Predictive Maintenance Without Sensors: A Small Factory Path

Picture a twenty-person plastic injection shop. The maintenance chief keeps a worn notebook in a drawer: which machine got fresh oil and when, which bearing failed last winter, which press was overhauled "once it started making that noise." That notebook is a treasure. The problem is that nobody except the chief can read it, and when he goes on holiday, the machines' entire history goes on holiday with him.

Say "predictive maintenance" to most small factory owners and they picture expensive sensor arrays, cloud platforms, and projects built for large plants. Yet the first step of this journey is not buying sensors; it is moving that notebook into a system. In this article we walk through how a small factory can progress from maintenance records to failure prediction step by step, what each stage roughly costs, and the limits that vendor slide decks rarely mention. For a broader view of which AI applications fit which type of business, see our industry-by-industry AI map.

What is predictive maintenance, and how does it differ from preventive?

Predictive maintenance uses data about a machine's actual condition (vibration, temperature, current draw, past failure records) to estimate when a failure is coming, then schedules the repair inside that window. Preventive maintenance is calendar-driven: every machine gets serviced at fixed intervals regardless of its condition. In short, preventive says "every six months"; predictive says "when needed, and just before it is needed."

The two approaches also rest on different assumptions. Calendar-based maintenance assumes failure arrives with time and usage. In reality, many failure modes are triggered not by the calendar but by stress: load spikes, installation errors, lubrication problems. That is why calendar maintenance wastes money in two directions at once: perfectly healthy parts get replaced, while off-schedule failures still arrive as surprises.

The market reflects the shift. Estimates from several research firms converge on a global predictive maintenance market of roughly 14-15 billion dollars as of 2025, growing at more than 25 percent a year. The driver is simple: unplanned downtime is one of a manufacturer's most expensive line items, and studies estimate it costs global manufacturing tens of billions of dollars every year.

Can you do predictive maintenance without sensors?

Yes, at the entry level. If your maintenance records, work orders, and failure history are entered into a digital system consistently, machine learning can find patterns in those records and flag which machines are prone to which failures. The first investment goes into record discipline, not hardware; sensors are a later stop on the route.

The industry even has a name for this road: starting from a computerized maintenance management system (CMMS). IBM, among others, lists work orders and maintenance logs alongside sensors as core data sources for predictive maintenance. One CMMS vendor claims, based on its own customer data, that small manufacturers with 20-200 assets see downtime drop 25-30 percent in the first year using record data alone. Treat the number as a vendor claim, but take the underlying message: a shop that keeps records starts seeing failures earlier than a shop that does not.

The most common obstacle we see in practice is not technical; it is habit. The chief's knowledge that "the press sounded different, so I adjusted it" makes it into the notebook loosely and into the system not at all. Yet those lines are exactly the raw material a prediction model needs. A model fed sparse or messy records produces false alarms, and after two false alarms the team stops trusting the system. That is why the goal of the first three months is not prediction; it is getting every intervention logged as a standard work order.

The small factory journey: three stops

The typical route passes through three stops: from firefighting to digital records, from records to condition-based rules, and finally to sensor-backed prediction on critical machines. Each stop pays for itself on its own, which means you start earning a return well before reaching the end.

Stop 1: from notebook to system. Every machine's identity, maintenance history, and failure log moves into one system. There is no AI at this stage; there is discipline. The gain is still concrete: for the first time you see, in numbers, how many hours each machine was down last year and how often each part was replaced. Spare parts get ordered from history rather than guesswork.

Stop 2: from records to rules. The accumulated data starts talking: alerts like "this press's bearing has been replaced every 4,000 hours on average over three years, and it is now at 3,600." A creeping increase in an injection machine's cycle time is another early signal; for many shops, this stage is the first time production data and maintenance data sit side by side. We covered another use of production-line data in our guide to reducing scrap rate with production data; the underlying data plumbing is the same.

Stop 3: adding sensors to the rules. Vibration and temperature sensors go onto the one to three most critical machines, and the model now listens to their live condition as well. A bearing's vibration signature changes weeks before it lets go; the system catches it and schedules the repair for a weekend shutdown. This is the version of predictive maintenance the brochures describe, but as you can see, it is the third stop, not the first. If you want a primer before getting this far, our starter guide to predictive maintenance and AI quality control for small manufacturers is a good place to begin.

What does predictive maintenance cost?

Entry costs are lower than most people expect. Digital maintenance software is typically priced per user, and sources quote entry-level plans from roughly 400 dollars per user per year. On the hardware side, temperature sensors start around 100 dollars, while vibration sensors can run around 1,000 dollars each. A three-user system plus sensors on two critical machines can often be set up for less than the cost of a single unplanned stop on one production line.

Those figures come from individual sources and vary a lot by configuration, so no budget is final until you have quotes. Still, the order of magnitude holds: this is not a five-figure industrial IoT project. Vendor content often mentions payback in 6-18 months; treat that with the same caution, because the only payback that matters is the one computed with your own downtime cost. It is also worth checking public support: many governments subsidize digitalization investments for small manufacturers, and maintenance digitization frequently qualifies.

What does one shift of downtime cost? A sample calculation

Downtime cost is the single real compass for this investment: add up the lost output, the idle labor, and the cost of the delayed order for every hour the machine stands still. A business that does not know this number makes maintenance decisions by feel; a business that knows it decides with arithmetic.

A hypothetical but realistic example: an injection machine producing 2,500 dollars of output per hour goes down for one 8-hour shift with a failed main bearing. Lost production: 20,000 dollars. If a three-person crew stood idle or worked overtime to catch up, add a few thousand more. Emergency parts and an outside service call usually cost more than a planned purchase, and the goodwill gesture to the customer whose order slipped is on top. A single unplanned stop easily lands in the 25,000-30,000 dollar range. For a shop that takes three or four of these hits a year, the entry investment above suddenly looks very different.

The same arithmetic has a quiet flip side: calendar maintenance is downtime too. Stopping a perfectly healthy machine to swap parts because the schedule said so costs both parts and production. The promise of the predictive approach is to shorten both ends at once: fewer surprise stops, and fewer unnecessary planned ones.

Can every failure be predicted?

No, and knowing this up front protects you from both disappointment and misdirected spending. Predictive maintenance works where a failure announces itself in advance: degradation begins, gives off a signal, and the machine fails some time later. The window between those two points (maintenance literature calls it the P-F interval) has to be wide enough to act in. Where there is no window, or it is too narrow, even the most expensive sensor cannot show the failure coming.

Window width varies enormously by failure mode. A pump bearing's decline can be caught months ahead through vibration measurement, while a crack driven by sudden load can take a machine down within a few hundred operating hours, and electronic board failures often give no warning at all. So the right question is not "how do we predict every failure" but "which of our failures are predictable." Look at your failure history: if the three to five failure types causing most of your downtime are the predictable kind, the focus of the investment picks itself.

One warning about the numbers in circulation: the oft-repeated claims that predictive maintenance "cuts maintenance cost 25-30 percent and failures 70-75 percent" trace largely back to a decades-old U.S. Department of Energy study, quoted context-free in every brochure since. They are not necessarily wrong, but they are a direction sign, not a commitment to your plant.

Frequently asked questions

How many machines should a pilot cover?

Industry practice points to starting with one to three machines, chosen for criticality or failure frequency. The aim is a measurable result within six months: one early warning caught, one stop avoided. Rolling out plant-wide before seeing a result risks both the budget and the team's trust.

Is existing ERP and production data enough?

Usually yes, for a start. Downtime logs, production counts, and the maintenance module in your ERP are enough to show which machines are the troublemakers. What usually falls short is not the data but its tidiness: the same failure logged once as "bearing" and once as "making noise" blinds the model. A standard failure-code list is the cheapest, highest-impact step you can take before any sensor.

Which sensors are needed?

The common trio: vibration for rotating equipment (motors, pumps, fans, bearings), temperature for heat-related failures, and current measurement for electrical problems. Which one you need is dictated not by the machine but by your failure history: start with the sensor that would have caught your most frequent failure type.

What should you do?

• Compute your downtime cost. Work out, once and concretely, what one lost shift on your most critical machine costs (lost output, overtime, delayed orders). Every investment decision leans on this number.
• Move the notebook into a system. Digitize maintenance records before buying sensors; every intervention becomes a standard work order. Make that discipline the sole target of the first three months.
• Classify your failure history. Sort the last two years of failures by type: which gave advance signals, which arrived cold? The predictable ones go to the top of the investment list.
• Keep the pilot small. Start with the one to three machines that stop most often, measure for six months, expand if it works. Instrumenting the whole plant in one go is both expensive and unnecessary.
• Check local support programs. Digitalization subsidies for small manufacturers exist in many countries, and maintenance digitization often qualifies.

That notebook in the drawer is your factory's most valuable dataset, written out by hand. The day you move it into a system, the predictive maintenance journey has begun; sensors, models, and forecasts are built on top of that foundation. If you would like to talk through where your own machine park should start, drop us a line.