How to Use Predictive Maintenance for Plastic Equipment to Reduce Long-Term Operation Costs

Predictive maintenance is often described as a software project, but in a plastics plant it is first an operating cost decision. A line that stops without warning does not only lose output. It loses the material left in the barrel, the energy used to bring the machine back to temperature, the labor of the crew pulled off other work, the freight premium on a part that should have been on the shelf, and in many cases the goodwill of a customer whose delivery moved by a day. A condition based program exists to move that spending from emergency work into planned work, and to move it earlier in the failure curve where the repair is smaller. This guide explains how to use predictive maintenance for plastic equipment to reduce long term operation costs, with the same level of detail a plant engineer needs to build a budget, choose measurement points, write acceptance values, and defend the investment to a finance team.

The scope is deliberately practical. It covers the cost structure that makes the case, the asset ranking that decides where the first sensors go, the failure modes that can actually be detected on extruders and molding and recycling equipment, the measurement channels and data architecture, the logic that turns a signal into a work order, the phased rollout, the failure modes of the program itself, and the financial tracking that proves the savings were real. Every machine and project element carries an estimated price in US dollars so that the numbers can be transferred into a capital request. Wanplas supplies complete plastic production lines across compounding, recycling, pipe, sheet, film, board, profile, molding and filling applications, and the equipment guidance at the end of this guide reflects the way those machines are specified, instrumented and handed over in a real plant.

1. Why Long-Term Cost Is Decided by Maintenance Strategy

The true cost structure of a plastics plant

Before a maintenance strategy can be compared, the baseline has to be honest. Most plants track maintenance as a single line called repairs, which hides the largest part of the loss. The table below shows an illustrative annual picture for a mid sized plant with eight production lines, roughly 22,000 metric tons of annual output, and a mix of extrusion, molding and recycling equipment. The figures are planning values in US dollars, not a quotation, and they are included so that every later saving can be expressed as a share of a real denominator.

The important observation is the ratio. The maintenance parts line is 3.7 percent of total cost, while the two lines that react most strongly to machine condition, material and energy, are 83.9 percent. A program that only reduces the repair budget has a small ceiling. A program that prevents hot starts, keeps barrel temperature within a narrow band, catches a gearbox bearing before it sheds metal into the oil, and holds cooling water flow at design value attacks the larger numbers. That is why predictive maintenance for plastic equipment is framed as a long term operating cost tool rather than a cheaper way to buy bearings.

Three maintenance strategies and where the money goes

Reactive maintenance waits for a functional failure. Preventive maintenance intervenes on a fixed calendar or a fixed counter of operating hours. Predictive maintenance intervenes when a measured condition crosses a threshold or a trend projects a crossing before the next planned window. The three are not mutually exclusive; a healthy plant uses all three, but the mix determines the cost.

These three numbers describe the same plant. The move from reactive to preventive saves about 42,000 US dollars a year, mostly because the stop is scheduled. The further move to predictive saves another 58,000 US dollars a year, because the repair is smaller and the interval between interventions becomes longer without becoming risky. Neither figure includes the material and energy effects in the first table, which is where the larger gain usually appears once heater bands, cooling circuits and screw condition are brought under trend control.

The compounding effect of a single unplanned stop

A single event makes the mechanism concrete. Consider a pipe extrusion line rated at 800 kilograms per hour that stops for 6.5 hours because a haul-off gearbox bearing fails. The plant has no spare in stock, so the part is flown in. The barrel stays hot for two hours before the crew decides to purge, and the line needs 1.2 metric tons of material to return to dimensional tolerance.

The bearing itself was 420 US dollars. Everything else, 4,874 US dollars, was the price of finding out at the wrong time. A vibration channel on that gearbox, installed for roughly 380 to 620 USD per point including cabling, would have shown a rising bearing defect frequency four to ten weeks before the stop. That single avoided event pays for a substantial part of a pilot system, which is the arithmetic that makes a phased program much easier to approve than a plant wide rollout.

What predictive maintenance actually changes

Most degradation follows a curve with a detectable point and a functional failure point. Between them lies the warning window. A new bearing runs quietly. A spall forms on the race, producing a small repeating impact. The impact grows, the vibration signature becomes unmistakable, the housing warms, wear metal appears in the oil, and finally the bearing seizes or loses clearance. A calendar based plan has to guess the interval and therefore replaces parts early or misses one. A condition based plan measures the progression and schedules the repair inside the window, at a moment chosen by production rather than by the failure.

Three practical changes follow. First, the repair becomes small: a bearing and a seal instead of a bearing, a shaft, a housing bore, a full oil change and a damaged coupling. Second, the spare part is bought through normal purchasing at normal freight and can be verified before the stop, which removes the 960 USD air freight in the example above. Third, the line is stopped at a planned changeover, so the purging and restart losses are shared with a change that was going to happen anyway. The same logic applies to a heater band drifting toward open circuit, a cooling circuit losing flow to scale, a screw losing output to wear, or a hydraulic pump losing volumetric efficiency.

A cost equation the finance team can audit

The value of a condition based program in a given year is the sum of avoided emergency events, avoided premature replacements, avoided quality and energy drift, and avoided excess spare inventory, minus the annualized instrument, platform and engineering cost. Write it as four positive terms and one negative term, and insist that each positive term has a measurement behind it. Avoided events come from the maintenance log. Avoided premature replacements come from the parts usage history. Quality and energy drift come from scrap records and meter readings. Inventory release comes from the storeroom valuation. The annualized cost comes from the capital request. Programs that skip this step tend to be cancelled in the second year, because the savings were asserted rather than measured. Everything in the remaining chapters is designed to produce those measurements.

2. Equipment Criticality and Failure Mode Ranking

Build an asset register before buying any hardware

The most common failure of a predictive program is buying sensors for the machines that are easiest to instrument rather than the machines that matter. The corrective is an asset register, and it costs almost nothing but discipline. For each asset, record a unique tag, the line and process position, the manufacturer and the machine family, the rated capacity, the running hours per year, the consequence of a stop, the mean time to repair, the cost of the last three failures, the spare parts with their lead times, and every existing measurement point already available in the control system. A plant with eight lines typically ends up with 300 to 600 assets, of which 40 to 70 will ever justify continuous monitoring. The rest stay on route based checks, operator rounds or simple run to failure with a spare on the shelf.

Do not underestimate what is already in the control system. Modern extrusion and molding lines expose motor current, barrel zone temperatures, melt pressure, melt temperature, screw speed, hydraulic pressure, cycle time, vacuum level, cooling water temperature and drive torque through the programmable controller. Extracting these values costs far less than adding new hardware, and in many cases a torque trend or a heater duty cycle trend is a better early indicator than an additional sensor. A discovery audit of existing tags for a mid sized plant takes 3 to 6 engineering days, at 480 to 720 USD per day, and frequently removes a third of the proposed sensor bill.

A scoring method that ranks assets consistently

Score each asset from 1 to 5 on five dimensions, then apply weights that reflect the plant’s own economics. The weighted total decides the monitoring tier. This is deliberately simple; a complicated scoring model that nobody can explain will be ignored by the people who have to act on the alarms.

Assets scoring above 4.0 go to continuous monitoring. Assets between 2.8 and 4.0 go to weekly or monthly route based measurement with a handheld instrument. Assets below 2.8 stay on preventive rounds or run to failure. The last row matters as much as the first four: a component with a huge consequence but no measurable warning belongs in redundancy, guarding, or a spare parts strategy, not in a sensor budget. A heater band that fails open in a single instant is a good example. You can detect the change in duty cycle before it fails, but once it is open, the useful response is a spare band and a fast changeover, not an alarm system.

Warning windows set the measurement method

The length of the warning window determines sampling frequency, and sampling frequency determines cost. Measuring once per week is useless for a defect that develops in three days and wasteful for one that develops over nine months. The table below gives realistic planning windows for common plastic equipment degradation. Use them to size the system, and adjust them after the first year with the plant’s own records.

Which assets earn the first measurement points

Applying the scoring above to a typical Wanplas style plant produces a short list. Continuous monitoring usually starts on the main extruder drive and gearbox, the haul-off or roll stack gearboxes, the pelletizer or granulator main bearings, the injection molding clamp and hydraulic power unit, the PET blow molding high pressure compressor, the BFC filler main drive, the chiller compressor, and the largest cooling pumps. These assets combine high hourly loss, long repair time, and a measurable precursor. The second wave adds vacuum pumps, screw feeders, film winding drives, printing and treating units, and downstream conveyors.

What to leave on route based checks

Not every asset deserves a permanent channel. Small gear motors, cabinet cooling fans, low pressure water pumps, conveyor drives on noncritical transfers, and individual heater bands are cheaper to inspect on a weekly walk with a handheld vibration pen, an infrared camera and a clamp meter. A capable handheld set costs 1,800 to 4,600 USD in total and covers 200 to 400 measurement points across a plant. A route of 120 points takes one technician about 3 hours, at roughly 30 to 45 USD per hour, so a weekly route costs 5,000 to 7,000 USD a year in labor and still gives early warning for the majority of slow developing faults. Continuous monitoring is then reserved for the assets where a weekly sample would miss the window or where the hourly loss is high enough to justify the hardware.

There is also a class of assets where the right investment is not measurement at all. Safety interlocks, guards, emergency stops, light curtains and pressure relief devices must be proof tested on a fixed schedule regardless of any trend, because a silent dangerous failure cannot be inferred from a gradually changing signal. Predictive techniques can support them, for example by tracking door cycle counts or stop time drift, but they never replace the functional test. Keeping this boundary clear is one of the markers of a mature program, and it prevents the uncomfortable situation where a sensor is installed and a mandated inspection is quietly dropped.

3. Detectable Failure Modes Across Plastic Machinery

Start from the failure, not from the sensor

A predictive program succeeds when every measurement channel is tied to a named failure mode that somebody in the plant has actually seen. If the team cannot state what a rising value predicts, the channel will eventually be muted. The discipline is to write one line per failure mode: what degrades, what signal changes first, how much warning is realistic, what the repair costs if caught early, and what it costs if the machine stops. The sections below work through the main equipment families in a plastics plant in that order.

Rotating drives: bearings, gears, couplings and alignment

Rolling bearings on extruder main drives, granulator rotors, haul-off gearboxes, roll stacks, cooling fans and pumps produce the clearest early signals. A spall on an outer race generates a repeating impact at a frequency set by the bearing geometry and shaft speed. Long before overall vibration rises, that impact shows up in an envelope or high frequency band. Gear tooth pitting shows up as a growing amplitude at the mesh frequency with sidebands spaced at shaft speed; a single cracked tooth produces a once per revolution impulse. Coupling wear and shaft misalignment appear as elevated vibration at one and two times running speed, often with a phase difference across the coupling that a two channel measurement can confirm. Mechanical looseness produces a distinctive series of harmonics and a directional reading that changes between the horizontal and vertical axes.

The financial case is strong because the repair escalates sharply. Replacing a bearing during a planned window costs 300 to 900 USD in parts plus 2 to 5 hours of labor. Letting it fail can score the shaft, damage the housing bore, contaminate the oil, and take a gearbox out for 18 to 36 hours, with a bill of 4,000 to 14,000 USD before any production loss is counted. On a high torque compounding gearbox the gear set alone can be 18,000 to 45,000 USD with a lead time of 10 to 20 weeks, which turns a 400 USD bearing into a business continuity problem.

Screw and barrel wear

Wear in an extruder is slow, which is exactly why a trend is more valuable than an absolute value. As the clearance between screw flight and barrel grows, backflow increases. For a given screw speed the output falls, melt temperature rises because the material spends longer under shear, and the drive torque needed for a given output increases. Track three ratios together: kilograms per hour per screw revolution, melt temperature at a fixed set point, and specific energy in kilowatt hours per kilogram. A 4 to 8 percent drift in output per revolution over several months, with stable formulation and screen pack condition, is a reliable indicator of clearance growth. A step change usually means a screen pack blockage, a feed problem or a heater fault, not wear.

A screw and barrel set for a mid sized compounding line costs 12,000 to 38,000 USD depending on diameter, alloy and nitriding or bimetallic treatment, with a lead time of 8 to 16 weeks. Planning the replacement during a scheduled shutdown avoids a rushed purchase and lets the plant verify the new parts before they are needed. Trending also answers the harder question of whether a rebuild, at 3,500 to 9,000 USD, is still possible, or whether the clearance has passed the point where hard facing makes economic sense.

Heating, temperature control and melt stability

Heater bands degrade in a predictable way. Resistance rises as the element oxidizes, the controller duty cycle increases to hold the set point, and the zone eventually fails open. Tracking duty cycle per zone alongside zone current is often cheaper than adding hardware, because both values exist in the controller. A healthy zone that used to run at 32 percent duty and now runs at 58 percent under the same recipe, ambient conditions and throughput is telling you something: the band is weakening, the insulation is damaged, a thermocouple has drifted, or the cooling water valve is leaking into the zone. A failing solid state relay produces a different pattern, with the zone toggling between full on and full off and a growing temperature oscillation.

Thermocouple drift is especially costly because it is invisible in the temperature reading itself. The controller believes the zone is at set point while the melt is actually 6 to 12 degrees C off, producing dimensional variation, gels, discoloration or poor weld line strength. Compare the zone reading against a calibrated handheld probe during a monthly round and track the deviation. A replacement thermocouple costs 25 to 90 USD; a single day of off-spec sheet at 2,400 kilograms and 1.4 USD per kilogram is 3,360 USD of material alone. A full set of calibrated probes and a small dry block calibrator costs 900 to 2,200 USD and pays for itself quickly on a multi zone line.

Hydraulic and pneumatic power

Injection molding clamps, blow molding mold carriages and many auxiliary actuators rely on hydraulic power. Pump wear reduces volumetric efficiency, which shows up as a longer pressure rise time at the start of each cycle, a higher case drain flow, and a gradual increase in oil temperature. Track pressure rise time from a fixed starting point rather than absolute pressure, because oil viscosity and temperature move with the season. A rise from 0.8 seconds to 1.25 seconds at the same oil temperature is a meaningful signal. Cylinder seal bypass appears as an inability to hold clamp pressure with the pump unloaded, and valve spool wear appears as a growing lag between the command and the measured movement.

Hydraulic contamination is the root cause behind many of these failures. An inline particle counter costs 2,800 to 6,500 USD installed, while a quarterly laboratory oil analysis costs 60 to 140 USD per sample. A plant with six hydraulic power units spending 2,400 USD a year on oil analysis can avoid a pump replacement at 3,200 to 8,500 USD and a valve rebuild at 1,200 to 3,000 USD. Compressed air leaks are the pneumatic equivalent: an ultrasonic leak survey with a 900 to 2,400 USD handheld instrument typically finds 15 to 40 leaks on a medium plant, each wasting 150 to 900 USD a year in electricity at industrial power rates.

Electric motors, drives and power quality

Motor current signature analysis can detect broken rotor bars, winding insulation decline, air gap eccentricity and load problems without adding a single sensor if the drive already reports current. Current imbalance between phases above 3 to 4 percent under balanced supply deserves investigation. Winding temperature trending catches cooling path blockage, which is common on extrusion lines where polymer dust and granules collect on fan covers and fins. A motor that runs 15 degrees C hotter than its own baseline at the same load has roughly half the insulation life it should have.

Rewinding a 55 kilowatt motor costs 1,800 to 3,600 USD with a lead time of 1 to 3 weeks; a new high efficiency replacement is 3,400 to 6,200 USD. An emergency bearing failure that damages the winding can double that. Power quality logging at the main distribution board, at 1,200 to 3,800 USD for a fixed analyzer, also protects the variable frequency drives that feed extruders and haul-offs, where a single drive module replacement runs 2,200 to 9,000 USD depending on rating.

Cooling circuits, molds and heat exchange

Scale, corrosion products and biological growth narrow cooling channels and reduce heat transfer. The measurable precursors are a falling flow at a fixed pump speed, a growing difference between supply and return temperature at constant load, and an increasing approach temperature at the heat exchanger. On a mold, the cycle time needed to reach ejection temperature creeps upward. On a sheet line, roll surface temperature becomes uneven across the width. A 10 percent flow loss on a cooling circuit can add 2 to 5 percent to cycle time on an injection machine, which on a 22 second cycle running 6,000 hours a year is 200 to 490 hours of lost capacity.

Chemical cleaning of a circuit costs 400 to 1,600 USD in chemicals, pumps and labor. Replacing a badly fouled roll with an internal spiral channel can cost 9,000 to 26,000 USD with a long lead time. Tracking flow and approach temperature costs very little when the plant already has temperature instrumentation, and it converts a cleaning decision from a calendar guess into a measured need.

Downstream and process specific equipment

On recycling lines, granulator knife edge rounding raises motor current, increases fines, and changes flake size distribution; a rising current at constant throughput plus a growing fraction of undersize flakes is the signal to rotate or sharpen knives. Knife sets cost 350 to 1,400 USD and a sharpening service 120 to 300 USD, while a damaged rotor from a hard foreign object can exceed 7,000 USD. On filling lines, fill volume deviation, valve cycle count and vacuum level decay predict valve seat wear and seal failure; a filling valve rebuild kit costs 180 to 520 USD, while an unplanned stop on a high speed line can exceed 1,800 USD per hour. On film and sheet lines, winding tension deviation and roll diameter growth predict dancer and brake wear, and thickness profile drift points to die bolt or lip damage. In every case the pattern is the same: a slowly changing ratio, measured under comparable operating conditions, beats a single absolute alarm.

A consolidated failure mode map

The table below is the working summary used to assign measurement channels. Costs are planning estimates in US dollars for a mid sized machine and include parts and internal labor but not lost production, which is deliberately kept separate so the two can be added honestly.

4. Sensor Selection and Measurement Channels

Vibration: the workhorse channel

For rotating assets, vibration is the most informative single measurement. A general purpose piezoelectric accelerometer with a sensitivity of 100 millivolts per g and a usable range up to 10 kilohertz costs 90 to 260 USD. A permanently mounted industrial unit with a protective housing, integral cable and a rated ingress protection class costs 220 to 520 USD. Wireless vibration nodes with an internal battery and a radio link cost 380 to 850 USD per point and remove most of the cabling cost, at the price of a lower sampling rate and a battery replacement every 2 to 4 years. For slow speed equipment such as a large cooling fan or a drum mixer, choose a low frequency sensor with a 0.5 hertz lower limit, which typically adds 120 to 300 USD per point.

Mounting quality matters more than sensor price. A stud mounted sensor on a clean flat machined pad gives repeatable data; a magnetic base on a painted curved housing can attenuate high frequency content by more than half and will drift as the mount loosens. Budget 60 to 180 USD per point for a proper mounting pad, adhesive or stud, cable route and junction box. On a gearbox, place sensors on the bearing housings in the radial direction, with one axial point on the output shaft, and record the exact position with a photo in the asset register so that a replacement sensor goes back in the same place.

Temperature: contact, infrared and imaging

Platinum resistance temperature devices cost 35 to 110 USD each and are the right choice for bearing housings, winding pockets, oil sumps and cooling water lines where a stable absolute value matters. Thermocouples cost 18 to 60 USD and suit higher temperature zones. Noncontact infrared point sensors, at 130 to 420 USD, work on moving surfaces and on parts that cannot be reached with a cable. A handheld thermal camera, at 1,600 to 5,800 USD for a model with sufficient resolution for electrical and mechanical surveys, is shared across the whole plant rather than installed per asset, and it is one of the best returns in a route based program.

Temperature alone is a weak early indicator for most mechanical faults; by the time a bearing housing is 10 degrees C hotter, the damage is often advanced. Its strength is as a confirming channel and as a process measurement: heating zone duty, cooling approach, oil sump temperature and mold surface balance. Treat it as part of a group, and require a second independent signal before scheduling a stop.

Current, power and torque

Current transformers with a signal conditioner cost 90 to 280 USD per three phase circuit, but on a modern line the drive already reports current, torque, power and speed over the control network, so the extraction cost is engineering time rather than hardware. This is the cheapest high value channel in most plastics plants. Motor current tracks load, knife condition, screw wear, conveyor drag, pump efficiency and feed consistency. Torque per kilogram of output is a sensitive measure of process health. Power factor and harmonic distortion protect the drive and the supply.

The caution is normalization. Current rises with throughput, with melt viscosity, with ambient temperature and with a formulation change. Store the recipe and the throughput alongside the current, and compare like with like. A channel that ignores operating context will generate alarms at every product changeover and will be ignored within a month.

Pressure, flow and level

Melt pressure transducers rated for high temperature cost 480 to 1,300 USD and are often already fitted. Hydraulic pressure transducers cost 120 to 420 USD. Cooling water flow meters cost 260 to 900 USD depending on pipe size and technology, with clamp on ultrasonic units at the upper end avoiding any pipe work. Differential pressure across a filter or a screen pack, at 180 to 520 USD for a transmitter, is a direct measure of blockage and one of the cleanest predictive signals on an extrusion line. Oil level, water level and tank temperature sensors cost 60 to 220 USD each and prevent a whole family of avoidable pump failures.

For a cooling circuit, the useful derived value is not flow but the heat balance: flow multiplied by the temperature rise across the circuit, compared with the electrical power going into the process. A falling ratio points to scaling or air binding; a rising ratio at constant production points to a control or valve problem. Both can be computed in software from three inexpensive measurements.

Oil condition and wear debris

Oil analysis is a scheduled laboratory measurement rather than a continuous channel, but it belongs in the plan. A standard package including viscosity, water content, acid number, particle count and elemental wear metals costs 60 to 140 USD per sample. Quarterly sampling on six gearboxes and four hydraulic units costs about 2,400 to 5,600 USD a year. An inline particle counter at 2,800 to 6,500 USD installed is justified on the highest consequence gearbox, where a single avoided gear set replacement is worth 18,000 to 45,000 USD. Inline water in oil sensors at 600 to 1,400 USD are worth fitting on outdoor or washdown equipment where condensation and hose contamination are common.

Interpret trends rather than single results. A jump in iron and chromium together points to wear of hardened surfaces; a jump in silicon and aluminum points to dirt ingress through a breather; a rising water content with stable particle count points to a cooler leak or condensation. Keep the same sampling point, the same bottle, the same running condition and the same laboratory, or the trend becomes noise.

Acoustic, ultrasonic and visual channels

High frequency acoustic emission sensors, at 700 to 1,900 USD per channel, detect early bearing defects and slow speed rubbing before conventional vibration does. Ultrasonic instruments are more commonly used as handheld survey tools for compressed air leaks, steam traps, electrical corona and vacuum leaks, at 900 to 2,400 USD for a capable unit. A fixed acoustic monitor on a critical slow speed bearing can be worth the cost where the warning window is short and the repair time is long.

Machine vision has a narrower but growing role: checking pellet shape at a die face, detecting film web edge position, reading a roll diameter, or watching for granule buildup at a feed throat. A basic industrial camera with lighting and a small edge processor costs 900 to 2,800 USD installed. Laser shaft alignment tools, at 2,400 to 6,500 USD, are not continuous channels but should be in the plant toolkit, since misalignment is a root cause behind a large share of repeat bearing and coupling failures.

Installed cost per measurement point

The table below gives realistic installed prices in US dollars, including the device, mounting hardware, cable and tray, termination, a share of the input module or gateway, configuration and documentation. These are the numbers to use when building the pilot budget in Chapter 8.

Specification rules that prevent wasted spending

Five rules keep the hardware budget honest. First, extract existing control system tags before ordering anything. Second, specify the failure mode and the decision that each channel will trigger, in writing, before purchase; a channel with no decision attached is a research project. Third, standardize on a small number of device types and mounting methods so that spares, training and calibration stay simple. Fourth, record installation geometry and firmware version in the asset register, because a sensor replaced in a slightly different position invalidates the baseline. Fifth, buy the calibration and the spare at the same time as the sensor. A 220 USD accelerometer without a 180 USD calibration check and a spare cable will eventually produce data that nobody trusts, and untrusted data is the most expensive kind.

5. Data Acquisition, Edge Processing and History Storage

A four layer architecture that stays maintainable

A durable predictive setup has four layers, and mixing them is the usual cause of a system that works at handover and fails a year later. Layer one is the measurement layer: sensors, transmitters, drive data and controller tags. Layer two is the edge layer: input modules, a gateway or industrial computer near the line, and the logic that time stamps, filters, buffers and compresses values. Layer three is the history layer: a time series store that keeps raw and derived values with their context and serves them to analysis. Layer four is the decision layer: rules, trend models, dashboards, alarms and the interface to the maintenance system.

The most important design choice is that the edge layer must keep working when the network to the history layer is down. A plastics plant has electrical noise, planned network maintenance, and occasional switch failures. An edge gateway with local buffering for 7 to 30 days of data, at 1,800 to 4,600 USD per line, prevents gaps that would otherwise destroy a trend at exactly the moment an incident is being investigated. Buffering also means a gateway reboot does not lose the vibration spectra captured during an overnight run.

Sampling rates and the data volume they create

Sampling everything at high speed is the fastest way to make a system unaffordable. Match the rate to the physics. Slow process values such as zone temperature, cooling flow, oil temperature and tank level change over minutes and need one sample every 1 to 30 seconds. Motor current, torque, melt pressure and speed change over fractions of a second and need 10 to 100 samples per second if a per cycle signature is wanted, or a one second aggregate if only a trend is needed. Vibration waveforms for bearing and gear analysis need 5,000 to 25,000 samples per second per channel, but only in short bursts; a 2 second waveform captured once per hour and a set of scalar features captured once per minute is enough for most programs.

A pilot with 42 measurement channels on two lines typically produces 8 to 22 gigabytes a year with this plan. That fits comfortably on a 2 to 4 terabyte redundant server costing 3,400 to 7,800 USD, with room for a plant wide rollout. A design that stores full waveforms continuously on 200 channels can produce 30 terabytes a year and forces a storage discussion into every budget meeting, with no improvement in fault detection.

Protocols, naming and context

Use open industrial interfaces rather than custom point to point wiring. Controller data commonly arrives over industrial Ethernet using OPC UA, Modbus TCP or a vendor neutral gateway, while sensor and message traffic often uses MQTT with a lightweight payload. The specific choice matters less than three rules: every value carries a UTC time stamp with millisecond resolution, every value carries a quality flag, and every value is attached to a stable tag name that follows a documented convention. A tag such as LINE02_EXTR_GBX_BRGVIB_01 tells an engineer the line, the asset, the component, the measurement and the position without opening a spreadsheet.

Context is what makes a signal interpretable. Alongside each measurement, store the recipe or product code, the throughput set point, the running state, the ambient temperature where relevant, and the shift identifier. Without these, a rising motor current is ambiguous: it could be a worn knife, a harder formulation, a higher throughput or a blocked screen. With them, the analysis can compare the current only against other periods running the same recipe at the same rate. Context tags usually come from the controller or the manufacturing execution system and cost almost nothing once the interface exists. This is the single highest value part of the data model, and it is the part most often left until last.

Edge processing versus central analysis

Push simple, time critical logic to the edge and keep heavy analysis central. The edge gateway can compute one minute aggregates, extract band energies from a vibration signal, detect a rate of change, apply a dead band, and raise a local alarm in under a second. Central analysis is where cross asset comparison, long term trend fitting, model training and reporting belong. Splitting the work this way keeps the network load low, keeps the alarm functioning during a server outage, and makes it possible to add a new line without rewriting the analysis layer.

A practical edge device for one production line costs 1,800 to 4,600 USD including enclosure, power supply, isolated inputs and configuration. A plant wide edge standardization program with eight lines, spare units and engineering documentation typically lands at 22,000 to 48,000 USD. Resist the temptation to run analysis directly on the controller; the controller’s job is deterministic machine control, and adding diagnostic code there creates a change management burden every time a rule is tuned.

History storage, retention and access

Keep three retention tiers. Raw high resolution data stays for 60 to 180 days, which covers the investigation window for any failure. Aggregated features and alarms stay for 5 to 10 years, which is what a long term degradation trend needs. Event records, work orders and the decisions taken stay for the life of the asset, because they are what makes the financial case auditable. A time series database with compression handles the first two tiers; the third usually belongs in the maintenance system rather than in the historian.

Access control deserves attention from the start. Operators need a simple view of current condition and the next planned action. Maintenance engineers need waveforms, spectra and history. Production planners need the projected intervention date. Finance needs a monthly savings summary. Four views on one data set, with role based permissions, avoids the common problem of a single dashboard that is too complex for the floor and too shallow for the engineer.

Architecture cost for a two line pilot

The table below is a realistic bill for a pilot covering one compounding line and one sheet or pipe line with about 42 measurement channels in total. Prices are in US dollars and include hardware, software licenses and the integration labor normally quoted separately by a system integrator.

6. Baselines, Thresholds and Trend Logic

Build the baseline from healthy running, not from a handbook

An absolute vibration limit taken from a general standard is a poor starting point for a specific gearbox mounted on a specific frame. The same machine running a filled compound at high torque produces a different signature than the same machine running an unfilled resin at low torque. The reliable approach is to build a plant specific baseline: collect 30 to 90 days of data after the asset is known to be in good condition, cover every recipe and throughput band that the asset normally runs, and store the distribution of each feature within each operating band rather than a single number.

Record the moment the baseline was created and what physical condition it represents. If the gearbox had 8,000 hours on its oil and a slightly worn coupling when the baseline was captured, that is the starting point, and it should be written down. A baseline with no condition note is impossible to interpret two years later, when somebody asks whether a 12 percent rise is degradation or simply a different product mix. Baseline documentation costs a few hours of engineering time and prevents months of argument.

Segment by operating condition before comparing anything

Create operating bands using the context tags from Chapter 5. Typical bands for an extruder are recipe family, throughput range, screw speed range and whether the line is in steady production, start-up, purge or idle. Compute a separate baseline for each combination that occurs often enough to matter, and discard transient states from the trend entirely. Start-up data is particularly destructive: a cold machine with a high torque spike will dominate a statistical limit if it is not excluded, and every subsequent alarm will be too insensitive to catch a real fault.

A simple implementation uses a state tag derived from the controller: the line is in steady state only when it has been running for more than 20 minutes, throughput is within 5 percent of set point, and no recipe change has occurred in the last 15 minutes. Values outside steady state are stored but excluded from baselines and from alarm evaluation. This one rule removes a large share of false alarms in extrusion and molding applications.

Four kinds of threshold and when each one works

Absolute limits are appropriate for safety and for physical constraints, such as a maximum bearing housing temperature or a maximum melt pressure. They are poor for early fault detection. Relative limits, expressed as a percentage change from the asset’s own baseline in the same operating band, are much more sensitive; a 30 percent rise in a bearing envelope band is meaningful even if the absolute value is still low. Statistical limits, such as three standard deviations from the band mean or a percentile of a rolling window, adapt to normal variation and are useful when the baseline distribution is stable. Rate of change limits catch acceleration: a value that has been creeping up 1 percent a week and suddenly rises 2 percent a day deserves attention even before it crosses any level threshold.

Alarm levels with an owner and a deadline

An alarm without a named owner and a response deadline is a log entry. Define four levels. An advisory is a trend note visible to the maintenance engineer, reviewed at the weekly meeting, with no immediate action. A warning requires a confirmation check within 5 working days and a note in the asset record. An action level requires a work order, a spare part check and a planned intervention date, normally within 2 to 4 weeks. A critical level requires production to be notified immediately and the machine stopped at the next safe point. Each level should have a distinct color and notification policy on the dashboard, and only critical levels should reach a phone after hours.

Set the escalation rule explicitly: a warning that is confirmed and then rises again within 30 days moves up one level automatically. This prevents the common pattern where a slow developing fault generates a warning every month and is acknowledged every month until it fails. Conversely, require that any alarm can be suppressed only with a reason code and an expiry date, so that a suppression never becomes permanent by accident.

Confirm with a second independent signal

Single channel alarms are the main source of wasted engineering time. A confirmation matrix pairs a primary indicator with an independent one that has a different failure mechanism, so a sensor fault cannot easily imitate a machine fault.

Rules that keep false alarms under control

Four inexpensive rules do most of the work. Require persistence: a threshold must be exceeded for a defined number of consecutive samples or a defined number of days, not once. Apply a dead band around the limit so a value hovering at the boundary does not toggle the alarm on and off. Suppress evaluation during start-up, purge, recipe change, cleaning and any period when the machine state tag is not steady. Require a quality flag of good on the measurement, so a disconnected sensor reports a data fault rather than a machine fault.

Track the false alarm rate as a program metric from the first month. A well tuned pilot should stay below 10 to 15 percent of alarms leading to no finding. If the rate is higher, resist adding more sensitivity; first check operating band segmentation, then sensor mounting, then context tags. Adding a second threshold to a badly normalized signal only doubles the noise.

Worked example: turning a trend into a date

Suppose the envelope band on a haul-off gearbox output bearing sits at 0.42 millimeters per second for eight weeks in the main operating band, then rises to 0.51, 0.58 and 0.67 over three consecutive weeks. The action level is 0.85, and the plant’s own history says this bearing family reaches functional failure between 1.1 and 1.4. A linear fit on the last four points gives roughly 0.08 per week, so the action level is about 2.3 weeks away and the failure region about 5 to 9 weeks away. That is enough to order a 420 USD bearing, book a 2 hour window at the next scheduled changeover, and confirm the seal and shaft dimensions while the machine is still running. Note what made this possible: a stable baseline in one operating band, a rate of change rule, and an independent confirmation from an oil check. The absolute value of 0.67 was never close to any generic alarm limit.

Change management after a repair

Every repair changes the machine. A new bearing, a rebuilt gearbox, a replaced sensor or a realigned coupling all create a step change in the signal. If the old baseline is kept, the program will either generate a false alarm or, worse, hide the next real deviation inside an artificially wide band. Make baseline reset part of the work order close out: capture the physical condition, allow a short run in period of 24 to 72 hours, then collect a fresh 14 to 30 day baseline and archive the old one with a link to the work order. The archive matters, because comparing the before and after distributions is the simplest proof that the repair actually worked, and it becomes the evidence base for the next threshold revision.

7. From Rules to Remaining Useful Life Models

Five levels of analytical maturity

Plants often assume that predictive maintenance requires machine learning. In practice most of the value arrives at the second and third levels, and a plant that starts at the fifth level without clean data usually ends up with a demonstration rather than a working program. Level zero is simple visibility: a dashboard showing current values and history. Level one is fixed rules and thresholds of the kind described in Chapter 6. Level two is feature extraction, where a waveform becomes a set of meaningful numbers such as band energies, kurtosis, harmonic ratios and cycle signatures. Level three is statistical or learned anomaly detection, which compares the current feature vector with the healthy distribution for that operating band. Level four is a remaining useful life estimate, which projects the degradation trajectory to a future crossing and gives a date range rather than an alarm.

The cost and the prerequisite data rise sharply with each level. Level one needs a few weeks of data and a controls engineer. Level three needs months of labeled or at least well segmented data across several operating conditions. Level four needs several complete degradation histories, which a single plant may not have; this is where pooling data across identical machines becomes valuable. A sensible program climbs the levels deliberately, with a written decision at each step about whether the next level would change an actual maintenance decision.

Frequency analysis and the features that actually matter

A vibration waveform in the time domain hides early bearing damage. Converting it to a spectrum with a fast Fourier transform separates the energy by frequency, so a small repeating impact can be seen at its own characteristic frequency even while the overall level is unchanged. Envelope demodulation goes further: it extracts the amplitude modulation caused by a defect striking once per revolution and reveals the defect frequency against a much quieter background. For a gearbox, track the amplitude at the gear mesh frequency and the sidebands spaced at shaft speed; for a bearing, track the ball pass frequencies computed from the bearing geometry; for a fan or rotor, track one and two times running speed for imbalance and misalignment.

Order tracking matters on variable speed lines. A pipe haul-off or a film winder changes speed constantly, so a fixed frequency bin smears the same fault across the spectrum. Synchronizing the sampling to a shaft encoder and expressing the spectrum in orders of running speed keeps the fault at a fixed position. A tachometer input costs 120 to 380 USD per asset and often improves detection more than a more expensive accelerometer.

The useful scalar features are few: root mean square velocity for general energy, peak value and crest factor for impacts, kurtosis for the impulsiveness that precedes a visible vibration rise, energy in three to six defined frequency bands, and the ratio of the defect frequency amplitude to the broadband floor. Store these once per minute per channel. They occupy little space, they are easy to explain to a maintenance technician, and they are sufficient for most alarm logic.

Statistical trending before machine learning

Exponentially weighted moving averages and cumulative sum charts are underused in this field. Both are designed to detect a small persistent shift in a noisy signal, which is exactly the shape of early degradation. An exponentially weighted moving average with a smoothing constant around 0.2 will reveal a 5 percent drift in motor current or specific energy within days, long before a simple limit is crossed. Cumulative sum charts are better when the shift is gradual and the noise is high. Both can be implemented in a few lines of code at the edge layer, both produce a number that an engineer can inspect, and neither needs labeled failure data.

Pair these with a multivariate view. A single signal drifting by 5 percent may be nothing; the same drift in output per revolution, specific energy and melt temperature together is strong evidence of screw wear. A simple health index formed from three to six normalized features, weighted by how strongly each one moved during known past events, is often all a plant needs. This is not a black box, which matters when a production manager asks why a line should stop two weeks early.

Where learned models help and where they hurt

Unsupervised anomaly detection is genuinely useful on complex assets with many interacting signals, such as a high speed filling line or a multi zone sheet line. It learns the shape of normal operation across all operating bands and flags combinations that have not been seen before. It needs no failure labels, only clean normal data, which makes it a realistic third level step after 3 to 9 months of collection. Supervised fault classification needs examples of each failure, and a single plant may see only two bearing failures in three years. Where a machine family is repeated across several lines, those histories can be pooled and the model becomes viable.

Avoid three common mistakes. Do not train on data that includes start-up, purge and changeover unless those states are labeled and modeled separately; the model will learn that a start-up is normal and become blind to a fault that develops during steady running. Do not validate on a random split of consecutive samples, because adjacent samples are nearly identical and the accuracy will be inflated; split by time period or by production campaign. Do not deploy a model whose output cannot be traced back to the features that drove it. If an engineer cannot explain an alarm to the person who has to stop the line, the alarm will be overridden.

Remaining useful life without false precision

A remaining useful life estimate should be a range with a confidence statement, not a date. Three approaches are practical in a plastics plant. The first is trajectory projection: fit a curve to a degradation feature, extend it to the action threshold, and report the crossing window using the spread of recent residuals. It needs only one degrading asset and works well for the linear or gently accelerating phase of bearing and filter degradation. The second is similarity based: compare the current trajectory with archived histories from the same asset family that reached a known outcome, and report the distribution of their remaining times. The third is a survival style model that estimates the probability of failure in the next 30, 60 or 90 days from condition features and operating hours, which is more useful for planning than a point estimate.

Report it in a form that production can act on: the earliest sensible intervention date, the latest safe date, and the confidence band. A result such as “plan this bearing between 12 and 26 days from now, with moderate confidence” is far more useful than “failure in 19.4 days”. Always attach the underlying trend plot and the confirmation measurement so the maintenance planner can judge the call. Where the band is wider than the lead time of the spare part, the correct action is to order the part, not to stop the machine.

Deciding how far to climb

Use a simple test: would a more sophisticated output change a decision? For a heater band with a 1 to 10 day warning window, a graded duty cycle rule plus a spare band on the shelf is the whole answer; a remaining useful life model adds cost and no value. For a compounding gearbox with a 20 week lead time on a gear set and an hourly loss above 900 USD, a trajectory projection with a confidence band is worth the engineering effort because it changes purchasing, scheduling and inventory decisions. Apply the expensive methods only where the decision is expensive too. Most plants find that 70 to 80 percent of their monitored assets are well served by levels one and two, and that the top 10 to 15 assets justify level three or four.

8. A Phased Implementation Roadmap With a Project Budget

Phase zero: preparation, weeks 1 to 6

The objective of phase zero is to know what you have before spending on hardware. Deliverables are a complete asset register for the two pilot lines, a criticality score for every asset, a written failure mode table with the signal and decision for each proposed channel, a discovery audit of tags already available in the control system, a review of the last 24 months of maintenance and downtime records, and an agreed savings measurement method signed off by finance. The last item is easy to skip and expensive to skip: without a frozen baseline of downtime hours, emergency freight, overtime and scrap, nobody can prove the program worked in twelve months.

Phase zero costs 9,000 to 20,000 USD, almost entirely engineering time at 480 to 720 USD per day, plus a handheld survey set at 1,800 to 4,600 USD if the plant does not already own one. A useful by product is that the failure mode table often reveals two or three chronic problems that can be fixed immediately with no new technology at all. Plants frequently recover the whole cost of phase zero from those fixes before a sensor is ordered.

Phase one: the pilot, weeks 7 to 22

Choose two lines with different failure profiles and a supportive production team. One compounding or extrusion line gives rotating machinery and thermal systems; one molding or filling line gives cyclic hydraulic and drive signatures. Install 35 to 50 measurement channels, favoring extracted controller tags over new hardware. Run the first 30 to 60 days in observation mode with no alarms sent to operations, so baselines can be built and segmentation errors corrected without training the team to ignore notifications. Then switch on graded alarms, with the maintenance engineer reviewing every alarm and recording the finding.

Pilot acceptance should be explicit and measurable. Require at least 95 percent data availability on critical channels, at least 70 percent of alarms confirmed by an independent check, at least two actionable findings that lead to a planned repair, and a documented false alarm rate below 15 percent. Also require a written operator and technician feedback round, because a technically correct system that the floor dislikes will not survive. The pilot hardware, software and integration budget is the 51,600 to 112,600 USD detailed in Chapter 5; add 8,000 to 18,000 USD for training, baseline work and the first two months of engineering review.

Phase two: expansion, months 6 to 18

Expand only to assets whose failure modes and economics were proven in the pilot. A typical second wave adds 90 to 160 channels across four to six more lines, reuses the same historian and dashboards, and focuses on the highest criticality assets listed in Chapter 2. Standardize the sensor types, mounting kits and tag conventions so that each additional line costs less than the last. In a well run rollout the incremental cost per line falls 25 to 40 percent after the first two, because the engineering templates, alarm logic and training material already exist.

Phase two is also when the maintenance workflow changes, which is the harder part. Work orders must be generated automatically at action level, spare parts must be linked to the predicted intervention, and the planning meeting must include the condition review as a standing item. A program that produces excellent predictions but still plans maintenance from a calendar will show no financial return. Budget 12,000 to 30,000 USD for the interface to the maintenance system and the process work around it, and 6,000 to 14,000 USD for training the wider team.

Phase three: optimization, month 18 onward

At this stage the plant has 12 to 24 months of history. Use it to tune intervals, retire channels that have never influenced a decision, add level three analysis to the few assets that justify it, and convert the remaining calendar based tasks into condition based tasks where the data supports the change. Re-score criticality annually, because product mix and line loading change. Review every avoided event and every missed event in a quarterly meeting, and feed both back into thresholds. The ongoing cost is typically 18,000 to 45,000 USD a year for software support, calibration, sensor replacement and a part time engineer, which is small compared with the savings line it protects.

Team, skills and training

A program of this size does not need a large permanent team, but it does need clear ownership. One maintenance engineer with condition monitoring responsibility at 0.3 to 0.5 of their time owns thresholds and findings. One controls engineer at 0.2 to 0.3 owns data acquisition and tag quality. A production planner spends two hours a week on the condition review. A maintenance technician owns calibration and sensor physical condition. Naming these four roles explicitly, with a deputy for each, matters more than the job titles. Training costs 900 to 2,400 USD per person for vibration analysis fundamentals, 600 to 1,500 USD for the software platform, and 300 to 800 USD per operator for a short session on what the alarms mean and what to do when one appears.

Risks that derail the rollout and the practical countermeasures

Six risks recur. The first is a pilot chosen on a quiet, well maintained line that produces no findings, which makes the program look worthless; choose a line with known chronic issues but supportive operators. The second is starting with alarms on day one, before baselines exist; run in observation mode. The third is a dashboard nobody opens; put the single most useful number on the existing production board instead of building a new screen. The fourth is an unfunded sensor replacement plan; budget 3 to 6 percent of hardware value per year for failed devices, cables and batteries. The fifth is losing the engineer who built the thresholds; require the tag dictionary, alarm rationale and baseline notes to be written down as a deliverable of every phase. The sixth is a savings claim that finance cannot verify; freeze the baseline in phase zero and keep the avoided event log with the work orders.

What to put in the purchase specification

Whether the equipment is new or retrofitted, the specification should require open data access with documented tag lists, a time synchronized controller, accessible bearing housing measurement pads on all drives above a defined rating, spare analog inputs in the control cabinet, a dedicated industrial network segment with a documented interface, and as built documentation of sensor positions. Ask for the machine to be delivered with a baseline vibration and thermal record from the factory test run; that record is worth more than any generic acceptance limit because it is a health baseline for that exact machine. These clauses add little to the purchase price when included at ordering, and they can add 8,000 to 25,000 USD in retrofit engineering if discovered after installation. Wanplas lines can be specified with these provisions during the quoting stage, and the plant engineering team should include them in the technical annex of the order.

9. Work Orders, Spare Parts and Production Scheduling

A prediction only saves money when it becomes a planned job

The most disappointing predictive programs are technically excellent and operationally inert. They detect a bearing defect six weeks early, display it on a well designed screen, and then watch the machine fail on schedule because nobody turned the finding into a job with a date, a part and a crew. The value chain has five links: signal, confirmed finding, work order, available part and planned window. Break any one and the earlier work has no financial effect. Design the workflow before the first alarm goes live, and test it with a simulated finding during the pilot.

A practical sequence is as follows. The analytics layer raises a warning. The maintenance engineer checks the confirmation matrix from Chapter 6 and records a verdict of confirmed, plausible, or dismissed with a reason code. A confirmed finding moves to action level and generates a work order automatically, pre filled with the asset tag, the suspected failure mode, the trend plot, the projected intervention window, the parts likely needed, the estimated labor hours and the safety requirements. The planner checks the production schedule and proposes a window. The storeroom confirms or orders the parts. The job is executed, the as found condition is recorded with a photograph, and the baseline is reset as described at the end of Chapter 6. The loop closes with a short review of whether the predicted condition matched what was found.

Fields that a condition based work order must contain

A standard preventive work order is not sufficient. Add the asset tag and measurement position, the alarm level and the date it was raised, the feature values at the time of the alarm and at the last healthy reading, the operating band and recipe during the alarm, the independent confirmation measurement and its result, the projected earliest and latest intervention dates with confidence, the parts list with storeroom location, the estimated repair duration, the required isolation and safety steps, and a mandatory as found field. The as found field is the one that makes the program learn. If the work order is closed with a note such as “bearing replaced, outer race spall confirmed”, the next model has a labeled event. If it is closed with “fixed”, the history is worthless.

Spare parts policy driven by the prediction

Condition monitoring changes inventory economics in a useful way. A reactive plant keeps a broad, expensive stock because it cannot know what will fail. A calendar based plant keeps a stock tied to the replacement schedule. A condition based plant can hold a smaller stock of high value parts and order most items against a known intervention date, provided the warning window exceeds the procurement lead time. Make this explicit in a table for every critical part: unit cost, normal lead time, expedited lead time, emergency freight premium, storage cost, and shelf life.

Where the warning window is shorter than the lead time, keep the part. Where the window is longer, order it. Where the part is very expensive and the window is long, negotiate a supplier hold or a frame agreement rather than tying up capital. On a plant with 8 lines, moving from a broad reactive stock to a condition driven policy typically releases 25,000 to 90,000 USD of working capital, although that is a one time cash benefit rather than an annual saving and should be reported separately from operating savings.

Fitting the repair into the production schedule

A predictive finding has value precisely because it is not urgent. Use that freedom. The planner should have a weekly view of all action level findings with their earliest and latest safe dates, and should match them to planned changeovers, material changes, cleaning stops, mold changes, scheduled shutdowns and low demand periods. Combining a bearing replacement with a planned screw pull or a color change can reduce the effective downtime to nearly zero, because the line was going to stop anyway. On extrusion lines, avoid a mid campaign stop when a product with tight tolerances is running; target the transition to a less sensitive product instead.

Give the planner a simple economic comparison rather than a raw alarm. If the projected window is 12 to 26 days, and a changeover is planned in 9 days, the choice is between stopping 3 days early at a small cost in lost campaign length, or waiting 21 days for the next changeover with a rising but still acceptable risk. Showing the hourly loss, the repair duration and the confidence band makes that a normal planning decision. Keep a record of the choice and the outcome; over a year these records show whether the program’s warning windows are being used well or routinely compressed into emergency work.

Closing the learning loop

Every completed job should produce one of four findings: the predicted condition was confirmed, a different condition was found, nothing abnormal was found, or the job could not be completed as planned. The second and third categories are the ones that improve the program. A different condition may reveal a failure mode that was missing from the register; nothing found may reveal a threshold that is too sensitive, a sensor that was installed in the wrong position, or an operating band that was not segmented. Review these cases monthly, with the trend plot on screen and the technician who opened the machine in the room. Budget 2 to 4 engineering hours a week, roughly 6,000 to 14,000 USD a year, for this review. No algorithm substitutes for it.

10. Data Quality, Alarm Fatigue and Model Validation

Data quality is the leading cause of program failure

When a predictive program loses credibility, the root cause is usually mundane: a sensor was knocked loose during a cleaning shift, a gateway clock drifted by four minutes after a power cut, a tag was renamed during a controller upgrade, a cable was rerouted and picked up interference, or an operating state was never labeled. The dashboard still shows plausible numbers, so nobody notices until a real fault is missed. Treat data quality as a maintained asset with its own checks, owner and budget.

Calibration and physical checks that keep the numbers trustworthy

Set an annual calibration cycle for vibration channels, temperature probes, pressure transmitters and flow meters, with a shorter six month check on instruments used for quality or safety decisions. A portable vibration calibrator costs 1,400 to 3,600 USD and can check an entire plant; a dry block temperature calibrator costs 900 to 2,400 USD; a pressure test pump with a master gauge costs 500 to 1,600 USD. Together these tools, at 2,800 to 7,600 USD, cover the majority of a mid sized plant’s channels and remove the need for a service contract on every device. Budget 90 to 260 USD per point per year if the work is outsourced instead.

Physical inspection is equally important and cheaper. During a monthly route, check that the sensor is tight, the cable is clipped away from hot surfaces and moving parts, the connector is clean and dry, and the mounting pad matches the photograph in the asset register. A two minute check per point prevents the single most common source of bad vibration data. Record every removal and refit in the asset register, because the baseline is only valid for a known mounting condition.

Classifying alarms so they can be improved

Every alarm should end in one of five categories: a true positive leading to a planned repair, a true positive where the decision was to keep running under watch, a false positive caused by operating context, a false positive caused by a data or sensor fault, or a false negative where a failure occurred without warning. Track all five counts monthly. A program with many context related false positives needs better operating band segmentation, not a higher threshold. A program with sensor related false positives needs a maintenance round, not new analytics. A false negative is the most valuable event of all: reconstruct the trend that preceded it, ask whether the signal existed but was filtered away, and add that failure mode to the register.

Aim for a confirmation rate above 70 percent at action level and a false negative count of zero or one per year on critical assets. Do not aim for zero false positives. A system tuned until it never produces a questionable alarm has almost certainly been tuned until it also misses slow degradation. A modest number of explainable warnings is the price of sensitivity, and the weekly review exists to absorb them.

Validating a model before it influences a stop decision

Validation has three stages. Back testing uses archived data to ask whether the rule would have caught known past events with enough lead time, and how many false alarms it would have produced in the same period. Shadow running applies the model to live data without sending alarms for 8 to 12 weeks, with an engineer reviewing its output each week. Controlled deployment then allows the model to raise warnings only, with action level still requiring human confirmation, for a further 2 to 3 months. Only after these stages should a model be allowed to generate work orders automatically on a critical asset.

Use time based splits, not random splits, when back testing. Hold out entire production campaigns so that the model is tested on periods it has never seen, and include at least one period with an unusual product or an extreme ambient condition. Report lead time, not just accuracy: a model that detects a bearing fault with 95 percent accuracy but only 2 days of warning is less useful than one with 80 percent accuracy and 6 weeks of warning, because the second one allows a planned repair. Record the version of every rule, the date it was changed and the reason, in a change log that production can read.

Alarm fatigue and the human side of the system

People stop responding to alarms that cry wolf. Cap the number of alarms reaching any one role. A maintenance engineer should not receive more than 8 to 12 actionable items a week from a pilot sized system; anything above that belongs in a digest, not in a notification. Never send an advisory to a phone. Show operators only what they can act on: current machine state, the next planned intervention, and whether any reading is outside its normal band. Give the maintenance engineer the detail view. Give management a monthly trend of avoided events and savings, not a live alarm feed.

Also protect the system’s reputation internally. When a prediction leads to a planned repair, say so explicitly in the shift briefing and in the monthly report. When a prediction is wrong, say that too, with the corrective action. A program that is seen as honest about its misses keeps the trust of the people whose judgment it depends on. The alternative, a program that claims every success and hides every miss, is usually switched off within two years of the engineer who championed it leaving.

Security, access and change control

Segment the condition monitoring network from the office network and from public access. Use read only data extraction from the controller wherever possible, so a diagnostic system can never write to a machine control tag. Keep the historian and dashboards behind authenticated access with role based permissions, and log every threshold change, suppression and manual override. Back up the tag dictionary, alarm rationale and baseline archives with the same discipline as the controller program. A segment firewall and a documented access policy for a mid sized plant costs 3,200 to 9,500 USD installed, and an annual review at 1,200 to 3,000 USD. This is not an optional extra: a system with access to production data and a path to the control network is part of the plant’s operational risk surface, and it should be treated that way from the pilot onward.

11. KPIs, Financial Tracking and Continuous Improvement

Separate leading indicators from lagging indicators

A predictive program is often judged on lagging indicators such as downtime hours, which take a year to move and are influenced by order mix, staffing and product changes. Leading indicators show whether the program itself is working and move within weeks: data availability, alarm confirmation rate, number of findings raised, average warning lead time, proportion of work orders generated from condition findings, and the share of repairs completed inside the predicted window. Report both, but manage the program on the leading ones. If leading indicators are healthy, the financial result follows. If they are not, a good quarter is probably luck.

A savings ledger finance will accept

Create one ledger from day one, with a line for every avoided event. Each line needs the date, the asset, the alarm level, the confirmed finding, the work order number, the actual repair cost, the avoided emergency cost estimated from the plant’s own history, and the name of the person who approved the estimate. Where a cost is avoided rather than incurred, finance will reasonably ask how you know. The answer must be a documented calculation, not an assertion. Use the plant’s own last 24 months of records to build an average emergency event cost per asset family, and freeze that value at the start so it cannot be adjusted later to flatter the result.

Keep four categories separate so nothing is counted twice: repair and downtime savings, material and quality savings, energy savings, and logistics savings such as freight and overtime. Working capital released from inventory is a one time cash effect and belongs in a separate line, never in annual operating savings. Report gross savings and net savings after the program’s own operating cost, and show both. A monthly one page report with these six figures, signed by maintenance and finance, is more persuasive than a long slide deck.

Reporting rhythm and ownership

Operators see current machine condition and the next planned intervention on the shift board. The maintenance engineer reviews alarms and data quality daily and leads a 30 minute weekly condition meeting with the planner. A monthly report goes to the plant manager and finance with the KPI table and the savings ledger. A quarterly review covers misses, threshold changes, criticality re-scoring and the retirement or addition of measurement channels. An annual review resets the baseline and decides whether to extend the program to new lines. Each of these meetings has a named chair and a written output; without that, the rhythm collapses within two months of a busy production period.

Avoiding the two ways KPIs get gamed

The first is reclassifying an unplanned stop as a planned stop after the fact. Prevent it by requiring that a stop be labeled at the moment it occurs, with the category locked in the maintenance record. The second is claiming every repair as a predictive success. Prevent it by requiring a condition finding dated at least 7 days before the work order for the event to count as an avoided failure. These two rules are simple, cost nothing, and protect the credibility of the whole financial case. A program that shows a modest honest saving survives budget reviews; one that shows an implausible large saving is usually cut in the second year.

12. Full US Dollar Cost, Savings and Payback Analysis

Three investment levels and what each one buys

The lean option concentrates on controller tag extraction, a handheld survey route and a light software layer. It suits a plant with 3 to 5 lines and a limited budget, and it delivers most of the value from thermal, load and process trends, but little early warning on bearings and gearboxes. The standard option is the two line pilot followed by measured expansion described in Chapter 8. The comprehensive option adds continuous vibration on every critical rotating asset, inline oil monitoring and level three analytics on the highest consequence machines.

Where the annual saving comes from

The following breakdown uses the eight line plant from Chapter 1 and the standard option. It deliberately keeps the categories separate, with the calculation shown so that a finance reviewer can challenge any line. The repair and downtime figure is the difference between the calendar based and condition based rows in the Chapter 1 strategy table. The other rows are additional and were explicitly excluded from that comparison.

Working capital released from a smaller spare parts inventory, estimated at 25,000 to 90,000 USD, is deliberately excluded from the annual figures above. It is a one time cash benefit and should appear as a separate line in the capital request, clearly labeled as non recurring.

Five year cash flow and simple payback

Using the standard option at 185,000 USD of capital spread across the first two years, with benefits ramping as the pilot and then the expansion come online, the cash flow looks as follows. Year one includes the pilot investment and a partial year of benefit. Year two includes the expansion investment and most of the benefit. Year three onward is steady state.

Simple payback falls in month 37 to 38, roughly three years from the start of phase zero. At a discount rate of 8 percent, the five year net present value is approximately 102,000 USD and the internal rate of return is in the range of 27 to 32 percent. Steady state return on the installed capital is about 45 percent a year after running costs. These are healthy but not spectacular numbers, which is exactly what a realistic maintenance case should look like. If a proposal shows a six month payback for a whole plant, check whether the same saving has been counted in two categories.

Sensitivity: what if the savings are smaller

Assume only 55 percent of the projected gross saving materializes, about 63,600 USD a year, because the plant has fewer failures than expected or the team cannot act on the findings quickly. Net steady state cash flow falls to 31,600 USD, the five year net present value turns slightly negative at roughly minus 12,000 USD, and payback moves beyond five years. Assume instead that the program performs better than expected at 140,000 USD a year, because a single avoided gearbox failure on a critical line was worth more than modeled: net present value rises to about 155,000 USD and payback arrives around month 30.

The sensitivity result is the most useful part of the analysis for a decision maker. It says the investment is not justified by the technology but by the organization’s ability to act on it. The three variables that matter are the hourly loss of the covered assets, the number of avoidable events per year, and the speed with which a confirmed finding becomes a planned job. Improve the third variable and the same hardware produces a much better return. This is why Chapter 9 and Chapter 11 are not administrative extras; they are where the financial result is actually produced.

Costing mistakes to avoid

Five errors appear repeatedly in capital requests. The first is counting the full cost of an avoided stop as saving when the production would have been recovered by overtime anyway; count only the true incremental loss. The second is double counting a repair saving inside both the maintenance budget and the downtime line. The third is omitting the annual running cost of calibration, software support and engineering time, which typically adds 15 to 25 percent to the five year cost. The fourth is assuming 100 percent detection on every failure mode, when sudden electrical failures and safety device faults are not predictable. The fifth is treating released inventory as a recurring benefit. Correct these five and the resulting case is usually still positive, and far more defensible.

13. Wanplas Equipment and Predictive-Ready Specification

Map the program to the machine families on the line

Wanplas covers the full plastic production chain, so a predictive program can be specified consistently across a plant rather than assembled from incompatible subsystems. On compounding and pelletizing equipment, including lab twin screw extruders, parallel twin screw extruders, triple screw extruders, double stage systems, Banbury kneader lines and micro pellet lines, the priority channels are gearbox vibration, main motor current and torque, barrel zone duty cycle, melt pressure and temperature, and specific energy per kilogram. On recycling equipment, including food grade PET crushing and washing lines, soft and hard PP and PE washing lines, rigid flake pelletizing and new generation pelletizing lines, add granulator rotor vibration, knife load current, wash water flow and contamination trend, and dryer thermal balance.

On extrusion lines for pipe, sheet, film, board and profile, focus on haul-off and roll stack drive vibration, encoder speed deviation, die zone temperature uniformity, cooling circuit flow and approach temperature, vacuum pump performance, and thickness or weight per meter drift. On molding equipment, including extrusion blow molding, PET bottle blowing, injection blow molding and injection molding machines, focus on hydraulic pressure rise time, clamp force balance, mold cooling balance, ejector force, heater duty and high pressure compressor condition. On filling equipment, including linear and rotary blowing-filling-capping combiblocks, focus on filling valve cycle counts, fill volume deviation, main drive vibration and vacuum decay. In every family the same rule holds: specify the measurement positions and the data interface at the ordering stage, not after commissioning.

Budget ranges by machine family

The ranges below are planning budgets in US dollars for complete equipment, not quotations, and they depend heavily on capacity, configuration, degree of automation and downstream scope. They are included because a predictive program is usually justified as a percentage of the asset it protects, and because a new line with monitoring provisions costs little more than one without them.

What to ask for in the technical annex

Six clauses make a new line predictive ready at very little cost. First, a documented tag list with engineering units and scaling for every drive, temperature zone, pressure, flow and speed value. Second, a controller with time synchronization and a documented read only interface. Third, machined measurement pads on all bearing housings of drives above an agreed rating, with photographs and positions recorded in the handover file. Fourth, oil sampling ports on gearboxes and hydraulic reservoirs that can be used while the machine is running safely. Fifth, spare analog and digital inputs in the control cabinet, plus a spare network port and a dedicated segment for diagnostic traffic. Sixth, a factory test run record containing baseline vibration, temperature, current and output values for the delivered machine, which becomes the starting baseline for the plant.

Wanplas supports this kind of specification work with water, electricity and gas design, three dimensional workshop layout, and a continuous running test before shipment. The group also provides USD 500 of free parts each year and free replacement for damaged parts within the warranty period, which fits well with a condition based spares policy because the parts that the data identifies as wearing can be planned into that allowance rather than bought in an emergency. Ask for the measurement provisions and the baseline record at the quotation stage; retrofitting them after installation typically costs 8,000 to 25,000 USD per line and involves a shutdown that a new order does not.

Choosing where to start on an existing Wanplas line

For plants already running Wanplas equipment, start with the highest torque rotating asset and the widest thermal system. On a compounding line that is the main gearbox and the barrel heating zones. On a pipe or sheet line it is the haul-off or roll stack drive and the cooling circuit. On a molding line it is the hydraulic power unit and the mold water balance. These are the positions where a modest number of channels produces the shortest payback, and they are the same positions where Wanplas lines can be retrofitted without disturbing the machine control logic. Once two of them have produced a documented avoided event, the internal case for extending the program to the rest of the plant becomes straightforward.

14. Frequently Asked Questions

Conclusion

Using predictive maintenance for plastic equipment to reduce long term operation costs is less about buying sensors than about changing when a decision is made. The expensive part of a failure is almost never the failed component. It is the hot barrel full of degrading material, the air freight on a part that should have been on the shelf, the crew pulled off other work, the out of tolerance product made while a zone drifted, and the delivery that moved by a day. A measured condition shifts that whole cluster of cost into a planned window where a smaller repair, a normally purchased part and an existing changeover absorb it.

The route is straightforward. Rank assets by consequence, tie each measurement channel to a named failure mode, extract the controller data before buying hardware, build baselines in defined operating bands, confirm every alarm with an independent signal, and make the finding into a work order with a part and a date. Track leading indicators weekly, keep a savings ledger that finance can audit, and review every miss as carefully as every success. The technology at levels one and two is mature and inexpensive; what determines the return is whether the organization acts early on what it learns.

For a typical eight line plant the standard program needs 163,000 to 250,000 USD of capital over 18 months, delivers about 115,700 USD of gross annual saving at steady state, and pays back in roughly 37 to 38 months. A lean start costs under 72,000 USD and can pay back in under two years. Whether the equipment is a compounding extruder, a recycling washing line, a pipe or sheet extrusion line, a blow molding machine or a filling combiblock, Wanplas machines can be specified at the ordering stage with the measurement pads, tag lists, data interface and factory baseline record that make this work straightforward. Ask for those provisions in the technical annex, start with the one asset whose failure would hurt most, and let the first documented avoided event fund the rest of the program.