Disclaimer

This content is provided for information only. The authors make no representation or warranty regarding the accuracy, completeness or currency of the content. No information in this whitepaper should be construed as medical advice. Readers should seek appropriate professional guidance before acting on any information contained in this document. The authors expressly disclaim all liability for any direct or indirect loss or damage arising from the use of or reliance on this information.

Introduction

This paper sets out how to implement predictive maintenance (PdM) for steam sterilisers and washer-disinfectors, including bedpan washer-disinfectors, using connected sensors, known as the Internet of Things (IoT), and data analytics. It defines the failure modes to monitor, the parameters and sensors that indicate them, analytics methods suited to the data available, the system architecture, compliance and cyber security controls, and a method to test whether PdM pays for a given fleet. PdM supplements the planned maintenance, requalification and routine monitoring required by AS 5369:2023. It does not replace them.

Failure Modes and Maintenance Requirements

Requirements: AS 5369:2023 requires planned preventive maintenance by a competent person to documented procedures, recalibration at intervals not exceeding 12 months, an annual electrical safety check and requalification at least annually. Preventive maintenance intervals follow the instructions for use (IFU). PdM can trigger additional maintenance and, where the IFU permits, support condition-based service intervals between the mandatory annual activities.

Washer-Disinfector Failure Modes: Heating element failure, steam generator scale, circulation and drain pump bearing and seal wear, solenoid valve failure, blocked inlet strainers and spray nozzles, dosing pump tube wear, door seal degradation and temperature sensor drift.

Steam Steriliser Failure Modes: Vacuum pump wear, door seal and valve leakage, steam generator element failure and scale, steam trap failure, and pressure and temperature sensor drift.

Effect of Failure: A failed bedpan washer-disinfector forces the ward to use a machine in another area or single-use containers. Do not clean bedpans manually. A failed CSSD steriliser reduces reprocessing capacity for every user of the department. Unplanned repairs add expedited freight, after-hours labour and outage time to the planned repair cost.

Monitored Parameters and Sensors

Extract one set of features for each cycle from the controller record and added sensors, and trend each feature against its baseline for that machine and cycle type.

• Heat-Up Time: Record the time from fill to disinfection or sterilisation temperature, e.g. a steriliser that normally reaches 134 deg C [273 deg F] in 5 min and now takes 7 min. At constant inlet water temperature, a rising heat-up time indicates loss of heating power, e.g. a failed element in a multi-element heater, a degraded contactor, low supply voltage or element over-temperature cut-outs caused by scale.

• Heater Current: Measure current on each heater circuit with a current transformer. A step reduction identifies a failed element.

• Thermal Disinfection Dose: Record temperature and time at disinfection temperature for each washer-disinfector cycle and calculate the A0 value achieved. A falling margin above the specified value, e.g. A0 600, equivalent to 90 deg C [194 deg F] for 1 min, indicates heater, sensor or control drift.

• Fill and Drain Time: At constant supply pressure, a rising fill time indicates a blocked inlet strainer or failing solenoid valve. Measure inlet pressure to separate supply faults from machine faults. A rising drain time indicates a blocked drain or failing drain pump.

• Pump Current and Vibration: Measure motor current and casing vibration on circulation and vacuum pumps. Rising vibration amplitude at running speed or bearing defect frequencies indicates bearing wear or imbalance. Rising current at constant duty indicates mechanical drag.

• Vacuum Performance: For sterilisers, record the time to reach each vacuum pulse set point and the air leakage test result. A rising pulse time indicates vacuum pump wear, high seal water temperature or a leak. A rising leak rate indicates door seal or valve leakage.

• Steam Saturation: During the holding time, compare measured chamber temperature with the saturation temperature for the measured chamber pressure. A growing difference indicates non-condensable gas, superheat or sensor drift.

• Water Quality: Measure supply water conductivity or hardness, and final rinse and steam generator feedwater conductivity. Cumulative hardness load indicates scale formation on heaters and nozzles before performance changes.

• Dosing Volume: Measure detergent volume per cycle with a flow sensor or container level. A falling volume indicates dosing pump tube wear.

• Cycle Count, Faults and Aborts: Record cycle count, fault codes and aborted cycles. Schedule wear part replacement on cycle count rather than calendar time.

Installation: Read controller data through existing data or printer ports in read-only mode. Do not change control parameters, firmware or the wiring of process sensors. Seal sensor penetrations to the ingress protection rating of the machine, perform an electrical safety test after installation, and requalify after any change that could affect the process.

Data Analytics and Machine Learning

Rules and Statistical Control: Start with rules derived from the validated cycle limits, e.g. disinfection temperature not reached or cycle time exceeded. Then apply statistical process control to each per-cycle feature, with control limits set from a baseline of at least 30 normal cycles for each machine and cycle type.

False Alert Rate: Size the alert logic to the planned false alert rate. At limits of 3 standard deviations, 0.27 % of normal cycles breach each limit, so 10 features at 20 cycles per day on 260 days per year produce about 140 false alerts per machine per year. Alert on persistence instead, e.g. 3 consecutive breaches, which reduces the theoretical rate to about 0.001 per machine per year for independent, normally distributed features. Confirm the actual rate in the pilot, because cycle data is autocorrelated and not normally distributed.

Degradation Trending: Fit a trend to each degradation feature, e.g. heat-up time or vacuum pulse time against cycle count, and project when it will reach its action limit. The projected time is the lead time available to plan parts and labour.

Machine Learning: Supervised models that predict failure or remaining life need labelled failure histories for each failure mode. A fleet of 40 machines at 0.75 failures per machine per year records 30 failures per year, spread over several failure modes, so few labelled examples accumulate for each mode. Use unsupervised anomaly detection on per-cycle features, e.g. isolation forest or autoencoder models, until enough failures are recorded. Validate every model on data recorded after its training period, and report precision, recall and lead time for each failure mode.

Published Results: Niyonambaza et al. (2020) logged heater and pump temperatures on one hospital steam steriliser for about 3 months and used a long short-term memory (LSTM) network to forecast them, reporting accuracy of about 90 % and 96 %. The study did not predict an actual failure. Shamayleh et al. (2020) used vibration features and a support vector machine to detect belt slippage in a laboratory analyser, and estimated diagnostic and repair cost savings of up to 25 % with a payback of 1 year for that application. Treat published accuracies and savings as specific to the equipment and data studied.

Diagnosis: Map each alert to a probable component from the feature that triggered it, e.g. heater current to the heating element and pulse time to the vacuum pump, so that the technician attends with the correct parts.

Implementation and Architecture

• Failure Mode Review: From the maintenance records of each machine, list the failure modes, their effect on service and their detectable precursors. Monitor only parameters that precede a failure mode with useful lead time.

• Network and Security: Connect gateways to a segregated network segment with no inbound access from the internet, and manage the connection under IEC 80001-1:2021. Apply the facility's cyber security controls to gateways and servers. The Therapeutic Goods Administration states that users share responsibility for providing a secure operating environment for networked medical devices.

• Platform and Integration: Store per-cycle features and alerts against machine, cycle number and a time stamp synchronised to the facility time source. Integrate alerts with the computerised maintenance management system (CMMS) so that each alert creates a work order. Require data export in an open format in the supply contract.

• Pilot: Pilot on a subset of machines for 3 to 6 months with rule and statistical control alerts. Record every alert outcome as true or false, and every failure without a prior alert as missed. Expand only when the measured detection rate and false alert rate meet the business case.

• Response Rules: Respond to an alert first by remote review of the trend, with local staff checking the machine in the field. Monitor until the trend confirms a fault, then schedule the repair, so that a technician does not attend without a confirmed fault. Withdraw a machine from service when a cycle fails its release criteria, regardless of its PdM status.

• Review: Review alert outcomes quarterly, retune control limits and retrain models on the extended data set.

Cost Model and Worked Examples

Method: Net annual value equals the detection rate multiplied by annual failures and by the saving per predicted failure, plus the cost of maintenance visits omitted, less the annual cost of the PdM system. The detection rate is the share of failures predicted with enough lead time to plan the repair. Annualise hardware and setup over their life at the real discount rate. Calculate the break-even detection rate, estimate the achievable rate from the failure history, and confirm it in the pilot. The following values are illustrative, in Australian dollars excluding GST.

Bedpan Washer-Disinfector Fleet: A hospital with an on-site technician operates 40 bedpan washer-disinfectors at 20 cycles per day.

Cost Structure: The PdM system has one-off capital costs and annual recurring costs. To compare capital with annual savings, spread it over the 5-year life at 5 % using a capital recovery factor of 0.231, so A$22,000 of capital is equivalent to A$5,081 per year. The only recurring cost to the facility is the platform licence of A$4,000 per year. The service provider absorbs data review and false alert review within its service.

Detection Rate Estimate: Estimate the achievable detection rate as the sum, over failure modes, of each mode's share of failures multiplied by the share of that mode with a detectable precursor. The following illustrative values give 51 % with retrofit sensors and 32 % with controller data only. Open-circuit element failures and electronic faults give little warning in either configuration. Replace the values with the failure history of the fleet.

Configurations: Each predicted failure saves A$1,300. The four configurations compare as follows at their estimated detection rates. Payback is capital divided by the annual benefit less the platform licence. A break-even rate of 0 % means the omitted visits alone cover the annual cost.

Result: All four configurations exceed their break-even detection rate. Without omitted visits, retrofit sensors give A$10,809 per year, a payback of 1.4 years and a 5-year net present value of A$46,795 at 5 %, and controller data gives A$5,246 per year with a payback of 1.7 years. With one omitted planned maintenance visit per machine per year, where the IFU permits, the omitted visits alone cover the annual cost. Retrofit sensors then give A$32,809 per year, a payback of 0.6 years and a 5-year net present value of A$142,044, and controller data gives A$27,246 per year with a payback of 0.5 years. Retrofit sensors give the higher net value because they detect more failure modes. Confirm that the IFU permits condition-based intervals before relying on omitted visits. Allow for a lower detection rate in the first year while baselines form.

Detection Sensitivity: The following table gives net annual value at detection rates of 25 %, 50 % and 75 %, in thousands of Australian dollars. Every configuration gives a net gain at each rate.

Input Sensitivity: The following table gives the break-even detection rate for retrofit sensors without omitted visits as one input varies.

CSSD Fleet: A CSSD operates 4 steam sterilisers and 4 washer-disinfectors with redundancy, with the following illustrative values.

Capital is A$6,000, equivalent to A$1,386 per year over 5 years at 5 %. With the platform licence of A$1,600 per year, the system costs A$2,986 per year. Each predicted failure saves A$5,900. The break-even detection rate is 6 % without omitted visits. At a detection rate of 50 %, net annual value is A$20,614 and payback is 0.3 years. At 25 %, net annual value is A$8,814.

Application: With low-cost retrofit hardware and data review absorbed by the service provider, PdM pays at modest detection rates in both fleets. It pays most readily on CSSD sterilisers and washer-disinfectors, where each predicted failure avoids more downtime cost. For bedpan washer-disinfectors, retrofit IoT hardware gives the higher net value, and condition-based service intervals that the IFU permits add the largest single saving. Review alerts remotely with local staff, and send a technician only when the trend confirms a fault.

Benefits and Limitations

• Planned Repairs: A predicted failure is repaired at a scheduled time with parts on normal freight and a shorter outage. The reduction in unplanned downtime equals the detection rate achieved, not the number of alerts issued.

• Condition-Based Maintenance: Where the IFU permits, service machines with stable trends at the mandatory annual visit and machines with deteriorating trends when the trend requires. Do not extend recalibration, electrical safety testing or requalification beyond 12 months.

• Records: Electronic cycle records support traceability and audit. AS 5369:2023 advises that facilities should be working towards electronic tracking and process record systems. The cycle record remains the basis for load release.

• Fleet Data: Compare failure rates by model, location and water supply to inform procurement, machine allocation and water treatment.

• Life Extension: Count life extension in the business case only where failure records show secondary damage that early intervention prevents.

• Limitations: PdM does not replace requalification, routine monitoring or cycle release. False alerts consume review time. Retrofit sensors can affect validation and electrical safety. Connected devices add cyber security risk. Supplier platforms create data dependence unless open data export is specified.

Conclusion

Predictive maintenance for sterilisers and washer-disinfectors uses per-cycle features from the controller record and targeted sensors, trended with rules, statistical process control and degradation projections, with machine learning added as labelled failure data accumulates. Its value depends on the measured detection rate, the false alert rate and the cost of an outage. In the illustrative 40-machine bedpan washer-disinfector fleet, with retrofit IoT hardware at A$500 per machine and data review absorbed by the service provider, retrofit sensors break even at a 23 % detection rate and return A$10,809 per year at the estimated 51 %. With one omitted planned visit per machine per year, where the IFU permits, the return rises to A$32,809 per year. In an 8-machine CSSD fleet, where each predicted failure saves A$5,900, retrofit IoT hardware breaks even at a 6 % detection rate. Pilot on a subset of machines, measure detection and false alert rates, and expand only where the measured rates meet the break-even values. Maintain all AS 5369:2023 maintenance, requalification and routine monitoring activities throughout.

Sources

1. Standards Australia. AS 5369:2023 Reprocessing of reusable medical devices and other devices in health and non-health related facilities.
2. Niyonambaza I, Zennaro M, Uwitonze A. Predictive Maintenance (PdM) Structure Using Internet of Things (IoT) for Mechanical Equipment Used into Hospitals in Rwanda. Future Internet 2020, 12(12), 224.
3. Shamayleh A, Awad M, Farhat J. IoT Based Predictive Maintenance Management of Medical Equipment. Journal of Medical Systems 2020, 44, 72.
4. IEC 80001-1:2021 Application of risk management for IT-networks incorporating medical devices, Part 1: Safety, effectiveness and security in the implementation and use of connected medical devices or connected health software.
5. Therapeutic Goods Administration. Medical device cyber security information for users.
6. AS/NZS 3551 Management programs for medical equipment.