Predictive Maintenance: A Practical Guide for Plant Teams

Predictive Maintenance: A Practical Guide for Industrial Plants

Predictive maintenance (PdM) is a maintenance strategy that uses measured condition data — vibration, temperature, current, pressure, ultrasound — to decide when a machine actually needs intervention, rather than when a calendar says so. For electrical engineers and plant buyers, the practical question is not whether condition-based decisions are useful, but how to build a predictive maintenance program that survives contact with a real production schedule: which assets to instrument, which sensors match which failure modes, how alarms reach the right person, and how to avoid drowning in data nobody acts on. This guide covers those decisions in the order most plants actually face them.

Reactive, Preventive, Predictive: What Actually Changes

The three strategies differ in when work is triggered and what information drives the trigger. They are not mutually exclusive; most plants run all three at once, applied to different asset classes.

• Reactive (run-to-failure): Work starts after a functional failure. Acceptable only where failure has no safety, environmental, or significant production consequence, and where a spare or swap is cheap. It is a legitimate choice for some ancillary equipment — the mistake is applying it by default to critical assets.
• Preventive (time-based): Work is scheduled at fixed intervals derived from OEM recommendations or experience. It reduces some failures but introduces others: components are replaced before end of life, and failures between intervals are not caught. It also consumes labor regardless of need.
• Predictive (condition-based): Work is triggered by measured deviation from a baseline or by a trend crossing a defined threshold. The goal is to intervene inside the degradation window — early enough to plan parts and labor, late enough to use most of the component's life.

For an engineering audience the distinction matters because preventive and predictive maintenance require different infrastructure. Preventive work needs a scheduler. Predictive work needs sensors, data paths, thresholds, and a workflow that converts a reading into a work order. Budget and staffing follow the infrastructure, not the label.

The P-F Curve and Why Detection Lead Time Is the Design Target

The P-F curve describes the interval between the point a potential failure becomes detectable (P) and the point the asset reaches functional failure (F). The length of that interval — the detection lead time — is the resource a predictive maintenance program actually buys. Every decision about sensor type, sampling interval, and alarm threshold either preserves or consumes that lead time.

What the curve implies in practice

• Different failure modes become detectable at different points. Bearing spalling may show in high-frequency vibration or ultrasound well before it is audible; a loose electrical connection may show as a thermal rise before any vibration signature appears. Sensor choice must follow the failure mode, not the other way round.
• Sampling interval must be short relative to the P-F interval. A monthly manual route on an asset with a short degradation window can miss the window entirely; continuous or high-frequency collection is justified where the P-F interval is short or the consequence is severe.
• Thresholds set too tight consume lead time in false alarms; thresholds set too loose consume it in late detection. Both are calibration problems, and both are fixable with baseline data.

Choosing Which Assets to Instrument First

Instrumenting everything at once is rarely funded and rarely wise. Criticality ranking keeps the first phase small enough to complete and visible enough to defend.

Criticality inputs that matter

• Consequence of failure: safety, environmental release, product quality, and lost throughput. A small motor driving a main process pump can outrank a large motor driving a redundant utility.
• Failure history and mode: assets with repeat failures, or with failure modes that give useful warning (bearing degradation, misalignment, imbalance, lubrication breakdown), are strong early candidates. Assets that fail instantly and without warning are poor PdM candidates regardless of cost.
• Repair lead time: long-lead spares and specialized repair turnarounds raise the value of early warning, because planned intervention is the only way to avoid extended downtime.
• Redundancy: an asset with a warm standby has a different consequence profile than a single point of failure. Rank the single points first.

A workable first-phase screen

• List assets by consequence, then remove those whose failure modes give no detectable warning.
• Remove assets where reactive maintenance is genuinely cheaper than instrumentation and planned repair.
• From what remains, start with the top group where a sensor can be installed without process shutdown or major modification.
• Define, for each instrumented asset, what action a given alarm will trigger. If no action can be named, the asset is not ready for the program yet.

Matching Sensors to Failure Modes

Sensor selection is a mapping exercise: failure mode to physical quantity to transducer to data path. Getting this mapping right is more important than the number of sensors deployed.

Common mappings

• Rotating machinery — bearings, imbalance, misalignment, looseness: vibration measurement, typically accelerometers. Wired piezoelectric sensors suit permanent installations on critical machines with existing cable routes; wireless vibration sensors suit assets spread across a plant where cabling cost or access would otherwise delay deployment. High-frequency techniques (envelope/demodulation) are used for early bearing defect detection.
• Electrical distribution — loose connections, overload, insulation stress: temperature (thermal imaging or contact sensors) and current monitoring. Vibration is generally not the primary indicator here.
• Pumps and piping — cavitation, flow anomalies: pressure and vibration, sometimes ultrasound for leak detection.
• Compressed air and gas systems — leaks: ultrasound inspection, often as a periodic survey rather than permanent instrumentation.
• Lubrication condition: oil analysis remains a laboratory technique; vibration and temperature trends complement it but do not replace it.

Selection criteria beyond the sensing element

• Mounting: the mechanical path between the fault and the sensor limits what can be measured. A poorly mounted accelerometer produces data that looks valid but is not. Flat, machined surfaces and stud or adhesive mounting are preferred over magnetic mounts for permanent installations.
• Environment: temperature range, ingress protection, washdown, and EMI exposure at the mounting point determine the housing and cabling specification.
• Data path: wired sensors require cable routing, junction boxes, and a data acquisition point; wireless sensors require gateway placement, radio coverage checks, and battery or power planning. Both require a defined sampling and transmission interval.
• Bandwidth and sampling: bearing defect frequencies sit well above shaft rotation rates. A sensor or acquisition setting that samples too slowly will not resolve them, regardless of sensor quality.

Alarms, Thresholds, and Workflow Integration

A predictive maintenance program fails quietly when alarms are generated but not acted on. The workflow, not the sensor, is the part that determines whether the program produces planned repairs or just dashboards.

Setting thresholds defensibly

• Establish a baseline per asset under normal operating conditions, including load and speed variation. A threshold copied from a generic table without a baseline is a guess.
• Use two or three levels: an alert level for trending and inspection, an alarm level for planning, and where justified a trip or shutdown level tied to safety. Each level should have a named owner.
• Trend and rate-of-change often carry more information than absolute amplitude. A vibration level rising steadily over weeks is a planning signal even before it crosses an absolute limit.
• Review and recalibrate thresholds after the first months of data. Initial settings are provisional by nature.

Connecting readings to work orders

• Route alarms into the existing CMMS or work-order process rather than a separate system. An alarm that lives only in a vendor portal is an alarm that will be ignored at shift change.
• Define response times per alarm level and assign them to roles, not individuals.
• Record the outcome of every alarm — confirmed fault, false positive, or no-fault-found — so thresholds and sensor placement can be improved.
• Keep the condition monitoring data and the maintenance history in a form that allows the two to be compared. Without outcome data, the program cannot demonstrate its own value.

Common Program Mistakes

• Starting with technology instead of assets. Purchasing sensors before ranking criticality usually results in instrumentation that does not match the plant's actual risk profile.
• Instrumenting assets with no detectable warning. Some failure modes give no usable lead time; condition monitoring on these assets produces cost without benefit.
• No baseline. Thresholds set without normal-condition data generate false alarms, and false alarms erode trust faster than missed detections.
• Alarms without owners. If no role is accountable for responding to an alarm within a defined time, the data path ends at the dashboard.
• Treating PdM as a replacement for preventive maintenance. Lubrication, inspection, and calibration tasks remain necessary; predictive techniques prioritize and supplement them.
• Ignoring the electrical and instrumentation side. Vibration programs often overlook temperature and current monitoring on the distribution system that feeds the machines.
• No feedback loop. Programs that never review alarm outcomes cannot improve thresholds or justify expansion.

Checklist: Is the Plant Ready for a Predictive Maintenance Program?

• Asset criticality ranking completed and documented.
• Failure modes identified for each candidate asset, with the detectable quantity named.
• Sensor type, mounting location, and sampling interval specified per asset.
• Baseline data plan defined for each instrumented asset.
• Alarm levels, thresholds, and responsible roles agreed before installation.
• Alarm routing into the CMMS or work-order process tested end to end.
• Spare parts and repair resources identified for likely planned interventions.
• Review cadence set for alarm outcomes and threshold adjustment.

A 90-Day Starter Outline

Days 1–30: Define and rank

• Complete the criticality screen and select a first-phase asset group small enough to instrument completely.
• For each asset, document failure modes and the measurement that would reveal them.
• Survey mounting locations, cable or radio paths, and gateway positions.
• Agree alarm levels and the response workflow with maintenance and operations before any purchase.

Days 31–60: Install and baseline

• Install sensors and data acquisition on the first-phase group; verify mounting and data quality on each channel.
• Collect baseline data across the full operating envelope, including load and speed variation.
• Configure provisional thresholds from the baseline and route alarms into the work-order process.
• Train the responding roles on what each alarm means and what action it triggers.

Days 61–90: Operate, review, adjust

• Run the program through at least one full review cycle at each alarm level.
• Record outcomes for every alarm: confirmed, false positive, or no-fault-found.
• Adjust thresholds, sampling intervals, and mounting where the data shows they are wrong.
• Produce a short internal review of what the first phase detected and what it missed, and use it to select the second-phase assets.

Where Condition Monitoring Fits

Predictive maintenance depends on a condition monitoring layer that turns physical measurements into usable signals. LOGGON TECH supplies industrial metering and monitoring equipment for this layer, including wireless vibration sensors for rotating machinery where cabling would otherwise delay or prevent installation. Engineers specifying a first phase can review the condition monitoring range for the general architecture, and the wireless vibration sensors page for the sensor side of a rotating-asset deployment. Selection should follow the failure-mode mapping and baseline plan described above, not the other way round.

FAQ

How many sensors does a first predictive maintenance phase need?

There is no fixed number. A first phase should be sized by the number of assets that can be fully instrumented, baselined, and supported by a defined alarm workflow within the available staffing. A small, complete deployment with working workflows is more useful than a large deployment with no response process.

Can wireless vibration sensors replace manual routes?

Wireless sensors change the sampling model: data can be collected continuously or at short intervals without a technician visiting the machine. Manual routes still have value for assets below the criticality threshold, for verification, and for measurements wireless sensors do not cover. The choice depends on the P-F interval of the failure modes involved and on access constraints at the machine.

What is the difference between condition monitoring and predictive maintenance?

Condition monitoring is the measurement and data layer: sensors, acquisition, and the display of trends. Predictive maintenance is the decision and workflow layer that uses those measurements to schedule intervention. A plant can have condition monitoring without predictive maintenance — data collected and never acted on — which is why the workflow and alarm ownership sections above are treated as part of the program, not as an afterthought.