Preventing a catastrophic system failure before it happens — what may sound like the plot line of a blockbuster action movie; is actually one of the most common and beneficial use cases in machine learning. Predictive maintenance (PdM) is the process of applying ML models to equipment degradation measurements to predict when the next failure will occur. PdM is a powerful tool that seeks to optimize the performance of other maintenance strategies. When applied to appropriate business scenarios it can not only answer crucial cost saving questions but also provide insights into the overall performance of assets.
Comparing Maintenance Strategies
Corrective (reactive) maintenance is the process by which parts are replaced after a failure occurs. Since corrective maintenance requires minimal startup cost and time investments, it is by default the most common strategy relied on by businesses. While corrective maintenance is straightforward to perform and ensures there is no waste in a component’s life, it can result in significant operation impediments. If several assets fail at once business downtime and unscheduled maintenance could result in lower production rates and higher cost requirements for labor. For businesses that operate at high production levels and adhere to strict schedules, corrective maintenance is a suboptimal strategy.
Preventive maintenance represents the next level in business strategies. This method operates on a preset maintenance or replacement schedule determined by the historical useful lifespan for the part. For instance, a car getting an oil change every six months is a preventive measure based not exactly on the performance of the oil but on the average known lifespan of oil. While this strategy avoids the unscheduled downtime and disastrous failures associated with corrective maintenance, it still suffers from part under-utilization and higher labor costs.
Predictive maintenance is the strategy that balances between corrective and preventive maintenance by relying on “just in time” replacement or maintenance of components. This approach replaces parts only when they are within a prediction window of failure. This prediction window can be based off several machine features including age, operational conditions, and telemetry or sensor reading data. Predictive maintenance maximizes the lifespan of components unlike preventive maintenance while minimizing unscheduled downtime and costs as compared to corrective maintenance.
Predictive Maintenance Criterion
While predictive maintenance can be extremely beneficial to business operations not all cases are applicable. There are several qualifying factors a use case must meet to ensure a successful PdM implementation.
First, the problem needs to have a target for prediction. In other words, it must fall into the category of supervised machine learning. Even with a target in mind there should also be a way to prevent the failures when predicted. Without a clear treatment plan, it is irrelevant whether a prediction is accurately made.
The problem also requires that the equipment has a sufficient record of operational history that reflects both positive outcomes and failures. Telemetry data, error reports, maintenance logs, and performance metrics are all examples of the types of features required to make accurate failure predictions.
Finally, subject matter experts (SME) in the field are crucial to the feature engineering process of predictive maintenance. SME’s guide data scientists in knowing what fail and lag windows to consider and what features are relevant to a machine’s possible failure.
Predictive Maintenance Benefits and Challenges
A successful PdM implementation can provide businesses with several insights and opportunities. Not only can a PdM model answer the question of whether a component will fail within a certain time window, but it can also reveal which features are most correlated with the failure as well as the remaining useful lifetime. Possessing this type of knowledge can lead to opportunities in schedule optimization. Schedule optimization can not only reduce downtime, labor, and tooling costs but it can also prevent major production delays.
However, as with any maintenance solution, PdM can also present challenges to businesses. If detailed machine monitoring is not already protocol the startup costs and training could be extensive. This includes not only equipment to monitor the machines, such as sensor readings, but also a method of efficiently storing and analyzing the possibly massive amounts of data. Technicians and managers must also possess training in the field of data ingestion and mapping as well as a willingness to implement new maintenance schedules. Developing a PdM model that supplies reliable predictions can also require a significant investment. Collaborating with SME’s to uncover the best feature engineering approach, tuning the model’s parameters, and testing various sets of prediction windows can all consume time and labor costs.
In conclusion, despite the possibility of a large overhead startup, predictive maintenance can provide businesses with a strategic advantage in cost, downtime, and labor minimization. Having the ability to balance maintenance between a complete failure and component under-utilization is crucial for not only schedule optimization but overall maximizing business operations. Plus, acquiring action hero status when a catastrophic failure is prevented is always a nice addition. If you’re interested in learning more, please contact us today!