Anticipatory maintenance prevents unplanned downtime

Unforeseen machine downtime is a production manager's nightmare, particularly in large facilities that operate around the clock. Machine downtime is almost inevitably associated with extremely high costs, causing missed order deadlines and the potential commercial and reputational issues associated with delivery delays.

In an iron ore mine, for example, the costs for downtime on a central conveyor belt can exceed five figures per hour - and any downtime is likely to last much longer than one hour, even extending into later shifts, with rapidly escalating costs. Even in other continuously producing facilities such as paper and sugar factories, rubber mixing plants or steelworks, a downtime cost of four figures per hour, and even higher, is not uncommon. Sudden unavailability of heavy load handling machinery that moves tree trunks into a paper production facility, for example, is likely to cause immediate and significant downtime.

The costs of unplanned downtime are generally exacerbated by urgency surcharges for repairs. Works managers are often prepared to pay almost any price without hesitation for replacement parts, as long as these are available as quickly as possible.

This economic risk has often been minimised by replacing critical machinery components during pre-scheduled maintenance breaks as a precautionary measure, and therefore much more frequently than necessary, meaning that components are discarded when they may have many months of life left in them.

To combat this issue, predictive maintenance based on condition monitoring has become a key tool in the drive to reduce the risk of expensive plant downtime without compromising reliability.

Typically, components such as hydraulic pumps are fitted with sensors, and then upper and lower limit values for the sensor signals are defined based on the operating instructions and from experience of past values. A warning is displayed if the values measured exceed or fall short of these thresholds.

However, these manually entered limit values are often not very meaningful in practice. This is because, particularly for dynamic operation, they lead to too many error warnings, meaning maintenance teams may not take the warnings very seriously - which is a major problem when a genuine machine issue does arise.

Anticipatory maintenance

However, the latest technologies are now moving this function from a predictive to a truly anticipatory approach to equipment maintenance.

The issue is to find a way of accounting for the complexity of making a realistic diagnosis of wear and tear on components across varied applications. For example, the load-bearing capacity and life expectancy of identical components can be completely different in an open-air mining operation at the Arctic Circle, for example, when compared with a foundry with very many more constant environmental conditions.

Statistically, an error is detected by chance with a probability of only 13 per cent, while an expert monitoring the system by traditional means has a 43 per cent chance of detecting it.

Any complex context where large volumes of data are generated by sensors is where 'machine learning' methods come into their own. Machine learning is based on a combination of improved algorithms, increased computer performance and the necessary domain knowledge to allow for the model-based monitoring of operating states rather than the previous value-based approach.

The latest systems harness the interplay of sensor systems, Cloud-based applications and machine learning methods in order to implement anticipatory maintenance measures - generating knowledge regarding the state of health of the plant from recorded sensor data and making far more reliable predictions. Customers then receive corresponding maintenance recommendations for their plants.

Learning the normal state

To do this, a machine learning algorithm determines a normal healthy state for each component from a variety of sensor signals during an initial learning phase. These can include criteria such as pressure, flow rate, vibration, temperature and oil quality, depending on the equipment being monitored.

The learning phase may only last a few days if the equipment being monitored carries out the operations under very similar conditions all the time. However, if the equipment is only seldom used, and in differing ways, or if different products are manufactured, this phase may be extended. This data is harnessed in the evaluation alongside drive expertise and prior knowledge of cause-and-effect relationships on wear and tear issues.

Following the learning phase, the system continuously defines a 'health index' for the components being monitored with its data-based model. If an individual measured value temporarily deviates from the tolerance range, this does not necessarily lead to an error warning, as wear and tear can rarely be detected from one signal. However, if the health index deteriorates because data from several sensors changes, even if this is within the pre-defined limits - because the behaviour of the machine has changed - then the system warns there is a problem. These systems have a typical error detection rate of 99 per cent.

The health index not only shows the state of the assembly currently being monitored, but also gradual changes to upstream and downstream mechanical or hydraulic systems. If movements take longer or require more power, this indicates wear and tear in the mechanical or hydraulic systems. The system can give corresponding instructions in its regular health index reports and help to create specific recommendations for action. Because the system combines all measurement data from connected plants, every dataset improves prediction accuracy.

In a typical example, an electric motor failed following the learning phase. The sensors did not report any critical development in the individual signals from the engine itself, but the health index had already shown four weeks previously that there was a problem. The algorithm learned when a warning must be issued from this case, so the same error does not lead to a plant downtime for a second time. Now, if a similar pattern occurs in this or another facility, a warning is issued to check the electric motor and replace it if necessary.

Critical components - for example, hydraulic systems and motors - can be fitted with a variety of different sensors that produce many gigabytes of unfiltered data each day, even for a relatively small system. Before this data is sent to the Cloud-based predictive maintenance system, it is pre-processed in a data acquisition unit at the plant. This reduces the data stream so even limited bandwidths are still sufficient for data transfer - which is a key requirement when, for example, facilities are located in very remote areas with only mobile or satellite telecommunications networks available.

The risk of plant downtime can never be eliminated fully, but it can be so significant that the costs for the system are frequently recouped the first time downtime is prevented.

Follow the link for more information about anticipatory maintenance.