Predictive maintenance using machine learning

Predictive maintenance determines the actual condition of the equipment and predicts when the company should perform maintenance.

This approach promises cost savings because the company only performs tasks when warranted.

Objectives.
Benefits.
Approach.
Conclusions.

Objectives
Benefits
Approach
Conclusions

Objectives

Predictive maintenance looks for anomalies, i.e., unexpected measurements that might indicate a problem – but are not yet so severe that they are a failure.

The challenge is to determine the condition of in-service equipment to predict when the company should perform maintenance and prevent unexpected failures.

Benefits

Predictive maintenance allows for managing problems that could arise in the present and preventing future unexpected events.

IMPROVE PLANNING

REDUCE REPAIR COSTS

REDUCE PRODUCTION LOSSES

REDUCE DOWNTIME

Approach

Predictive maintenance is to model equipment failures based on observations of past machine runs and failures.

Neural networks can model the correct operation of the equipment at a given condition and detect when this operation is an anomaly.

That allows early spot potential equipment failures and fix them before they happen.

Conclusions

Predictive maintenance saves companies money since they will have shorter downtime and less lost production, better planning of people and materials, and reduced repair costs.

Neural Designer uses machine learning to build predictive models that represent a broad range of variables associated with the failure of equipment.

Building a reliable predictive maintenance model

Begin with time-stamped operating conditions, sensor measurements, maintenance events and a clearly defined outcome. The observation window and prediction horizon must be fixed before training so that future information cannot leak into the model.

Validation and deployment

Use chronological validation when equipment behaviour changes over time. Evaluate alert lead time, missed failures and false alarms alongside the usual model metrics, because each has a different operational cost. After deployment, monitor input drift and prediction quality as equipment, processes and maintenance policies change.

Contents