Scheduled vs Predictive maintenance
Reliability of the products is a key success factor for every company dealing with high volumes of products distributed over many different sites.
Scheduled maintenance is widely used to ensure the correct operation of the equipment and to avoid unexpected breakdowns, despite its disadvantages: it is a labor-intensive practice involving high costs and possible shortage of human experts, it is ineffective in detecting problems that develop between technician’s visits.
Predictive maintenance techniques can help to characterize the conditions of the equipment during every-day operations, giving room to the possibility to tell in advance when to intervene and what repair or maintenance intervention should be performed.
The main goal of predictive maintenance is to enable pro-active scheduling of intervention, thereby preventing equipment failures, while achieving cost-effectiveness of the maintenance. Data coming from all the monitored equipments / component / systems are concentrated in a central data hub where the predictive system is operating. Due to the variety of events it can be impossible to distinguish every single one, with the consequence of poor predictive performance.
Given such conditions we approach the problem as a multi-instance learning where, instead of using independent labeled instances as in standard predictive analysis, the model receives a set of bags containing multiple instances (segments of data) labeled as positive or negative:
- positive: it contains at least a failure (event, in general)
- negative: in the absence of events.
Data coming from event logs or from sensors can be integrated with service data (i.e., maintenance data) giving room both for improving the predictive power and for monitoring the action taken in response to signals issued by the predictive system.
Data key features to be defined are the following:
• Predictive Interval: a preset time window before the event (in the case of supervised learning as described afterwards). Usually it is bounded by the minimum anticipation required for an effective intervention.
• Wash-out Interval: a time window just before the failure that is not to be used in the predictive system.
• Refreshing time: depends on the data flow, on the goal of the system and on computation considerations. It is the rate at which data pattern are updated, real-time streaming data is possible, but not always necessary.
We have two ways to build a predictive maintenance system:
It requires to have events of failure and potential predictors obtained either from logs of the application software (as is often the case in medical equipments), or from sensor measurements.
It consists in detecting behaviors not congruent with those seen in conditions comparable with the current one. Without a set of events, models will not learn from the patterns preceding the events in a suitable defined time window. In this condition we can discover and point out significant changes in the data (e.g., in the time series) implying one or more variables. Often a sophisticated analysis is required to discover non-trivial differences in time series pointing to novelty discovery. The next step is to classify such novelty either as event-related or not. This process generates predictive examples, necessary to build and apply the approach (1) in a progressive, evolutionary way.
The two approaches are both necessary, the first to predict the occurrence of already seen events, the second to highlight anomalies perhaps related to so far unseen events.
Some specific circumastances where this approach is applicable:
- Car fleets endowed with remote sensing device, integrated with repair and maintenance data
- Hot drinks vending machines
- Medical equipments such as body scanners, Computerized Axial Tomography, Magnetic Resonance Imaging Apparatouses, Eco scanners, Positron Emission Tomography.
- Major component of a network like water or power distribution networks as well as an optical fiber network.