Predictive Maintenance

Machine maintenance is vital for many organisations, particularly in aerospace. Airlines spend a substantial share of their total operating costs on maintenance, repair and overhaul (MRO). According to the International Air Transport Association[1] (IATA), global aviation is expected to double over the next twenty years. MRO costs will rise along with it. Airlines face a significant challenge in keeping those costs down while improving operational efficiency. Predictive maintenance makes it possible to reduce MRO costs and improve operational efficiency using sensor technology and AI. It helps organisations make the most of equipment’s remaining service life by carrying out maintenance and overhauls at the right time and preventing costly repairs.

Introduction

For years, the prospect of lower MRO costs and greater profitability through intelligent maintenance techniques has been music to many organisations’ ears. The idea behind predictive maintenance is fairly simple: use sensor data and AI to maintain machines at the right time, reducing MRO costs and unplanned downtime. Until recently, the algorithms needed for predictive maintenance had not been developed, were not sufficiently advanced or were too costly to use.

In recent years, however, computing power has increased considerably, advances in AI have come in rapid succession, and sensors have become more accurate and less expensive. This lowers the initially high investment barrier and makes predictive maintenance more accessible and attractive to a growing number of organisations.

The energy, transport, manufacturing and information technology sectors are leading the way in predictive maintenance. This is hardly surprising, especially in aviation. According to an annual report by IATA’s Maintenance Cost Technical Group, airlines worldwide spent a combined 69 billion dollars on MRO in 2018[2], accounting for approximately 9% of their operating costs. Understandably, airlines are trying to reduce MRO costs with state-of-the-art predictive maintenance algorithms. The Dutch aviation sector is actively working on this: KLM’s most recent annual report[3] states that predictive maintenance delivered the largest reduction in its MRO costs.

In manufacturing, MRO costs are even higher[4] and, depending on the industry, can account for between 15-60% of operating costs. It is therefore crucial for these organisations to optimise MRO.

Maintenance approaches

There are different approaches to inspecting and maintaining equipment. These are generally categorised as reactive, preventive and predictive. The main difference lies in how they use the equipment’s service life, a typical pattern of which is shown in the figure below.

The earliest maintenance activities were reactive: equipment was repaired only when a fault occurred. The drawback was the high costs and risks. You could not know when a machine would fail or how much damage it would cause. If a machine was not due to be used for the next few days (for example, over the weekend) and could be repaired in time, the financial impact would be limited. But if it was a production machine in a car factory producing one car every minute, or an engine on an aircraft flying near the top of the troposphere, the financial and/or societal consequences could be considerable. Today, much maintenance is preventive, with inspections and maintenance carried out at regular intervals. Preventive maintenance helps you avoid unexpected failures that bring high costs and risks. Yet this approach is not optimal from a cost perspective either: parts and machines are inspected or maintained even when they have not reached the end of their service life. Their remaining useful life is therefore not fully used, resulting in unnecessary costs. By inspecting and maintaining parts and equipment at the right time, you can avoid unnecessary costs and risks. Predictive maintenance makes this possible by aligning maintenance with the actual condition of parts and machines.

What is predictive maintenance?

Predictive maintenance is an approach that aims to detect machine failure mechanisms at the right time and prevent further degradation through inspections and maintenance.

Intelligent algorithms detect failure mechanisms by predicting the current condition or remaining useful life of machines based on sensor data. The sensors measure vibrations, temperature, sound, lubricant composition, machine performance and many other operating characteristics. These algorithms use complex mathematical formulas whose parameters continuously adapt to new sensor data and the accuracy of their predictions. As a result, their predictions about the actual condition of machines can be highly accurate, provided enough high-quality data is available.

There are two ways to predict maintenance needs. The first is to calculate the probability that components or machines will fail within a given period. The second is to predict when components will fail. The first provides information about the current condition of components and machines, and therefore indirectly about their remaining useful life. The second estimates their remaining useful life directly.

Applying predictive maintenance requires a great deal of high-quality data, and machines need to be fitted with sensors. The initial investment is therefore relatively high. In most cases, however, organisations already have extensive historical data that they use to manage and control their processes. This data can help assess the potential of predictive maintenance for specific applications. While the initial investment may be relatively high, the benefits are greater. A study[5] by the US Department of Energy shows that an effective predictive maintenance programme can reduce maintenance costs by 30%, cut downtime by 45% and reduce failures by as much as 75%.

The advantages and disadvantages.

By now, it is clear that predictive maintenance offers many benefits, particularly in cost savings and risk management. The benefits of using intelligent algorithms to predict maintenance needs can be summarised as follows:

• Longer operational life for components and machines
• Lower costs for energy, labour, maintenance and parts
• Less unexpected (catastrophic) downtime for components and machines
• Better service and product quality
• Greater safety for employees and the environment
• Less maintenance equipment needed

Predictive maintenance therefore leads to greater operational efficiency through lower costs, higher quality, fewer risks and more sustainable practices. There are also drawbacks, however, which can be summarised as:

• Management is often unaware of the potential savings
• High initial investment
• Higher training costs for operating staff

Example of a predictive model.

This section demonstrates the accuracy of predictive maintenance models using a scientific paper[6] published at the International Conference on Circuits and Systems.

The paper applies predictive maintenance algorithms to sensor data to predict the number of flights remaining before aircraft engines fail, known as Remaining-Useful-Lifetime (RUL). It uses NASA’s Turbofan Degradation Simulation dataset. This four-part dataset consists of time series covering three operating settings and 21 sensor measurements from a fleet of aircraft engines over a period of time. The engines are of the same type but vary in how they were manufactured and in the amount of wear they have sustained. These differences are not recorded in the dataset.

At the start of the time series, the aircraft engines function as they should, but over time, wear causes the first faults to develop. These failure mechanisms progress until the engines eventually fail. The time series used to train the algorithms end some time before the engines fail. The algorithms are then trained to predict as accurately as possible how much time remains until failure in each time series.

The article compares several algorithms, but Random Forest performs best. Figure 1 shows the Random Forest algorithm’s predictions for different aircraft engines in the first part of the dataset. Accurate predictions lie close to the diagonal, where the predicted RUL equals the actual RUL.

Figure 1: Random Forest algorithm for predicting remaining useful life.

The Random Forest algorithm’s mean squared error on the first part of the dataset is approximately 25. This means its predictions of how many flights an aircraft engine can make before it fails differ from the actual number by an average of 5 flights. Figure 1 also shows that the algorithm is more accurate when failure is imminent. This is as expected: predicting the near future is easier than predicting the distant future. With predictions this accurate, you can intervene in time and extend the operational life of your machines. The predicted number of remaining flights can also be higher than the actual number. In this unwanted scenario, an aircraft engine fails unexpectedly, bringing significant risks and costs. It is therefore important that human expertise in MRO and artificial intelligence complement rather than replace each other. For equipment whose operation directly affects the wellbeing of people and nature, preventive maintenance and predictive maintenance need to go hand in hand.

Data Science Lab and predictive maintenance.

Does your organisation rely heavily on machines? Are you looking for a more sustainable approach to maintenance? Do you want to reduce MRO costs and increase operational efficiency? We’d be happy to welcome you to our Lab to explore together which techniques your organisation needs and which are right for you.

References

For more information about the possibilities of predictive maintenance, see the sources below.

• Mobley, R. Keith. An introduction to predictive maintenance. Elsevier, 2002.
• Selcuk, Sule. "Predictive maintenance, its implementation and latest trends." Proceedings of the Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture9 (2017): 1670-1679.

1. https://www.iata.org/en/pressroom/pr/2018-10-24-02/
2. Airline Maintenance Cost Executive Commentary Edition 2019 - https://www.iata.org/contentassets/bf8ca67c8bcd4358b3d004b0d6d0916f/mctg-fy2018-report-public.pdf
3. KLM Annual Report 2019 - https://www.klm.com/travel/nl_nl/images/KLM-Jaarverslag-2019_tcm541-1063986.pdf
4. Mobley, R. Keith. An introduction to predictive maintenance. Elsevier, 2002.
5. “Operations & Maintenance Best Practices: A Guide to Achieving Operational Effi‐ciency,” US Department of Energy Federal Energy Management Program, August 2010
6. Mathew, Vimala, et al. "Prediction of Remaining Useful Lifetime (RUL) of turbofan engine using machine learning." 2017 IEEE International Conference on Circuits and Systems (ICCS). IEEE, 2017.

020-3031414