Within the burgeoning field of Advanced Analytics, there is an application that holds particular potential: Predictive Maintenance (PM). The promise of PM is that, by using predictive analytics, companies can probabilistically anticipate which equipment is in need of repair, new parts or otherwise at risk of failure. This helps managers shift their upkeep programs from being reactive and routine to being proactive and fixing problems before they arise.
Leveraging the ongoing huge investments in Internet of Things (IoT), companies are developing the capability to collect vast data sets of conditions from connected assets ranging from environmental variables like ambient temperature to process variables like pressure, vibrations and product flow. Tetra Pak, for example, invested in sensors to monitor every step in the manufacture of its packaging materials to fully capture all of the potential areas of production disruption. The amount of data being collected is staggering. Per Dmitry Smolin, Director of the Smart and Connected Factory Program noted: “We record, on a daily basis, one billion data points from all of our machines. For example, a laminator has 400 sensors which are constantly recording information.”
Combined with more direct data like production speeds, line stoppages and breakdowns, analytics and machine learning are then applied to the data to generate predictions of equipment failure, for example of an automotive manufacturing line. Rather than perform routine maintenance, over-maintaining or wait until the failure occurs, the predictions are used to anticipate issues and resolve them before they escalate.
The potential benefits of anticipating equipment maintenance or repair needs becomes apparent when one considers equipment difficult to access, such as an offshore wind turbine, or equipment with very high stakes of failure, like an oil pipeline or a gas turbine. Assets that can be repaired before they fail are safer for people and the environment, and the overall maintenance costs can be considerably lower by having less and shorter downtimes, lower technician time and fewer spare parts in stock. The opportunity for savings are staggering, with a recent report from McKinsey estimating that by 2025 the annual benefits of predictive maintenance could be as much as $US 630 billion by increasing asset availability by up to 18% and reducing maintenance costs by as much as 25%.
The predictive maintenance sweet spot
PM seems like the perfect real-world opportunity for Industry 4.0. The benefits are visible and measurable, and it rests on a foundation of some of the most prominent Industry 4.0 technologies that otherwise may seem nebulous, like machine learning and Internet of Things. But closer consideration leads to the realization that there is a set of circumstances that must come together for PM to be effective; a ‘sweet spot’ of conditions that will enable PM to be a true game-changer. PM will be either difficult to implement or will not fully harvest benefits should one of the key elements be missing.
The first element of the sweet spot is that the necessary data be collectible and connectable. Data sources may include IoT connected sensors, soft data like repair logs and local root cause analysis trees. In almost any sort of data collection there is a challenge of standardization of data (some may be in local units, handwritten or in a different language!). Data availability is far more likely when the organization already has a culture of discipline in lean or kaizen to push for continuous improvement. From there, cloud capability and data lakes are needed to store the data, so the asset must have opportunities for connectivity. Oil pipelines, for example, might be a challenge and require more investment for connectivity.
As with other machine learning applications, PM will have a difficult time developing reliable, actionable predictions if there are too many variables or too many potential root causes of failures. The system will be too complex to develop algorithms economically and each discrete failure cause will lack a sufficient training dataset. Yet if there are very few variables or root causes, the asset may not warrant inclusion in a PM program, since failure predictions could likely be done by traditional statistical methods without turning to more expensive machine learning. To be a candidate for PM, then, the asset must have a moderate quantity of impacting variables and failure root causes.
The final element of the sweet spot is that the economic stakes must be sufficiently high. Designing and implementing a PM program will require mapping out a strategy, investment in sensors and connectivity, recruiting and training to develop expertise in machine learning and managing a transformation initiative. The benefits of PM must warrant the investment, be they in the form of preventable equipment down time, increased efficiency, lower labor and diminished safety or environmental risk. For this to be the case, the targeted asset must either be highly strategic, such as a critical production line, or there must be multiple, very similar assets that can benefit as a whole from a single PM program, such as turbines in a windfarm.