Asset health AI for Australian manufacturers uses sensor and historian data to detect bearing wear, lubrication breakdown, electrical anomalies and thermal drift weeks before a failure, then raises a work order in the CMMS for a planner to approve. It pays off when it sits inside an ISO 55001 asset management system, measures OEE to ISO 22400, and respects IEC 62443 security.
When a critical asset fails, a line can be down for a full shift. Compressed air, CIP skids, hydraulic presses, conveyor drives, robotics and refrigeration all carry that risk, and running them until they break is hard to defend.
This guide is for operations directors, plant managers and maintenance managers in food and beverage, contract manufacturing, metal fabrication, plastics, packaging and automotive parts. Asset health AI gives your maintenance crew earlier warning, better diagnostics and an audit trail. It does not replace them.
Key takeaways
- Measure OEE per asset to ISO 22400 before buying any AI. AI on bad baselines produces confident nonsense.
- Most plants sit between preventive and condition-based maintenance. Jumping straight to prescriptive AI is the most common pilot failure.
- Every connection from OT to IT is attack surface: segment to IEC 62443, read-only data out, human approval on safety-critical work orders.
- Integration with the historian and CMMS matters more than the algorithm.
- Start on one asset class with a documented exit criterion.
How does ISO 55001 frame asset health AI?
AS/NZS ISO 55001:2024 defines the asset management system. It sits above the maintenance manual and ties objectives (uptime, cost, safety, quality, sustainability) to lifecycle decisions on every asset.
| Stage | What it covers | Where AI contributes |
|---|---|---|
| Asset strategy | Criticality, replace vs refurbish, capex | Failure-mode prediction informs replacement timing |
| Acquisition and commissioning | Specification, FAT/SAT, baseline capture | Baseline vibration, thermal and current signatures for model training |
| Operation | Run-time, OEE, energy | Real-time anomaly detection on telemetry |
| Maintenance | PM schedules, condition monitoring, work orders | Predictive triggers feed CMMS work orders |
| Disposal | End-of-life, residual value | Asset history informs reinvestment decisions |
ISO 22400-2:2014 is the KPI layer underneath, defining Overall Equipment Effectiveness (OEE) as availability times performance times quality. Asset health AI mainly improves availability and performance, with quality gains where machine condition correlates with defects.
What does asset health mean on the factory floor?
Asset health is the gap between a machine's current condition and its design condition. Closing it early is cheaper than running to failure.
- SensorsVibration, thermal, ultrasonic, current, oil quality, acoustic, machine vision
- Edge ComputePLC, gateway, or edge box. Local feature extraction, hot-path anomaly checks
- HistorianAVEVA PI, Ignition, or equivalent. Time-series store for training and audit
- ML ModelsAnomaly detection, remaining useful life, root-cause clustering
- MES and CMMSDraft work orders, OEE feeds, parts reservation
- Maintenance ActionPlanned intervention with full evidence trail under ISO 55001
Wireless condition sensors are now far cheaper than the wired installations of a decade ago. The hard problem has moved from measuring to making sense of data inside existing operations systems. AI earns its place by finding patterns across thousands of historian hours, correlated with maintenance records and quality outcomes, that technicians know exist but cannot watch across hundreds of assets.
Where does predictive maintenance sit on the maturity curve?
Plants do not jump from reactive maintenance to prescriptive AI in one step.
| How it works | Typical trade-off | |
|---|---|---|
| Reactive (run to failure) | Fix when it breaks | Cheap to set up, costly unplanned downtime and scrap |
| Preventive (time-based PM) | Fixed schedule | Predictable, over-maintains many assets, parts waste |
| Condition-based | Sensor thresholds trigger inspection | Less waste, still reactive to thresholds |
| Predictive (AI/ML) | Models forecast failure days to weeks ahead | Needs clean baselines and failure history |
| Prescriptive (AI + decision support) | Recommends intervention, parts and skills | Only works on mature data and processes |
Most Australian manufacturers sit between preventive and condition-based. Starting prescriptive AI on assets without clean baseline data is the most common reason pilots fail. See our AI pilot success factors.
Why is IT/OT security non-negotiable for asset health AI?
Asset health AI needs data from PLCs, SCADA, historians and MES to reach ML platforms, the CMMS, ERP and reporting. Every connection is new attack surface.
The IEC 62443 series is the reference for industrial control system security that Australian manufacturers are expected to align with. Purdue-model segmentation, data diodes for one-way flows, role-based OT access and unidirectional gateways are table stakes.
If your plant is part of a critical infrastructure asset class (food and grocery, water, energy, communications), the Security of Critical Infrastructure Act 2018 may impose risk management obligations, and asset health platforms touch data in that risk register.
- Treat every wireless gateway as untrusted on a segmented network.
- Never expose PLC programming ports to the AI platform. Read-only OPC UA or MQTT egress only.
- Route every AI-raised work order on safety-critical assets through human approval.
- Document model versions, training data lineage and decision logic.
How does asset health AI integrate with SCADA, historian, MES and CMMS?
A typical plant runs some mix of:
- SCADA (AVEVA, Ignition, Rockwell FactoryTalk, Siemens WinCC)
- a historian (AVEVA PI, Canary, or the SCADA's own)
- an MES, often partial or in-house
- a CMMS (Maximo, SAP PM, Pronto Xi, MEX, or a spreadsheet someone calls a CMMS)
- an ERP (SAP, Pronto, MYOB Acumatica, Microsoft Dynamics).
A proven pattern:
- Read telemetry from the historian via OPC UA or REST.
- Stream high-frequency vibration and current data through MQTT to edge gateways.
- Extract features at the edge to cut bandwidth and protect raw data.
- Send features and anomaly scores to a central model store.
- Push work order recommendations to the CMMS through a documented API with human approval.
- Show OEE-aware health scores in the MES or plant dashboard.
Many plants run inference at the edge and send only aggregated scores for retraining and reporting, on infrastructure they control. See our data sovereignty guide and system integration service.
What changes for food and beverage plants under HACCP?
Hygiene zones constrain sensor choice: IP69K-rated sensors that survive caustic CIP regimes. Models watching pump performance during CIP can flag valve seat wear, blocked spray balls or cavitation before a non-conforming batch.
HACCP critical control points suit AI-augmented monitoring: chiller temperature drift, cooker pressure variance, tunnel pasteuriser dwell deviation. Food Standards Code 3.2.1 and 3.2.2 require documented evidence that CCPs are controlled. Pre-empting a CCP excursion strengthens your HACCP plan, in line with FSANZ expectations.
Any AI output used in a food safety decision needs version control, lineage and human sign-off. Pharmaceutical contract manufacturers under TGA PIC/S GMP apply the same rigour through ALCOA+ data integrity principles.
How does asset health AI support energy and NGER reporting?
Compressed-air leaks, over-cycling refrigeration, unloaded oversized motors and degraded heat exchangers show up as anomalies in current draw, kWh or thermal signatures. The same telemetry that drives maintenance feeds scope 1 and 2 reporting.
If your organisation meets NGER thresholds (25 kt CO2-e or 100 TJ at a facility, 50 kt CO2-e or 200 TJ at corporate group level), asset-level energy data supports your reporting and AASB S2 climate disclosures that lenders and customers increasingly request. Our own product Carbonly.ai automates NGER and ESG data processing using the same document and data patterns.
Which asset should you start with?
Starting on the wrong asset means twelve months proving nothing.
What is the single biggest operational pain point right now?
- One asset is the constant bottleneck on production throughputStart with the constraint asset
- Energy bill is increasing faster than outputStart with the top three energy consumers (motors, compressors, refrigeration)
- Customer complaints about product quality are risingStart with the quality-critical asset and pair AI health with QC inspection
- Recent safety incident or near-miss on a specific machineStart with safety-critical assets, align with AS/NZS 4024 risk assessment
- Insurance premium hit related to fire or equipment failureStart with insurer-flagged assets
Where quality is the driver, pair asset health with AI visual inspection.
What does a realistic 12-week asset health pilot look like?
- Week 1 to 2Asset selection and criticalityConfirm scope: one asset class or one line. Document failure modes, current PM regime, OEE baseline.
- Week 3Network and cyber posture reviewMap data flows against IEC 62443. Approve segmentation, gateway placement, OPC UA read-only.
- Week 4 to 5Sensor and gateway installationInstall during planned maintenance windows. Validate baseline data quality before training.
- Week 6 to 8Baseline capture and model trainingCollect three to six weeks of data. Train the first anomaly model with maintenance team input.
- Week 9 to 10CMMS integration and workflowConnect model outputs to draft work orders with human approval. Document decision logic.
- Week 11Run alongside the teamAI advises beside the existing PM regime. Compare predicted events with actual outcomes.
- Week 12Decision gate and scale planQuantify avoided downtime, integration debt, and go or no-go for rollout.
Three things make or break it:
- an accountable plant sponsor who owns the success criteria
- a baseline with equipment data and matching maintenance records, so the model learns from real failures
- a documented exit: if the AI does not beat the current PM regime on the chosen asset within 12 weeks, stop and reassess.
- Unplanned downtime hours on the pilot asset classBaseline vs pilot
- OEE on the monitored line (ISO 22400)Baseline vs pilot
- Scrap and rework on quality-correlated assetsBaseline vs pilot
- kWh on monitored assetsBaseline vs pilot
- Predicted events confirmed by techniciansHit rate
Be wary of any vendor promising dramatic downtime reductions in year one; ask how they defined the baseline. Our automation business case template and AI vendor selection questions help here.
What contract risks should boards check?
OEM telematics. Supply contracts for CNC machines, packaging lines and robots may let the OEM collect operating telemetry for "service improvement", sometimes extending to benchmark resale. Audit them before adding third-party sensors.
Model ownership. If a vendor keeps model IP "for the platform", you cannot take your maintenance intelligence with you. Negotiate export rights up front.
Data location. Machine data carries little Privacy Act risk, but production rates, recipes and batch records are commercially sensitive. Government, Defence and critical infrastructure customers may require onshore data. The National AI Centre's Guidance for AI Adoption (October 2025) sets transparency and data governance expectations across sectors.
How do you choose an asset health AI approach?
- Sensor-led vendors lead with hardware plus basic anomaly detection. Good for condition monitoring.
- SCADA/MES platform extensions (AVEVA, Ignition, Rockwell, Siemens) are often lowest friction if you already run that stack.
- CMMS vendor AI modules help mature preventive into condition-based, usually weaker on raw telemetry.
- Specialist platforms carry deep model libraries for those willing to commit at platform level.
- Independent integrators combine your existing CMMS and SCADA with vendor-neutral ML, avoiding lock-in to one roadmap.
See our manufacturing AI use cases for the wider picture.
What does good look like at 18 months?
- one dashboard for health, OEE, energy and quality per asset
- AI-recommended work orders flowing into the CMMS through human approval
- technician time shifted from callouts to planned work
- energy savings traced to assets and feeding NGER and AASB S2 reporting
- an audit trail under ISO 55001, IEC 62443 and AS/NZS 4024
- a second wave of asset classes in pilot.
If the AI is not in the actual maintenance workflow and the real schedule still lives in a spreadsheet, it is theatre.
How Solve8 can help
Across our consultants' careers we have worked inside industrial environments at BHP, Rio Tinto and Schneider Electric, and we connect historians, SCADA and CMMS so asset health models feed work orders a planner approves. See our manufacturing and resources work and process automation service, or contact us to scope a pilot.
