Context
Vacuum pumps are crucial equipment in many industrial processes. However, these pumps are subject to wear and tear, and can break down over time, leading to costly production interruptions. Faced with this challenge, the customer, the manufacturer of these vacuum pumps, performs preventive maintenance based on its knowledge to avoid unwanted stoppages. Today, it wants to turn to predictive maintenance to improve customer service and maximize pump life.
The project focused on analyzing data from 10 vacuum pumps, with the main aim of identifying the warning signs of end-of-life.
Need
The customer wanted to develop a solution capable of detecting pump failures in advance, so as to react in time to avoid unforeseen production stoppages. This solution would also enable better control of field interventions and maximize pump service life.
Solution
To meet this need, DiagFit has created an anticipation model based on historical pump data. This model, capable of detecting weak signals present in the equipment’s time series, will enable the first signs of equipment ageing to be identified. Thanks to this model, it has been possible to anticipate certain types of breakdown to enable effective preventive maintenance.
Although the model developed does not yet detect all types of failure, it has demonstrated an excellent ability to identify end-of-life pumps.
What’s more, thanks to the indications provided in the alerts (sensor contributions, temporality, etc.), the customer can focus his interventions.
Results
The model enabled the customer to benefit from early warning detection from 2 months to 7 days before breakdown (50% probability at 2 months, 88% probability at 1 month and 100% probability at 7 days).
This anticipation capability has not only reduced the risk of production stoppages, but has also enabled interventions to be focused.
The solution brings considerable added value in terms of safety and operational efficiency.
Fault detection based on a vacuum pump database:
In this figure, the red line indicates the model’s detection threshold. One month before the breakdown, the model detected a drift, as shown by the blue curve representing the prediction score.
