Data Mining of Very Large Data Bases and Commercial Solutions to Predictive Maintenance

Rolls Royce Commercial Engines

Commercial engines are very complex units: their tolerances for performing efficiently and effectively are very small. One example is the combusters, located in the engine where Jet A1 fuel is injected in a vaporised sate. Once ignited and fully burned, this is naturally the hottest location in the gas turbine. The problem, however, with such high-temperature locations is the material science. The very high temperatures, along with rapid changes in temperature and gas flows, result in the combustor components being subjected to extreme wear, resulting in failure. Over the last 50 years, new applications for materials and coatings have led to improved on-wing performances, but the limitations of this science are very apparent, with the need to regularly inspect such hot areas, recording all visible defects. The engine manufacturers design the engine to perform with a given mass to pre-defined tolerances and service intervals. Therefore, if a new engine for an aircraft is too heavy (exceeding the airframe OEMs initial specification) or does not produce satisfactory on-wing performance (because it requires component changes), there will be a detrimental financial effect on the engine OEM..

Rolls Royce estimates that the total operational cost for an aircraft turbine engine is approximately 4% of total costs, represented in Figure 5.7. The high cost of engine operation has always been a great concern for the aviation business. Rolls Royce understood the need to follow engine performance trends for many years, and in the late 1980s onwards Rolls Royce worked very closely with Cathay Pacific Airways and American Airlines to better improve performance and reduce operating costs. In the UK Derby site, a new mathematical approach was led by Dr Michael J Provost (PhD December

FIGURE 5.7

Rolls Royce estimation of the engine maintenance costs versus total airlines operating costs. (r2pi.)

FIGURE 5.8

Rolls Royce correlation between component changes and sensor bias. (Rolls Royce.)

1994, Cranfield University), to utilise a new mathematical approach to predict engine performance. Provost recognised the difficulties with the infinite possibilities of variables with the engines, especially if components are removed and replaced (thus introducing new elements of change). However, instead of considering component performance as a linear representation, Provost considered multiple components as vectors. Additionally, a bias was incorporated to address such variables, as illustrated in Figure 5.8.

The total system is defined as an engine, having unknown component changes (as associated effects) plus the unknown sensor biases. This engine system will be operated, and performance data sensed/recorded. Within the engine performance data, there will be randomly generated noise (from sensors, recording media, etc.), and ultimately the final data will be a combination of all of the aforementioned items. While this is considered to have an infinite number of solutions to a very complex problem, Provost's team used a mathematical solution from the 1960s known as a Kalman filter, to provide the best 'estimate' of the state of the total engine system. A representation of this predictive system is illustrated in Figure 5.9. Blocks 'B' and 'C' are compared to the values of the ideal performance, and this variance between performance values gives a measurable indication of the recorded changes.

The use of this mathematically modelled data, compared with, for example, performance data captured with a test engine in flight-laboratory conditions, has been pivotal in permitting accurate forecasting and predictive maintenance schedules. Rolls Royce has successfully used Provost's application of the Kalman filter to gain world-wide patent exclusivity for this methodology, because the value of estimating equipment longevity is so financially valuable. Initially, Rolls Royce partnered with a Canadian organisation in a joint venture to form DS & S, which was the predictive maintenance holding company.

FIGURE 5.9

Flow diagram illustrating the application of the method to estimate levels and trends in Rolls Royce predictive maintenance tool. (Rolls Royce.)

This recognition of the value of reliability was further demonstrated in 2006, when Rolls Royce purchased all the shares of the joint venture and became the sole owner of DS & S.

Rolls Royce has recognised the importance of predictive maintenance capabilities, as although it values initial equipment sales, the income generation from service support (parts, maintenance management) has much greater financial longevity, which can be predicted by the industry up to 50 years from point of manufacture. This ultra-long-term ambition is only made possible by combining the use of embedded sensors within the engine and the ability of to retrieve live 'on-wing' data communications (via ACARS with satellite communication), so that real performance can be compared with the VLDB to optimise and achieve maximum performance from the on-wing asset.

Airbus, Palantir and Skywise

As previously mentioned, it is noted that all the major OEMs have commercial offerings for predictive maintenance. Airbus has partnered with a US-based software analytics organisation, Palantir, to capture and analyse aircraft performance data to form a VLDB that can be analysed. Partly founded by Paypal co-founder Peter Thiel, the early function of Palantir's products included counter-terrorism strategies designed to disrupt the financial advantages of electronic commerce. Palantir currently has two distinct products based around data processing algorithms. The first is known as Gotham, and is aimed at Government agencies, including the intelligence services, to map, network and analyse VLDBs. The second major product is Foundry, which includes their proprietary algorithm. This is the analytical technique that the 'Skywise' product is based upon.

The Skywise analytical product is marketed as a confidential tool that allows the operator to collate and analyse performance data (Figure 5.10). While the inner workings of the Foundry algorithm that powers Skywise are not published, the service states that the VLDB is confidential and only visible to the operator and the manufacturer (Airbus). This assurance implies that competing airlines that also use the Skywise product will not be able to view the other operator's data. While Skywise does not explain how such predictions or savings are calculated, the Skywise website has numerous video testimonials from prestige carriers, confirming the success of the service and implying the financial savings that have been made by the carrier.

FIGURE 5.10

Airbus and Palantir's Skywise data analysis indicating all the actors.

Skywise relies on vast quantities of data being captured, and Palantir suggests that this collection of data can be measured in pentabytes. The system records operational data values from up to 20,000 different sensors per individual aircraft, with a recording frequency of between 20 and 100 values per second. Each flight is expected to generate approximately 1,000,000 data values per sector, which for a European operator is approximately 1 ТВ per sector. Operationally, because of the vast data sizes acquired and stored within each aircraft, the operator still requires an engineer to visit each individual aircraft with a laptop to download the stored data. Once in the office, the recovered data can be uploaded to the Skywise server for analysis.

Skywise claims (at the time of publishing) that over 100 airlines are contracted to the sendee, with 9,000 + aircraft being monitored and a VLDB in excess of 10+ PB.

In addition to the predictive maintenance function Skywise performs, the Skywise service is also able to compare previous operational data to allow for accurate forecasting of future operations, including mass and balance, fuel planning, and revenue management. For partner component OEMs, the service also allows for an accurate Root Cause Analysis of in-service components that would previously only be seen during the removal and return for overhaul to the OEM. Now the respective OEM can review how their components are performing/operating on-wing.

While Skywise states that the operator can choose what data to share with Airbus, and what to withhold, it is also helpful to note that Palantir participated in the British Parliament Inquiry in 2018, as Palantir had 'operational connections' including meetings with Cambridge Analytica (CA) with confirmation in the inquiry that employees of Palantir and CA had been sharing offices with other third parties. This admission by CA and others to the parliamentary inquiry highlights how extremely large data sets that are collected can be shared, sold or accessed by third parties, without the knowledge or approval of the end user.

Conclusions

The concepts of aircraft maintenance have been explained, including the differences between scheduled and unscheduled maintenance, in addition to ongoing maintenance checks. The use of aircraft performance data is a very important and valuable tool for the Airframe OEMs, component OEMs and lastly the airlines themselves. The origins of data collection lay in logging entries into paper-based engineering Technical Logs that were completed before and after each flight. However, paper-based systems required the data to be transposed into a very large database, and these practical delays resulted in faults in monitored items going unnoticed and working beyond their intended operation. Later, a new source of flight performance data was brought in with the DFDR, and new devices (Quick Access Recorders and Cartridges) were introduced to transfer the information from the aircraft to the computer database.

With modern aircraft in the 1990s, ground stations could request the airborne communication system (ACARS) to transmit a small number of aircraft system performances in-flight. This system performance was fully autonomous, and did not require assistance from flight crews. Another development was replacing QARs with the introduction of portable laptops with cables and modules, with the ground engineer regularly downloading the information onto the computer's hard drive.

After the events of 9/11, it became more practical for airlines to consider leasing operational time from engine manufacturers in a 'power by the hour' agreement. However, the engine manufacturers needed to remotely monitor the ongoing performances of their assets, so, data communication modules were added to the engines' FADEC computers. Rolls Royce developed their own predictive reliability software, and patented it - as did the other engine manufacturers.

Lastly, the collection of big data by the Airbus Group in conjunction with Palantir has demonstrated that significant levels of financial savings can be made. Some of this data is transmitted by ACARS. However, the vast majority of the system performances are still being transferred once a week using a laptop, with the data uploaded to a server at Airbus's and Palantir's facilities.

The next chapter will explain the background of human error, or 'Human Factors'. It will explore important events that are now used as case studies to demonstrate the need to improve safety management, and why this is a mandatory subject for airline staff.

6