A manufacturer should log a quarter of maintenance downtime by machine before buying AI that predicts machine failures, put the first model on the machine whose sensor-detectable failures cost the most, and score its alerts for a second quarter before they become planned jobs.
A June 2020 NIST report defines predictive maintenance as maintenance initiated based on predictions of failure made using observed data such as temperature, noise, and vibration. Its Machinery Maintenance Survey, used for 2016 estimates, asked managers of machinery maintenance to estimate what percent of planned production time was downtime and what percent of that was due to reactive maintenance, meaning unplanned maintenance and repair, called maintenance downtime here. NIST notes the survey asks about issues many firms do not formally track.
A journal version published April 25, 2021 counts 71 respondents at US discrete manufacturing establishments, excluding food, beverage and tobacco, textile, apparel and leather, petroleum and chemical makers. In the 2020 report, the top 25 percent by reactive share of maintenance spending estimated, on average, 6.7 percent of planned production time lost to maintenance downtime, against 0.9 percent for the bottom 25 percent, which averaged 3.3 times the top group's shipments and relied more on preventive and/or predictive maintenance, so a preventive schedule alone can place a plant there. NIST found that gap statistically significant at the 10 percent level, calls many other group differences anecdotal, given small samples, and says lower losses may not be completely caused by advanced maintenance.
For a 13-week quarter, the maintenance manager should close every unplanned work order with the machine, the hours it stood, whether a maintenance issue caused it and which part failed. The plant controller divides each machine's maintenance downtime by its planned production hours, a share an operating partner can compare across portfolio plants logging the same fields on similar machines. The controller then prices those hours at the cost of idle energy, labor and capital, as NIST does, or a lost order's contribution margin where larger, plus expediting and penalties on late orders, the production planner tying each to its stop.
The CFO funds one model, on the machine whose sensor-detectable failures, as the maintenance manager judges them, cost the most. Its fixed fee, for a stated term, itemizes hardware and fitting, paid upfront and capitalized, and software and training, paid from the maintenance budget only after a second-quarter pass; the plant keeps hardware, model and data.
The second quarter runs 13 weeks from a start date in the fee, longer if maintenance stops fall below a minimum; the plant manager writes that minimum and the pass marks into the fee, the maintenance manager the window for confirming an alert. The machine keeps its maintenance schedule; the maintenance manager logs each alert and a technician checks the machine within the shift, repairing at once what is unsafe; each scheduled job records any fault an open alert named. The controller counts alerts whose named fault a stop, check or scheduled job confirmed inside that window, alerts nothing confirmed, and maintenance stops no alert named; the plant manager switches the model off on any missed pass mark. From the third quarter the maintenance manager turns alerts into planned jobs, dropping scheduled tasks they duplicate. Each quarter the controller totals the machine's maintenance spend, quarterly fee share and planned and unplanned maintenance stops priced as in the first quarter, per planned production hour; the CFO funds the next costliest machine only once that total falls below the second quarter's, which carries no fee share.
NIST says predictive maintenance usually requires a larger upfront investment than preventive maintenance.
