Unplanned machine downtime costs the German manufacturing industry more than €50 billion per year according to Statista. Predictive maintenance — condition-based maintenance driven by sensor data and machine learning — promises to reduce these costs dramatically. AWS provides a complete platform with IoT SiteWise, Amazon Lookout for Equipment and SageMaker that requires no in-house data science expertise. This article describes the reference architecture, implementation path and AWS MAP funding opportunity for mid-sized manufacturers.
The Problem: Reactive Maintenance Is Too Expensive
Most manufacturing companies still operate reactively: a machine fails, the maintenance team is called, spare parts are ordered. Or they rely on fixed maintenance intervals — replacement every six months regardless of whether the component has 90% life remaining or is already on the verge of failure.
Both approaches are costly. Reactive maintenance leads to unplanned stoppages that can cost anywhere from €5,000 to €50,000 per hour depending on the production line. Preventive maintenance replaces components too early and ties up maintenance resources unnecessarily.
Predictive maintenance resolves this dilemma by continuously monitoring the actual condition of each machine and predicting failures before they occur.
Maintenance Strategies Compared
| Strategy | Trigger | Advantage | Disadvantage | Typical Cost |
|---|---|---|---|---|
| Reactive (Run-to-Failure) | Machine breakdown | No planning effort | Unplanned downtime, secondary damage, high emergency costs | High (emergency) |
| Preventive (Time-Based) | Fixed schedule | Predictable, no surprises | Early part replacement, unnecessary production interruption | Medium |
| Condition-Based (CBM) | Manual inspection | Condition-dependent | Labour-intensive, limited sensor coverage | Medium-high |
| Predictive | ML forecast on real-time sensor data | Optimal maintenance timing, minimal downtime | Initial data infrastructure required | Low (long-term) |
The AWS Platform for Predictive Maintenance
- AWS IoT SiteWise
- Collects and structures industrial time-series data from OPC UA sources. Automatically calculates metrics like vibration RMS, average temperature and pressure variance. Stores data in an integrated time-series data store consumed directly by Lookout for Equipment.
- Amazon Lookout for Equipment
- Fully managed ML service purpose-built for industrial anomaly detection. Lookout trains without data science effort: upload sensor data, describe the asset model, train the model. Output: real-time anomaly scores per asset with explainability (which sensor contributed most).
- Amazon SageMaker
- For advanced use cases: training custom ML models on historical production data. SageMaker offers AutoML (Autopilot), notebook environments for data scientists and managed model deployment. Trained models can be deployed back to the edge via Greengrass.
- Amazon Kinesis Data Streams
- Real-time data streams for high-frequency sensor data (down to millisecond resolution). Kinesis buffers data and routes it in parallel to Lookout for Equipment, S3 and Lambda-based alerting functions.
- Amazon SNS + AWS Lambda
- Alerting stack: when Lookout for Equipment detects an anomaly, a Lambda function triggers an SNS alert via email, SMS or integration with the company's CMMS.
Reference Architecture: From Vibration to Work Order
- Sensor acquisition: Vibration, temperature, current and pressure sensors on drives, bearings and hydraulic components measure continuously at high sampling rates.
- Edge aggregation (AWS IoT Greengrass): Greengrass aggregates high-frequency data locally (averages, RMS, peak-to-peak), reduces data volume and forwards relevant features to IoT Core. Critical thresholds trigger local alerts even without a WAN connection.
- Data ingestion (AWS IoT Core → SiteWise): IoT Core receives aggregated time-series data and routes it via IoT Rules to AWS IoT SiteWise, which stores it in the factory asset model.
- Historical data storage (Amazon S3): SiteWise exports raw data hourly into an S3 bucket (data lake) as the training foundation for ML models.
- Anomaly detection (Amazon Lookout for Equipment): The trained Lookout model runs in real time on new sensor data, producing anomaly scores between 0 and 1 with explanations of which sensor values deviate most.
- Alerting (Lambda + SNS): When the anomaly score exceeds the configured threshold (typically 0.6–0.8), Lambda triggers an SNS alert containing the affected asset, anomaly score, contributing sensors and recommended maintenance action.
- CMMS integration: Lambda automatically creates a work order in the CMMS (SAP PM, IFS, Maximo). The maintenance planner prioritizes based on the anomaly score and assigns time and resources.
- Feedback loop: Maintenance findings are fed back into the system. This feedback improves the model at each retraining cycle.
AWS MAP: Up to 50% of Project Costs as AWS Credits
The AWS Migration Acceleration Program (MAP) is a comprehensive funding program for cloud migrations and modernizations. For predictive maintenance projects that are part of a broader cloud modernization, AWS provides credits that are applied directly against AWS service costs.
- Assess phase: Free assessment tooling (AWS Migration Evaluator) and access to AWS experts
- Mobilize phase: AWS credits for the pilot phase (typically USD 10,000–50,000 depending on project size)
- Migrate/Modernize phase: Up to 50% of qualifying migration costs as AWS credits
Storm Reply is a MAP-certified partner and manages the complete application process — from project qualification to credit redemption.
Key Performance Indicators for Predictive Maintenance
- Mean Time Between Failures (MTBF)
- Average time between two failures of an asset. Predictive maintenance should measurably increase MTBF — typically 20–40% in the first year after implementation.
- Mean Time To Repair (MTTR)
- Average time to resolve a failure. Decreases with predictive maintenance because maintenance teams arrive prepared (diagnosis known, spare parts available).
- Planned vs. Unplanned Maintenance Ratio
- Ratio of planned to unplanned maintenance interventions. Target: shift towards planned. An 80:20 ratio (planned:unplanned) is considered excellent; many companies start at 50:50 or worse.
- Cost of Maintenance per Unit of Output
- Maintenance costs relative to production volume. This metric shows whether predictive maintenance actually reduces costs or merely shifts expenditure.
Storm Reply: Predictive Maintenance from a Single Source
Storm Reply, as an AWS Premier Consulting Partner, combines manufacturing expertise with data science competency. Storm Reply's offer for predictive maintenance includes use-case assessment, pilot implementation in 8–12 weeks, MAP application support and full rollout and operations management.
Frequently Asked Questions
- What is the difference between preventive and predictive maintenance?
- Preventive maintenance follows a fixed schedule (e.g. every six months) regardless of actual asset condition. Predictive maintenance is based on real-time sensor data and ML models: maintenance actions are triggered exactly when a failure is imminent — not too early and not too late.
- How much sensor data do I need for predictive maintenance?
- Amazon Lookout for Equipment requires at least 14 days of historical sensor data for initial training, ideally 6–12 months. The more failure history in the dataset, the more precise the predictions. Even without failure labels, the model can detect anomalies.
- Does AWS MAP fund predictive maintenance projects?
- Yes. The AWS Migration Acceleration Program funds migration costs with up to 50% as AWS credits. Predictive maintenance built on AWS IoT and SageMaker qualifies as part of a broader cloud modernization. Storm Reply is a MAP-certified partner.
- How long does a predictive maintenance implementation take?
- A first pilot with one asset is achievable in 8–12 weeks: sensor connectivity via Greengrass, data in SiteWise, first Lookout model. Rollout to multiple assets typically takes 4–9 months.
Predictive Maintenance Assessment
Which of your assets are the best candidates for predictive maintenance? Let us find out together — including MAP funding opportunity.
Get in touch