[Manufacturing](https://bfcs.ai/industry/manufacturing/) · [Better](https://bfcs.ai/better/)

# Can AI Predict Maintenance Needs?

Much factory maintenance still runs on two clocks: the breakdown and the calendar. Condition monitoring offers a third, servicing a machine when its own signals say it needs it. The evidence favours the shift, but it is thinner than the sales pitch, and the hard part is the plant, the data and the people, not the model.

Updated **October 2026** · 7 min read · Evidence-graded · 7 sources

## The BFCS verdict

Every article scored on the same four dimensions

| Dimension | Verdict | Why |
| --- | --- | --- |
| Better | Partly · Low to moderate confidence | Evidence beats the calendar; the AI layer is less proven · In a small US survey, plants relying less on reactive maintenance lost about half as much production time to downtime, but its authors call the comparison anecdotal. A 2022 review found only 44 machine learning studies tested on real industrial manufacturing cases from 2015 to 2021. |
| Faster | Partly · Low to moderate confidence | More warning, slower to set up · A detected fault gives planners time to book parts and labour before the line stops. Surveyed US manufacturers reported difficulty setting baselines and too many false alarms; in our reading that work, not fitting sensors, sets the pace. |
| Cheaper | Conditional · Low confidence | Costs rise before they fall · Surveyed US plants leaning on predictive maintenance spent more on direct maintenance relative to shipments, not less. Savings come from avoided downtime and defects, and they can take longer to show than a pilot lasts. |
| Safer | Conditional · Governance-gated | Fewer surprises, if the alert has an owner · We found no factory study measuring injury outcomes from predictive maintenance. Fewer unplanned breakdowns and fewer unnecessary teardowns reduce exposure only when alerts are triaged and every intervention still follows isolation procedures. |

## The problem, in numbers

Maintenance is a large cost, and one a manufacturer can influence. Using a survey of US manufacturers, researchers at the National Institute of Standards and Technology (NIST) estimated total annual maintenance costs and losses at $222.0 billion on average, in 2016 dollars, with about half of it lost sales. (Source 1) The cost is not only the repair. It is the order that ships late and the batch that comes off a drifting machine out of tolerance.

Much of that maintenance is still done after something breaks. Across the surveyed establishments, maintenance was on average 45.7% reactive, 31.8% preventive and only 17.3% predictive. (Source 2) The half of respondents leaning hardest on reactive work lost a mean 10.38% of planned production time to downtime, against 4.91% for everyone else. (Source 1)

### How US manufacturers maintain their machinery (Source 2)

Reactive · 45.7%

Preventive · 31.8%

Predictive · 17.3%

Average share of maintenance by type across surveyed US manufacturing establishments (NIST, 2016 data; other types make up the remainder). Predictive maintenance was the smallest of the three.

### The evidence, in four numbers

45.7% · of machinery maintenance was reactive · average across surveyed US manufacturing establishments; 17.3% was predictive (NIST, 2020)

10.38% vs 4.91% · planned production time lost to downtime · heaviest reactive users vs the rest, 71 US respondents; the authors treat it as anecdotal (NIST)

77% to 92% · of failures were random, not age-related · three older studies (labelled UAL 1968, Sweden 1973, US Navy 1982) as summarised by NASA, 2008; not presented as factory data

44 of 4,549 · records kept as primary studies in a review of machine learning on real factory cases · records published 2015 to 2021, duplicates included; inclusion required empirical results from an industrial case (Applied Intelligence, 2022)

## What improves, and why

Preventive maintenance puts the calendar in charge: change the bearing every so many hours whether it needs it or not. That works when failure tracks age. Often it does not. NASA's maintenance guide summarises three older studies, labelled UAL (1968), Sweden (1973) and US Navy (1982), in which random failures made up 77% to 92% of the total and age-related failures only 8% to 23%. It adds that in many cases scheduled overhaul raises the failure rate by introducing early-life faults into equipment that was running well. (Source 3) Those studies are decades old and NASA does not present them as factory data, but in our reading the logic carries: a fixed interval cannot see a fault that arrives at random.

Condition-based maintenance waits for evidence instead. NASA's guide draws it as a P-F curve. In the usual reading, a fault starts, becomes detectable as a potential failure (P), and only later becomes a functional failure (F) that stops the machine. The guide notes that the point where degradation starts and the point where it is detected rarely coincide, and advises setting monitoring intervals conservatively at first. (Source 3) The window between P and F is the lead time a planner can use to order parts, book a fitter and stop the line at changeover rather than mid-run.

Machine learning extends what can be read inside that window. In principle, instead of one alarm threshold per sensor, a model can combine vibration, temperature, current and process signals and learn what normal looks like for each machine. It can also make the trade-off explicit. One peer-reviewed method, tested on a simulated example and a semiconductor manufacturing benchmark, sets the balance between unexpected breaks and unused component life as a cost decision rather than a fixed rule. (Source 4)

> The calendar replaces parts that were fine and misses faults that arrive early. Condition monitoring is only better if someone trusts the signal enough to act on it.

## What the evidence actually says

The manufacturing evidence is suggestive rather than conclusive. In the NIST survey of 71 usable responses, respondents relying less on reactive maintenance had 52.7% less unplanned downtime and 78.5% fewer defects. Among those, the ones leaning most on predictive maintenance had 18.5% less downtime and 87.3% fewer defects than those leaning on preventive. (Source 1) But they also spent 212.3% more on direct maintenance relative to their shipments, and reported higher lost sales. (Source 1) The authors say the comparisons should be treated as anecdotal because the groups are small. (Source 1) Their earlier report also found that plants relying less on reactive work were more likely to make to order and to compete on differentiation rather than cost, (Source 2) so in our reading the kind of business may explain part of the gap.

On machine learning specifically, the base is narrower than the volume of papers implies. A 2022 systematic review screened 4,549 records from 2015 to 2021 and selected 44 primary studies, a set limited to work with empirical results from an industrial manufacturing case. The equipment ranged from CNC machine tools and injection moulding machines to robot arms and conveyor motors. Only two studies used online learning, and the authors named loss of model performance as the data shift over time (concept drift) as the most important limitation. (Source 5)

An earlier university pilot survey of 15 manufacturers and 8 technology or consulting firms, with NIST case studies, helps explain why adoption lags. It found very few predictive or proactive practices in use, although the paper reports 72.7% of the surveyed manufacturers considering condition-based maintenance, without saying how many answered. The common barriers were cost, workforce skills and organisational and technology readiness; manufacturers also reported too many false alarms, technology providers pointed to a lack of failure data, and some smaller firms' operators distrusted the technology. (Source 6) One aerospace manufacturer ran a machine tool health monitor for a year, saw no failures, switched it off to save production time, and then had a similar axis failure. (Source 6)

## Where it falls short

-   **Few failures to learn from.** In our reading, a well-maintained plant has little failure history on record, and technology providers name the lack of failure data as a barrier. (Source 6)
    
-   **Costs come first.** NASA's guide notes that test equipment, software, training and sometimes new staff raise spending for two to three years on average. (Source 3) A year-long pilot may see no failure at all. (Source 6)
    
-   **Models drift.** Factory data change over time, so what normal looks like shifts, and the review names the resulting loss of model performance as the main limitation. (Source 5)
    
-   **Skills are scarce.** In Ai Group's 2024 survey of 251 Australian businesses, 79% reported difficulty finding or training technicians and trades staff, though the release does not state that share's base or break out manufacturing. (Source 7)
    

## Where governance decides the outcome

Moving from calendar to condition changes who decides when a machine is serviced. That decision needs a named owner, documented thresholds for pulling equipment, and a record of every alert and what was done about it. Deferring an alert to protect a production run should be a visible decision, not a quiet one. Retire a calendar task only when the monitoring that replaces it has been shown to catch the same failure mode, and keep statutory inspections on their schedule. A predictive alert changes when work happens, not how it is done: isolation and lockout apply exactly as before. Sensor changes and model retraining need change control, because a model that has stopped seeing a fault is worse than the calendar it replaced.

## How to think about adoption

Start with a handful of machines where an unplanned stop is expensive and the failure gives a measurable early sign, such as spindles, gearboxes, pumps and motors. Record a baseline first: downtime as a share of planned production time, defect rate, and the share of maintenance that is reactive, so any change can be tested on your own floor. Fix work order coding before buying analytics, because the model learns from the failures you wrote down. Budget for skills as well as sensors, and run any pilot long enough to see failures, or agree in advance how its value will be judged if none occur. Treat vendor uptime claims as hypotheses to test, not results.

## How we read this evidence

BFCS.ai does not fabricate figures. The strongest manufacturing evidence here is a small US survey whose authors call their group comparisons anecdotal, and the failure-pattern studies are decades old, from an airline, a Swedish study and the US Navy, as summarised by NASA; we label both and grade confidence accordingly.

## Sources

1.  Thomas, D. & Weiss, B. (US National Institute of Standards and Technology), "Maintenance Costs and Advanced Maintenance Techniques in Manufacturing Machinery: Survey and Analysis," International Journal of Prognostics and Health Management 12(1), April 2021 (peer-reviewed, open access). [doi.org](https://doi.org/10.36001/ijphm.2021.v12i1.2883)
    
    **What we took from it:** Machinery Maintenance Survey of US manufacturers: 85 responses, 71 used. Total annual maintenance costs and losses averaged $222.0 billion (2016 dollars, Monte Carlo estimate), about half from lost sales. The top 50% of respondents by reactive maintenance spending lost a mean 10.38% of planned production time to downtime against 4.91% for the rest (Monte Carlo means, Table 6); the rest had 52.7% less unplanned downtime and 78.5% fewer defects. Among the rest, those leaning most on predictive maintenance had 18.5% less downtime and 87.3% fewer defects than those leaning on preventive, but 212.3% higher direct maintenance costs relative to shipments and higher lost sales. The authors say these comparisons should be treated as anecdotal evidence because the groups are small.
    
2.  Thomas, D. S. & Weiss, B. A., "Economics of Manufacturing Machinery Maintenance: A Survey and Analysis of U.S. Costs and Benefits," NIST Advanced Manufacturing Series 100-34 (June 2020), the technical report behind footnote 1, cited for figures not repeated in the journal version. [doi.org](https://doi.org/10.6028/NIST.AMS.100-34)
    
    **What we took from it:** On average, establishment maintenance practices were 17.3% predictive, 31.8% preventive and 45.7% reactive; establishments relying less on reactive maintenance were more likely to make to order and to compete as differentiators rather than on cost.
    
3.  US National Aeronautics and Space Administration, "Reliability-Centered Maintenance Guide for Facilities and Collateral Equipment" (September 2008). [nasa.gov](https://www.nasa.gov/wp-content/uploads/2023/06/nasa-rcmguide.pdf)
    
    **What we took from it:** Figure 4-2 (P-F curve): the point of initial degradation and the point of initial detection rarely coincide, and monitoring intervals should be set conservatively at first. Section 4: in three studies (labelled UAL 1968, after the Nowlan and Heap work; Sweden 1973; US Navy 1982) random failures were 77% to 92% of total failures and age-related failures 8% to 23%; in many cases scheduled overhaul increases the failure rate by introducing infant mortality. Section 3: maintenance costs typically rise at first under a new reliability-centred maintenance program, for an average of two to three years, because of test equipment, software, training and sometimes new staff. The guide is written for NASA facilities and collateral equipment and does not describe the three studies as factory data.
    
4.  Susto, G. A., Schirru, A., Pampuri, S., McLoone, S. & Beghi, A., "Machine Learning for Predictive Maintenance: A Multiple Classifier Approach," IEEE Transactions on Industrial Informatics 11(3), 2015, pp. 812-820 (peer-reviewed; abstract only, read on the Queen's University Belfast repository). [doi.org](https://doi.org/10.1109/TII.2014.2349359)
    
    **What we took from it:** Classifiers trained on different prediction horizons trade off the frequency of unexpected breaks against unexploited lifetime, feeding a cost-based decision rule; demonstrated on a simulated example and a benchmark semiconductor manufacturing maintenance problem.
    
5.  Fernandes, M., Corchado, J. M. & Marreiros, G., "Machine learning techniques applied to mechanical fault diagnosis and fault prognosis in the context of real industrial manufacturing use-cases: a systematic literature review," Applied Intelligence 52(12), 2022, pp. 14246-14280 (peer-reviewed; free full text on PubMed Central, PMC8894092). [doi.org](https://doi.org/10.1007/s10489-022-03344-3)
    
    **What we took from it:** Of 4,549 records published January 2015 to October 2021 (3,377 without duplicates), 44 primary studies met its criteria, which required empirical results from an industrial case; equipment includes CNC machine tools, injection moulding and cold forging machines, robot arms and conveyor motors; only two studies used online learning methods, and the authors name loss of model performance under concept drift as the most important limitation.
    
6.  Jin, X., Weiss, B. A., Siegel, D. & Lee, J., "Present Status and Future Growth of Advanced Maintenance Technology and Strategy in US Manufacturing," International Journal of Prognostics and Health Management 7 (Special Issue on Smart Manufacturing PHM), 2016, article 012 (peer-reviewed; read as the author manuscript on PubMed Central). [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC5207222)
    
    **What we took from it:** Pilot survey by University of Cincinnati and University of Michigan researchers of 23 companies (15 manufacturers, 8 technology or consulting firms), 2015, plus NIST case studies of smaller firms. Very few predictive or proactive practices were adopted by the surveyed manufacturers; the paper gives 72.7% of the surveyed manufacturers as considering condition-based maintenance without stating the base (72.7% is not a whole number of 15); common barriers were costs, workforce and skill levels, organisational and technology readiness and complexity of system design changes; manufacturers reported unsatisfactory false alarms, technology providers noted a lack of failure data, and some smaller firms' operators distrusted the technology. An aerospace manufacturer ran a machine tool health monitoring system for a year, stopped it after no failures occurred, then had a similar axis failure.
    
7.  Australian Industry Group, Centre for Education and Training, "Technical and trade skills in high demand and hard to find in 2024" (media release, 10 June 2024), reporting its survey Listening to Australian businesses on workforce and skills 2024: [australianindustrygroup.com.au](https://www.australianindustrygroup.com.au/news/media-centre/2024/technical-and-trade-skills-in-high-demand-and-hard-to-find-in-2024/)
    
    **What we took from it:** 251 businesses responded; 77% reported an increased need for technicians and trades skills and 79% reported difficulty finding or training such staff, up from 39% in 2020. The release does not state whether the 79% is of all respondents or of those reporting increased need, and does not break out manufacturing or maintenance roles.
    

Putting this to work

## Working through this in your own organisation?

BFCS.ai is published by Inhouse CX. We help small and mid-sized organisations decide where AI earns its place, and put the governance in place that makes it safe to use.

[Talk to Inhouse CX →](https://inhouse.cx/engage/)

Continue the thread

- [Can AI Improve Manufacturing Quality?](https://bfcs.ai/articles/can-ai-improve-manufacturing-quality/): More in Manufacturing · MANUFACTURING · BETTER · 6 min · Read →
- [Can AI Make Factories Safer?](https://bfcs.ai/articles/can-ai-make-factories-safer/): More in Manufacturing · MANUFACTURING · SAFER · 6 min · Read →
- [Can AI Improve Bid Accuracy?](https://bfcs.ai/articles/can-ai-improve-bid-accuracy/): Also Better · CONSTRUCTION · BETTER · 7 min · Read →

---

Canonical page: https://bfcs.ai/articles/can-ai-predict-maintenance-needs/
This is the Markdown version of the page, generated from its HTML at build time.
Site guide for agents: https://bfcs.ai/llms.txt · Every page: https://bfcs.ai/sitemap.xml