Predictive maintenance for agency fleets, facilities, and equipment

Reference article · 6 min read

Predictive maintenance means using operating data to estimate when a specific asset is likely to fail, so the repair is scheduled before the failure rather than after it. In a public works or general services department, the asset in question might be a refuse truck transmission, a sewer lift station pump, a rooftop air handler, a standby generator, or a chlorine feed pump. The appeal is straightforward: fewer road calls, less emergency overtime, fewer bypass events, and parts ordered before the unit is down. The difficulty is that most of the value depends on data the agency already owns and has probably never curated, and on a workflow that ends in a dispatched work order rather than a screen somebody checks on Tuesdays.

It also helps to be honest about the comparison being made. The alternative to a predictive model is rarely doing nothing. It is a preventive maintenance schedule that could simply be tightened, or a condition inspection that could be added at low cost. A model has to beat that, not beat neglect.

The data that has to exist before a model is worth building

Vendors will ask what systems you have. The more useful question is what condition the contents are in.

• An asset register that matches reality. Every unit needs a stable identifier, a class, a make and model, an install or in-service date, and a location. If the same pump appears three times under three spellings, nothing downstream works.
• Work order history tied to those identifiers. Several years is better than several months, because failures are rare events and a model learns from the failures, not from the routine services.
• Failure labels, not just work orders. Most CMMS histories cannot distinguish a planned PM from a breakdown repair from a warranty claim. Fixing the reason codes and the failure codes is usually the single highest value data project an agency can do here, and it pays off even if no model is ever built.
• Meter readings. Odometer, engine hours, pump runtime, and cycle counts are what convert calendar time into usage. Automated capture is far more reliable than a technician typing hours into a form.
• Telematics and fault codes. Fleet telematics and engine diagnostic codes are already streaming in many agencies. Confirm you can export the raw record, not just view it in the vendor portal.
• Sensor data where it exists. Vibration, temperature, pressure, amp draw, and run/fail signals from SCADA give the clearest degradation signal for rotating equipment. SCADA historian data is often the richest source an agency has and the least used outside operations.

Two cautions. Telematics and vehicle location data can raise employee monitoring questions that may be governed by labor agreements or state law, so involve human resources and counsel early. And treat operational technology connections carefully: adding sensors or data links to SCADA touches control networks, and should go through the same segmentation and review you would apply to any other OT change.

Which assets are actually good candidates

A model needs a population, a signal, and a usable lead time. Good candidates share a few traits: you own many similar units, failures happen often enough to learn from, degradation shows up in something you measure before the unit stops, and you have enough warning to act. Lift station pumps, HVAC compressors and air handlers, and medium and heavy fleet drivetrains tend to fit. A single one-of-a-kind pump station, a fifteen year old asset class you are about to replace, or equipment where failure is instantaneous and random tends not to.

Compare every prediction to a tightened PM schedule

Before accepting any model, require an evaluation against the simple alternative. Ask for a backtest on held-out history: train on earlier years, then check what the model would have predicted in later years, with no access to the outcome.

• Lead time. How many days or engine hours of warning does the prediction give? Less warning than your parts lead time is not actionable.
• How often the alarm is right. Of the units flagged, how many actually failed in the window? Of the units that failed, how many were flagged?
• The baseline. Run the same evaluation for a rule as simple as "service this class every X hours instead of every Y hours" or "inspect anything past Z hours." If the model does not clearly beat that, adopt the tightened schedule and save the money.
• The cost of being wrong in each direction. An unnecessary PM costs labor, a part, and downtime. A missed failure costs a tow, overtime, and sometimes a regulatory consequence. Those are not symmetric, and the threshold should reflect that rather than defaulting to a generic setting.

Route the prediction into the work order system

A prediction that lands on a dashboard produces a dashboard. A prediction that creates a record in the CMMS produces work.

1. Write the prediction into the maintenance system as a work request against the specific asset, with a recommended task, a due window, and the evidence behind it.
2. Name one person or role who triages new predictions on a defined cadence, with authority to schedule or dismiss.
3. Require a disposition code on close: fault confirmed, fault not found, deferred, other work performed. This is the feedback loop, and without it the model cannot improve.
4. Suppress duplicates. One open item per asset until it is closed, or technicians will stop reading them.
5. Check parts before alerting. A warning you cannot act on erodes trust faster than no warning.

What to ask vendors and what to put in the contract

Ask what minimum history the approach needs, how it behaves when meter readings are missing or late, how the model was validated and against what baseline, how often it is retrained and who does that, and whether each prediction comes with a human readable reason a supervisor can evaluate.

In the contract, require that the agency owns and can export its telematics, sensor, and work order data in a documented format, along with the model outputs and scores. Require integration into the existing CMMS or EAM rather than a separate portal. Set acceptance criteria in terms of work order outcomes over a defined pilot period, not in terms of software delivery. Confirm records retention treatment with your records officer and counsel, since maintenance records and the analytics derived from them may fall within public records obligations, and requirements vary by state.

A reasonable way to start

Pick one asset class. Clean its register, fix its failure coding, and automate meter capture. Establish the current baseline: failure rate, PM interval, and cost per failure. Then run any model in shadow mode alongside the existing schedule long enough to compare. Expand only to classes where the comparison came out in the model's favor.

HQT works with state and local agencies on data engineering, systems integration, and analytics for asset and maintenance programs.

Frequently asked questions