The scope of this experiment is the optimisation of the maintenance plan of aircraft (structure, engines, equipments and parts) based on a large amount of data and correlation in order to enable:
- Aircraft operators as airlines to fly more with minimum operational and maintenance costs.
- Aircraft maintenance centres to be more efficient and decrease costs as they can optimize their resources allocation
- (Aircraft, engines, equipments and parts) manufacturers to improve their current and future products (design, reliability, etc.) by a better knowledge of their fleet while in service.
The state of the art highlights that forecasting tools exist on the market but the optimisation of scheduled and unscheduled maintenance tasks dates taking into account a large amount of data like several constraints set by the aircraft operators and/or the maintenance centres (as qualified resources availability based on the aircraft configuration) is done manually.
These tools compare the maintenance schedule set by manufacturers with available historical in-service data (as flight hours) in order to forecast next stops for aircraft maintenance. This forecast represents the current aircraft maintenance plan.
2MoRO Solutions existing tool uses an in-house frequent pattern algorithm to give a precise indication whether the next maintenance tasks should be performed ahead or after of maintenance schedule.
Description of the infrastructure
Current status: algorithm works on small data subset
Current limitations: data size, and analysis depth
Our data-mining application extracts straightforward information from a set of built data. It provides a set of temporal interval sequential patterns from a set of discrete temporal sequences in regards to extracting parameters (temporal, frequency, accuracy, interested data dimension) specified by user.
Data are loaded from a data base and pre-processed regards to user mining needs. It can be organised on several sequences database connected there between each other in regards to an arbitrary choice (made by user regards to mining needs).
Extraction algorithm is performed on each sequences database in order to identify straightforward patterns. The extraction process applies a growth pattern approach by performing a vertical extraction based on database projection. Recursively the algorithm first extract frequent patterns and then project data base.
The extraction implement specific (time) parameter constraint and apply a sliding window in order to consider different data merging combination. For instance, for aircraft lives data, let Vi refer to the flight i, Mj refer to a maintenance task j and S = <S1;S2> be a set of historic sequences where S1 =<(0;V1)(2;V2)(3;V3)(5;M1)> and S2 =<(0;V1)(2;V3)(3;V2)(6;M1)>. Let the minimal support constraint be equal to 2. Our method returns the sequence: < ([0;0]V1)([2;3]V2 V3)([5;6]M1) >. On this sequence intervals express an uncertainty for the exact moment when data occur.
According to the application field the accuracy of extracted patterns can be computed or performed by using linkage between sequences database. The provided result consist is set of frequent patterns associated to a frequent and accuracy rank.
The set of patterns can be used on a computation application in order to perform maintenance plan and predict maintenance task.
Currently, our solution is deployed on a local machine and works on small data. However, the parallelization of the algorithm is feasible (correctness theoretically proofed) and not coasty. the Figure infrastructure describe the flow and the interraction between the diffrent modules of our application:
Infrastructure : Interraction between the diffrent modules of our application and the description of the data flow
To provide scheduled maintenance tasks:
- Aircraft are delivered by manufacturers with a lot of documentations (aircraft maintenance manual (AMM), maintenance planning document (MPD), etc.). The format of the documentations is either paper-based or electronic versions.
- Maintenance tasks are described in the AMM and the scheduling in the MPD.
- During the aftermarket phase (operational phase) that can last more than30 years, manufacturers provide update of these documentations to the operators.
To perform inspections:
- Inspections of aircraft (airframe, engines, equipments and other systems) are performed by aircraft operators and/or Maintenance Repair and Overhaul centres based on their qualifications and the level of tasks to be performed.
To analyse damages:
- Damages are compared to the limits set by manufacturers in the aircraft documentations. Depending on the damages that are inside the limits, the outcomes may be doing nothing or repairing, replacing the parts. If damages are beyond the limits, manufacturers should be contacted to search for solutions to repair the parts. New solutions have to be certified by authorities (i.e. FAA, EASA, etc.) before being distributed by manufacturers and implemented by operators.
To determine tasks:
- Based on the analyses of damages and theoretical scheduled tasks, a list of tasks to be performed is determined by the operator and/or MRO centre.
To display the planning:
- At last, all determined tasks have to be planned. The aircraft availability has to be optimised, that is the reason why some tasks are proposed to be performed ahead of time.
Comparison between current and targeted scenarios
- Comparison of measures with limits set by manufacturers is almost done manually
In the targeted scenario, the creation of a global knowledge database will be enabled. Therefore, analysis of damages will be more reliable and the determined tasks to be performed more adequate. At last, it will be possible to make proposals to perform tasks beyond the limits set by manufacturers and then increase aircraft availability. Those proposals will be provided to operators and/or MRO centres with justifications. Hence, it will be easier for them to obtain approval by manufacturers and indirectly by authorities.
Detailed description of the demonstrator
Integrating our Solution on Cloud SME project will help us to perform and validate our solution on large scale data.
Facing a big amount of data will allow us to overcome the scalability of our algorithm, to test and perform the accuracy of its results in real situations including a larger set of parameters and configurations served by data.
Using Big Data will also allow us to adapt, perform and expand predictive functionality that uses our data-mining results.
This experiment should be able to use our algorithm in a parallelised way. We already know the mathematical formula that gives us how many processors we can parallelise our program depending on the quantity and the definition of data.
The capacity to define dynamically where the data is going to be stored is another new functionality, today we need to work this out manually for each client (choice, maintenance, backups...). The objective is to define it easily once for good.
Obtaining enough data logs and traces to understand how our algorithm works in order to anticipate the need for more resources. This would be a way to master our scalability.
Make our program work seamlessly on a fleet of 10 aircraft without any problems and ensuring the resources consumption can be kept inside a coherent business model.