Cautions to avoid double counting

A common error in this kind of calculation is counting the same benefit twice. The model calculation takes care on the following points.

• “Reducing premature changes” and “extending tool life” are the same benefit. The portion of use beyond the fixed life appears in (1) as a reduction in the number of tools. If you set up a separate item called “reduction in premature changes,” you count the same tools twice.
• The loss per breakage does not include the tool cost. The cost of the broken tool is already included on the (1) side as part of the number of tools used. Putting the tool cost into (2) as well would double count it.
• Prediction will also cause some tools to be changed earlier than the fixed life. If tools that wear quickly are changed before the criterion line, the number of tools increases by that amount. The tools reduced in (1) need to be measured as a net value, after offsetting the tools changed later against those changed earlier.
• There is only one baseline. If you mix in a different world, such as “even without AI, some improvement would have been achieved by reviewing the counts,” there are two things being compared against, and the explanation of the benefit no longer holds.

Part of the investment may be eligible for BOI incentives if conditions are met. This is discussed in the section on issues specific to Thailand and ASEAN.

Items to include in an RFP for tool life prediction AI

In an RFP for introducing tool life prediction AI, it matters more to write in the decision rules and the verification methods than to have the vendor commit to model accuracy as a single number. We recommend including at least the following 12 items.

If a vendor’s proposal contains accuracy figures, always ask “on what data, and with what split, was this figure evaluated?” As we saw in the earlier section, an evaluation split within the records of the same tool means something different from figures evaluated on a different tool or a different machine. It is safer not to use figures without stated evaluation conditions as material for comparison.

What to confirm in FAT/SAT

What acceptance testing confirms is whether tool change decisions come out according to the rule you defined, within the tolerated ranges for false alarms and missed detections. In FAT (testing before factory shipment, or in the vendor’s verification environment) and SAT (testing after on-site installation), confirm the following items.

In SAT, it is important to decide the test period and number of cases in advance. Because breakages do not occur frequently, a short period may not yield enough cases to evaluate the missed detection rate. In that case, one approach is to write a staged acceptance into the contract: confirm the false alarm rate and the reproducibility of change proposals in SAT, and evaluate the missed detection rate over a set period after operation starts. The idea of deciding exit criteria in advance at the PoC stage is covered in detail in “IoT PoC Roadmap 2026: Decide the Exit Criteria Before You Start.”

Issues specific to Thailand and ASEAN

1. Mixed machines and manufacturers

In Japanese plants in Thailand, it is not unusual to see machine tools made in Japan, Europe, Taiwan and elsewhere side by side. Because CNC built-in functions are separate for each manufacturer, installing the same mechanism on all machines becomes difficult. It is more realistic to narrow down to the machines in the target process first, confirm results there, and then roll out horizontally.

2. A structure that can sustain wear measurement

Collecting ground-truth labels is the least glamorous and least sustainable task in introducing AI. Decide who measures the wear of replaced tools, the measuring tools and the recording format, and prepare work instructions in Thai. It is important not to make the procedure depend on specific individuals, so that records continue even when staff are transferred or leave. This is why the annual operating cost includes wear measurement labor.

3. Shop-floor judgment and language

If operators look at change proposals and make decisions, a precondition is that the screen wording and the procedure for what to do on receiving a proposal can be read in Thai without confusion. Write it down to the level of concrete actions, such as “when a proposal appears, change the tool at the next setup break and place the replaced tool on the measuring table.”

4. BOI investment promotion measures

The BOI’s “Smart and Sustainable Industry” measure (upgrading to smart and sustainable industry) is listed as one of the special measures in a BOI document dated August 4, 2026. According to the BOI Investment Promotion Guide 2025, it covers existing businesses, whether or not they are BOI-promoted, and for BOI-promoted projects the condition is that the corporate income tax exemption has already ended or was not granted. The investment must be at least 1 million baht, and land and working capital are not counted in the investment amount.

• One of the cases under adoption of digital technology is the application of AI, machine learning, and big data and data analytics, with a 3-year corporate income tax exemption (capped at 50% of the investment). The August 2026 BOI document states that the adoption of AI, machine learning and big data, cloud and data centers in Thailand, and certified Thai-made software are counted at the full investment amount, while uncertified software, foreign-made software and overseas cloud are counted at half.
• Upgrading of machinery and automation offers import duty exemption and a 3-year corporate income tax exemption (capped at 50% of the investment; the cap is 100% if automation or robots are used and linkage with Thailand’s automation industry accounts for 30% or more of the value of the upgraded machinery).
• Transformation to Industry 4.0 (a plan approved by NSTDA) offers import duty exemption and a 3-year corporate income tax exemption (capped at 100% of the investment).

The guide also lists excluded businesses, and “manufacture of general automobiles” is one of them. Manufacture of finished vehicles is listed as excluded, but machining of automotive parts does not appear by name in this exclusion list. However, whether an individual business qualifies needs to be confirmed with the BOI. In addition, according to a law firm’s commentary, the BOI announced a renewal of its investment promotion measures on January 15, 2026, with many measures accepting applications from the first business day of 2026 to the last business day of 2027 (whether each individual measure is included in this could not be confirmed from that commentary). Because conditions can be revised, please confirm the latest conditions with the BOI when considering an application. The initial cost of 1,800,000 THB in the model calculation exceeds 1 million baht on its own, but the scope and counting of eligible investment vary depending on the content of the application.

5. Division of roles with the head office in Japan

Even if life criteria and modeling approaches are shared with the head office or mother plant in Japan, the Thai plant does not necessarily use the same tools and the same conditions. As we saw with the research datasets, different machines produce different data. Rather than bringing in the head office’s models and thresholds as-is, plan on the premise of re-collecting data with the Thai machines and conditions.

Frequently asked questions

What is needed to introduce tool life prediction AI?

What is needed first is not AI but a change rule: “what counts as end of life, and who changes the tool, when, and on what basis.” On top of that, put in place a structure for collecting the signals the rule requires, CNC context data, and ground-truth labels such as actual wear at the time of change. A way to proceed with less rework is to narrow the scope to expensive tools or processes with high breakage risk, measure the benefits in a PoC, and then expand.

Which signals are best for tool wear detection?

There is no single right answer. According to the organization in a 2026 review paper, cutting force correlates strongly with wear but is invasive and expensive to install; vibration is low cost and easy to install but strongly dependent on mounting position; and AE is sensitive to incipient damage but data-heavy and requires retraining when the machine configuration changes. The recommended order is to first test whether spindle load or current from the controller is enough, then vibration if not, and to consider AE or dynamometers only for processes that need even earlier signs.

If we only need tool breakage detection, is AI unnecessary?

It depends on the use. For breakage detection, approaches that create detection conditions based on the cutting load captured during a reference cut are offered as products. For example, according to the manufacturer’s description, FANUC’s cutting load monitoring function measures cutting load during a reference cut and automatically generates breakage detection conditions. It is a good idea to first check whether such mechanisms are sufficient, and to consider prediction models at the stage where you also require wear estimation or change timing proposals.

How much can tool change timing optimization reduce tooling costs?

There is no reliable figure that can be presented as a general reduction rate. In this article’s model calculation (hypothetical), simply varying the reduction rate in tools used to 5%, 10% and 15% moved the simple payback period substantially, to 6.0 years, 3.0 years and 2.0 years. Because the benefit changes with tool type, cutting conditions and how the current fixed life is set, we recommend running a PoC on your own processes and recalculating with measured values.

Is the CNC life counter alone insufficient for cutting tool life management?

The counter is simple and reliable, so it is worth keeping as a safe-side upper limit. However, some studies indicate that tool life varies widely even under the same conditions, and with counts alone you have no choice but to set a short life matched to the bad side of that variation. A realistic way to layer them is to keep the counter as the upper limit and use monitoring and prediction within it to move the change timing.

Can a tool condition monitoring system be retrofitted to older machine tools?

If data can be obtained from the controller, retrofit mechanisms can be considered. Even if it cannot, there is the option of externally mounting vibration sensors and the like. However, vibration results change greatly with mounting position, and external sensors alone may not capture CNC context such as tool number and program. It is safer to first take stock of which machines can provide how much data, and then choose the approach.

Summary

If the introduction of tool life prediction AI starts with comparing model accuracy, it tends to produce a system the shop floor does not use. First decide the life criterion and the change rule, and build a structure for continuously collecting the signals the rule requires, the CNC context and the ground-truth labels of actual wear. Narrow the scope to expensive tools and processes with high breakage risk, and keep the fixed-life counter as a safe-side upper limit. Deciding the order this way makes it clear what to buy and what to sustain yourselves.

The high accuracy reported on research datasets is a figure from experiments with fixed conditions. Check vendors’ accuracy claims together with the evaluation data and how it was split. Because cost-effectiveness moves greatly with assumptions such as the reduction rate in tools used and how much breakage falls, judge it after substituting values measured in a PoC. Then, in FAT/SAT, confirm one by one the breakage detection response, the false alarm and missed detection rates, the reproducibility of change instructions, and consistency with the CNC counters.

TOMAS TECH helps plants in Thailand build tool monitoring and life prediction mechanisms tailored to their machine configurations and operations. You are welcome to consult us even before your implementation policy is settled, for example while you are still narrowing down target processes or considering which signals to use. Please contact us via this form.

References

• ISO 3685:1993 project information (title and scope)
• ISO 8688-1:1989 project information (title and scope)
• Materials (2025): example of a VBB 0.3 mm life criterion based on ISO 3685
• Materials (2025): tool life defined as VB 0.3 mm based on PN-ISO 3685
• Materials (2025): replacement above 0.3 mm following ISO 3685 guidelines
• Materials (2023): example of VB3 0.5 mm under ISO 8688-2
• Materials (2024): choosing a life criterion under ISO 8688-2
• Data in Brief (2023): public dataset with 9 tools and 3 machines
• Sensors (2026): review of sensor technologies for machining
• PHM Society: 2010 Data Challenge
• Scientific Reports (2024): tool wear prediction with the PSD-CVT model
• Scientific Data (2025): MU-TCM face milling dataset
• NASA: Milling Wear dataset
• FANUC: Cutting Load Monitor Function (June 2022)
• Cutting Tool Engineering: Marposs Genior Modular (August 8, 2017)
• Marposs: acquisition of ARTIS
• MONOist: Okuma OSP-AI Machining Diagnosis (November 16, 2018)
• MachineMetrics: announcement of Tool Anomaly Detection (March 4, 2025)
• Caron Engineering: IMTS 2026
• Siemens: expansion of Sinumerik Edge apps (May 14, 2019)
• Capital.com (Reuters): Thailand’s August 2026 car production (September 28, 2026)
• BOI: document dated August 4, 2026
• BOI: Investment Promotion Guide 2025
• Mondaq: commentary on the renewed BOI incentives (2026 to 2027)
• Thailand Business News: growth in AI adoption in Thailand (June 10, 2026)
• IMTS: announcement of IMTS 2026 programs (July 14, 2026)
• Tokyo Big Sight: JIMTOF 2026 event notice (June 19, 2026)
• ORNL: tool wear detection with a tactile sensor (May 6, 2026)