Fleet dashboards are good at telling you what happened. Battery levels, coverage per shift, error logs, mission completions, distance travelled — the data is plentiful and the visualisations are pleasant. The problem is that all of it describes the past. A dashboard that tells you a brush motor drew more current yesterday has not given you a maintenance decision; it has given you a chart. The transition from monitoring to prediction is a specific and much narrower exercise: finding the channels whose movement precedes a failure by a useful interval, establishing a threshold that fires early enough to act on, and confirming the intervention costs less than the failure it prevents.

This guide does that exercise for cleaning robot fleets specifically, because the failure modes on a floor scrubber are different from those on a manipulator or a vehicle, and the channels that matter follow from those modes.

The Failure Modes You Are Actually Predicting

Predictive maintenance only works against failure modes that have a measurable degradation path. A controller board that fails instantly from a solder defect has no precursor and no telemetry will anticipate it. Everything else on a cleaning robot — the ones driven by wear, contamination, alignment drift or thermal stress — does have a path, and there are six that account for most unplanned stops on commercial fleets.

Failure Mode	Degradation Path	Time From First Signal to Failure	Predictable?
Brush motor brush wear	Commutation degrades, current draw rises for the same load, then thermal trip	80–200 operating hours	Yes — current and temperature trend
Squeegee blade wear and deformation	Recovery efficiency drops, wet streak appears, vacuum load falls	60–150 hours	Yes — recovery-to-applied water ratio
Filter and recovery tank blockage	Vacuum motor works harder, airflow falls, suction drops before trip	20–60 hours	Yes — vacuum motor current at constant speed
Battery capacity fade	Usable capacity declines per cycle; charge time to 80% shortens	Months — an order of magnitude ahead	Yes — capacity-per-cycle trend is the cleanest signal on the platform
Wheel and drive assembly wear	Position error grows, drive current rises for the same path, odometry drifts	100–300 hours	Yes — but requires a consistent reference route
Sensor contamination	Lidar return intensity drops, localisation confidence falls, map mismatches increase	10–40 hours	Partially — confidence is confounded by site changes

The time-to-failure column is the one that determines whether prediction is worth anything. A signal that surfaces six months ahead of a battery failure lets you plan capital. A signal that surfaces twenty hours ahead of a filter blockage is genuinely useful for scheduling, because it converts a mid-shift breakdown into a planned intervention at shift end. A signal that surfaces ten minutes ahead of anything is just an alarm, and an alarm is not predictive maintenance.

The Signals With Real Predictive Value

Of the channels a fleet reports, roughly a dozen carry predictive information. The rest are lagging indicators — they describe a state that has already been reached, which is useful for reporting and useless for prevention. The distinction is not about which channel it is, but whether the channel moves before the failure or because of it.

Signal	Type	Threshold to Watch	Leading Typical Failure By
Brush motor current at fixed load	Leading	+15–20% above commissioning baseline	60–120 h
Brush motor temperature rise over ambient	Leading	+12 °C over commissioning delta	20–50 h
Vacuum motor current at constant speed	Leading	+20% sustained across three shifts	20–40 h
Water applied to water recovered ratio	Leading	Fall of more than 8 points from baseline	30–80 h
Usable capacity per full charge	Leading	Decline of more than 12% from first-year median	3–9 months
Charge time to 80% state of charge	Leading	Shortening by more than 25%	1–6 months
Drive current per metre at constant speed	Leading	+18% on a repeatable route	80–200 h
Localisation confidence variance	Leading, noisy	Sustained 2σ drop on unchanged routes	5–30 h
Distance and hours since last service	Context	Not a threshold — a multiplier on every other signal	n/a
Mission completion rate	Lagging	Any fall is already a symptom	Already failed
Error log frequency	Lagging	Counts past events; useful for triage, not prediction	Already failed
Coverage per shift	Lagging	Affected by schedule and site, not just machine condition	Already failed

The bottom three rows are where most fleet analytics effort is wasted. Mission completion rate, error frequency and coverage are the metrics that appear on dashboards because they are easy to compute and satisfying to look at, and none of them predict anything. By the time completion rate falls, the failure has occurred. A predictive programme built on them is an alarm system with a history function.

Turning a Threshold Into a Decision

A threshold is not a decision. The gap between "brush motor current is up 18 percent" and "replace the brush assembly this week" requires two more quantities: the expected remaining operating hours before failure, and the cost comparison between intervening now and failing in service. The first comes from the degradation rate on your own fleet — how many operating hours the signal took to move from baseline to threshold — projected forward to the failure point. The second is the arithmetic that most programmes skip.

Scenario	Direct Cost	Consequential Cost	Total
Planned intervention (replace brush assembly at threshold)	Item €180 + 1.5 h labour €135	Scheduled at shift end; zero coverage lost	€315
Run to failure (motor thermal trip mid-shift)	Item €180 + emergency callout €260 + 3 h labour €270	Shift coverage lost on 3,000 m² at €0.42/m² = €1,260, plus a substitute manual pass	€1,970+
False positive (replace at threshold, item had 40 h left)	Item €180 + 1.5 h labour €135	None — the part was consumed early, not discarded	€315, of which €60–90 is premature

The table contains the argument for predictive maintenance on cleaning fleets, and it is not the parts cost. The exposed cost is the consequential one: a machine that trips mid-shift on a committed cleaning programme creates a coverage gap with a contractual consequence, and that gap is typically four to six times the value of the planned intervention. The same arithmetic also shows why false positives are tolerable. A spurious intervention costs the labour and the remaining life of a cheap consumable. A missed prediction costs a disrupted shift. The asymmetry means a fleet programme should tolerate a relatively high false-positive rate and should not be tuned to minimise false alarms at the expense of missed ones — the opposite of the instinct that tends to drive alarm tuning. The downtime-cost framework behind those numbers is developed in downtime cost and OEE.

Getting the Baselines Right

None of the thresholds above are absolute, and treating them as absolute is the second most common way these programmes fail. Every threshold is a deviation from a per-unit, per-site baseline, and the baseline has to be captured correctly at commissioning or it will never be trustworthy.

Four requirements make a baseline usable. First, it must be captured under defined conditions: a specific floor type, a specific pass pattern, a specific soil load category. Brush motor current on a smooth sealed concrete floor and on a rough unsealed surface differ by more than the degradation you are trying to detect, so a baseline that averages across floor types has no predictive power. Second, it must be captured in the first days after commissioning, before wear begins but after any run-in period. Third, it must be per unit rather than per fleet model, because manufacturing variation between two machines of the same type can exceed the degradation signal over the early months — a fleet-wide threshold will produce persistent false positives on the machine at the high end of the distribution. Fourth, it must be re-captured after any change with a mechanical consequence: brush replacement, squeegee change, motor replacement, wheel assembly work. The hours immediately after a component change belong to the new component, and carrying forward the old baseline will show a step change that is real but not a fault.

The practical consequence is that the first month of fleet operation should be treated as a baseline campaign rather than a performance period. Fleets that skip it spend the following year arguing with false alarms generated by comparing machines to each other instead of to themselves. The measurement conventions that make those comparisons defensible, including how to normalise for area and pass count, are set out in service robot KPI benchmarks.

What Telemetry Cannot Predict — and Why That Matters

Being explicit about the limits prevents a fleet programme from over-promising internally and then being judged against the promise. Four categories of failure are effectively invisible to onboard telemetry.

The first is any sudden electronic failure: a controller, sensor board or power stage that fails without a precursor. No degradation path, no prediction, and the only mitigation is spares and a response SLA. The second is site-caused damage: impact, cable ingestion, water ingress at a dock, a door closing on a machine. These are discrete events, not trends, and the useful data is post-hoc for triage rather than predictive. The third is failures originating in work the operator does not report — a machine that is handled roughly during manual intervention degrades faster than its signals suggest, and since nothing in the fleet reports how it was manhandled, the model is blind to the cause and will over-attribute the degradation to component life. The fourth is anything caused by a firmware or configuration change, where the machine is not degrading at all but behaving differently by design.

The honest framing for a fleet programme is that predictive maintenance covers the wear-driven majority of unplanned stops and does nothing for the sudden minority. That split is still a favourable trade, because the wear-driven failures are the ones that occur mid-shift and carry consequential cost, while sudden electronic failures are frequently covered under a parts contract. This is where predictive maintenance and contract coverage are complements rather than substitutes, which is the analysis in service contract versus warranty economics.

What AOMAN's Platform Reports

AOMAN FUTURE's cleaning platforms — the C1 large-format scrubber and the C2 Pro compact cleaner — expose the leading signals rather than only the completion statistics. Per-unit telemetry includes brush and vacuum motor current and temperature, water applied and recovered volumes, drive current, charge capacity per cycle, charge time to state of charge, and localisation confidence, logged per mission with the route and floor zone attached.

The design decision worth noting is that baselines are captured per unit per zone at commissioning rather than assigned from a fleet-wide model, which is the requirement that makes the thresholds usable without a false-alarm campaign. Motor current and water recovery are logged continuously rather than sampled at mission end, so a trend is visible within a shift instead of a week. What AOMAN does not provide is an automated prediction service that will tell you when to replace a component; the platform supplies the signals and the per-unit baselines, and the interpretation — including the failure-versus-intervention cost comparison above — remains yours or your service partner's. That boundary is deliberate, because the intervention economics depend on your labour rates and your coverage obligations, not on the machine. If you want the baseline captured during a trial on your own floors, request a telemetry baseline assessment and we will report the leading signals for your site class alongside the machine's consumption figures. For the wider operating-cost context, see the failure modes and business continuity guide and the battery and charging technology overview.