Definition: Predictive analytics is a set of methods that use statistics and machine learning to calculate the probability of future events from historical and current data. In mechanical engineering, the most important application is predictive maintenance: measured values from sensors are used to estimate when a component will probably fail, so that you can plan the maintenance in advance.

In B2B mechanical engineering, predictive analytics is used for: predictive maintenance of drives, bearings and pumps, spare parts planning, service contracts with availability guarantees, quality predictions in production, and the visualization of machine conditions on the digital twin.

Advantage: Maintenance takes place when the condition of the machine requires it, not on a fixed schedule and not only after a failure.

English: Predictive Analytics, Predictive Maintenance
German: Prädiktive Analytik, vorausschauende Wartung

How It Works

From measured value to prediction

A typical predictive maintenance system works in five steps:

  1. Collect data: Sensors measure vibration, temperature, current, pressure or speed. In addition, there are control data (run time, load cycles, error codes) and the service history.
  2. Transfer data: IoT gateways and industrial protocols such as OPC UA or MQTT send the values to a data platform (edge or cloud).
  3. Extract features: Characteristic values are derived from the raw signals, for example the RMS value of the vibration, frequency spectra or temperature trends.
  4. Apply the model: A model compares the current condition with learned patterns and estimates the remaining useful life (RUL) or the probability of failure.
  5. Take action: When the prediction exceeds a threshold, the system creates a maintenance order, a spare parts order or a notice to the operator.

Comparison of maintenance strategies

Strategy Trigger Typical risk
Reactive Failure Unplanned downtime
Preventive Calendar, operating hours Parts are replaced too early
Condition-based Current measured value above the limit Short warning time
Predictive Prediction of the future condition Data quality and model quality

Model types

  • Physical models: Describe wear with known laws (for example bearing fatigue). Easy to explain, but much work to create.
  • Data-driven models: Regression, time series analysis or neural networks learn from operating data. They need many examples, also of failures.
  • Hybrid models: Combine both and are common in mechanical engineering, because real failure data is often rare.

Digital Twin and 3D Visualization

Why predictions need a spatial view

A prediction such as "Bearing 3 on axis 2: remaining useful life approx. 400 operating hours" is clear to data scientists. But a service technician or a customer must know where this bearing is, how to get access to it and which parts are affected by the replacement.

For this, the digital twin connects the 3D model of the machine with the live data and the prediction data. Each component in the model has an ID that links it to its sensor values and predictions.

Display types for each platform

  • Web: Condition dashboard with a 3D viewer in the browser, critical assemblies marked in color (green, yellow, red).
  • Mobile and tablet: The technician on site gets the condition of individual components.
  • Kiosk and control room: A large screen in the hall shows the condition of all systems as a 3D overview.
  • Augmented reality: Prediction values appear directly on the real machine, including the removal steps.
  • Virtual reality: Technicians practice maintenance tasks on the virtual machine before the spare part arrives.

One 3D model serves all output channels. A clean 3D pipeline is essential: it prepares the CAD data for real-time rendering and keeps the component IDs.

B2B Applications

Sales: availability as a sales argument

More and more manufacturers sell service contracts with availability guarantees, not only machines. Predictive analytics supplies the data for this. In the sales conversation, a 3D visualization shows which components are monitored and what a warning looks like. This makes an abstract service promise tangible.

Trade show: live data on the exhibit

A packaging machine on the booth supplies real sensor data. A screen or tablet next to it shows the 3D model with a condition display. Visitors see how a temperature trend changes the prediction, without having to reach into the machine.

Service: targeted visits instead of troubleshooting

Before the trip, the technician knows which component is affected and brings the correct spare part. On site, AR instructions guide the technician to the affected location. Remote experts see the same model and the same data.

Training: practice maintenance before it is necessary

New service technicians practice the replacement of a monitored component on the virtual model. The scenario starts with the prediction message and ends with the restart of the machine. This way, they practice the full process, not only the manual task.

Design review: use findings from the field

Prediction and failure data from the field show which assemblies wear frequently. In the design review, these hotspots are marked on the 3D model of the next machine generation, for example to improve access or the bearing design.

Challenges

  • Data quality: Missing timestamps, incorrectly calibrated sensors or gaps make models unusable.
  • Little failure data: Well-maintained machines rarely fail, so models have few examples of failures.
  • Integration: Predictions have an effect only when they flow into ERP, service and spare parts processes.
  • Data security: Machine data are trade secrets. Access rights and encryption between the customer and the manufacturer must be defined in the contract and in the technology.
  • Acceptance: Operators trust predictions only when they can understand them. A spatial view helps to show the reason.

Introduction in Practice

Start small, then scale

Successful projects rarely start with the full machine fleet. A step-by-step approach has proven effective:

  1. Select a critical component: A component whose failure causes expensive downtime and whose wear you can measure, for example spindle bearings or vacuum pumps.
  2. Build the data basis: Retrofit sensors or make existing control data available, and document maintenance events correctly.
  3. Run a pilot: Run predictions for some months in parallel with the current maintenance and compare them with the findings.
  4. Connect the processes: Integrate warnings into service planning, spare parts logistics and customer communication.
  5. Visualize and roll out: Prepare the results on the 3D model for service and customers, then transfer them to more machines.

Cost-effectiveness: The costs for sensors, the data platform and model maintenance are compared with the avoided downtime, fewer emergency visits and predictable spare parts stock. The result is very different for each machine and its failure costs. You should prove it with real numbers in the pilot project.

Frequently Asked Questions (FAQ)

What is the difference between predictive analytics and predictive maintenance?

Predictive analytics is the general term for data-based predictions in all areas, for example demand or quality. Predictive maintenance is its application to maintenance: it predicts the condition or the time of failure of a machine or a component.

Which data do you need to get started?

At a minimum, continuous measured values of the critical components (often vibration and temperature), operating data from the control system and a documented maintenance history. Without a history of failures and repairs, you can hardly validate the models.

Do you need a digital twin for predictive maintenance?

Not necessarily. Predictions also work as a pure time series analysis. But a digital twin with a 3D model makes the results clear to service, sales and customers, and links the data clearly to a component.

Can you retrofit existing machines?

Yes. Retrofit sensors for vibration and temperature are installed on the outside of motors, gearboxes or pumps and send data by radio or cable to a gateway. Often, you do not have to change the control system itself.

How reliable are the predictions?

This depends on the amount of data, the data quality and the type of failure. Gradual wear is usually easy to predict, sudden breaks or operating errors are almost impossible to predict. For this reason, always give predictions with a probability or a time window, and check them continuously against real findings.

Industry Standards and Technologies

  • ISO 17359 – General guidelines for condition monitoring and diagnostics of machines
  • ISO 13381-1 – Prognostics in condition monitoring and diagnostics
  • ISO 20816-1 – Measurement and evaluation of machine vibration (successor to ISO 10816)
  • OPC UA (OPC Foundation) – Vendor-independent data exchange in industry (IEC 62541)
  • Asset Administration Shell (IDTA) – Standard for the industrial digital twin, including a submodel for predictive maintenance
  • MQTT (OASIS) – Lightweight protocol for IoT data transfer

Data and connectivity: - IoT – Connected machines - Sensors – Measurement technology - Digital twin – Digital copy of a machine - Condition monitoring – Monitoring of the machine condition - Edge computing – Data processing at the machine - Machine learning – Models that learn

Visualization: - 3D model – Geometric basis - 3D pipeline – From CAD to real-time 3D - Real-time rendering – Display in real time - WebGL – 3D in the browser - Dashboards – Overview of key figures

AR/VR in service: - Augmented reality – Overlays on the machine - Virtual reality – Virtual training - Mixed reality – Combination of real and virtual content - Remote assistance – Support from a distance - AR maintenance instructions – Step-by-step guidance

Security and law: - Data security – Protection of machine data - Data privacy – Personal data - Data Act – EU rules for access to machine data

Maintenance: - Condition-based maintenance – Maintenance based on measured values - Remaining useful life – Remaining service life - Spare parts management – Parts planning - Service contracts – Availability guarantees