Aircraft IT MRO – March/April 2014

Subscribe
Aircraft IT MRO – March/April 2014 Cover

Articles

Name Author
Case Study: Stepping up with MRO Software Rob Vogel, Senior Manager, National Airways Corporation View article
Early Adaptors S1000D David Boyer, VP of Aerospace Operations, and Tim Larson, Global Product Manager, Flatirons Solutions View article
Using Data to Improve MRO Schedule Management View article
Compliance – Stronger when tackled together Geoff Zuber, Director, Holocentric View article
Column How I See IT – 2014 Michael Wm. Denis View article

Using Data to Improve MRO Schedule Management

Author:

Subscribe

  This article appears in Issue 13: the March/April 2014 edition of the Aircraft IT MRO eJournal. For your own free subscription to the eJournal – click on ‘SUBSCRIBE FOR FREE’ for full details. 

Using Data to Improve MRO Schedule Management

Enhance aircraft fleet management and equipment performance by using data-mining methodology, says A. Ben Zakour, Research and Development engineer at 2MoRO Solutions

Aircraft manufacturers and operators have to differentiate their product from that of their competitors by offering customers ways to fly more often and/or further at lower cost. Towards this end, reducing maintenance costs can be a way to reduce the costs of aircraft use while ensuring a high level of service and safety.

Part of our research work at 2MoRO Solutions focuses on forecasting aircraft maintenance tasks and optimizing the planning of those tasks. This paper looks at the prototype prognostic tool that has grown out of this. The solution described takes advantage of heterogeneous raw data available around the aircraft lifecycle to better predict maintenance tasks and improve aircraft health monitoring.

Prediction of maintenance tasks helps to reduce diagnostic work as well as delays and costs caused by unscheduled maintenance events. Being able to forecast potential defaults, malfunctions, failures and errors is of considerable help to operators and maintenance professionals. It provides the right diagnosis and supplies a list of resources necessary to reduce maintenance time. Available data are made up of in-service data (derived from the aftermarket phase) and referential data (created mostly during the design phase). On the prototype presented in this paper, both kinds of information are pre-selected, analyzed and pre-processed in order to apply data-mining tasks. The mining task provides the discretional or predictive goal in accordance with user needs.

The prognostic tool combines and aligns both types of data to highlight timing differences between the maintenance plan and the real-life maintenance application. This correlation provides the link between the referential and in-service data: a link that can assist manufacturers in monitoring aircraft configurations and help with supplying the right resources in the right place at minimal cost. The proposed correlation adapts maintenance plans in accordance with the operating environment. It is made by identifying specific frequent patterns called ‘critical behaviors’ or ‘typical behaviors’.

‘Critical behaviors’ are identified by performing a data-mining task that extracts frequent patterns. It uses several extraction constraints, including a sliding window, to extract straightforward patterns, satisfying users’ needs. The extracted patterns represent frequent ‘similar’ contextual uses and contextual behaviors associated with maintenance tasks, and specific inspection reports identifying default, malfunction, fault or failure.

We’ll deal with the components of this methodology as follows. The first section details the types of information on which the analysis is based and presents the way they are merged and pre-processed. The second section describes the data-mining task used to identify critical behaviors and explains how the resulting patterns are used to predict aircraft maintenance events and analyze equipment performance trends. Finally, we conclude by explaining the next improvement in and ideas for industrial applications of the proposed framework.

Data description and pre-processing

The types of information on which the analysis is based and presents the way they are merged and pre-processed.

Before applying a data-mining task, a pre-processing step is needed to adapt raw data and improve the mining (description and/or predictive) results. The pre-process step consists in analyzing, strengthening and organizing data.

In order to identify critical behaviors, our prognostic prototype uses in-service and referential data.

  • In-service data are characterized by the context in which the aircraft is used. It includes information about aircraft operations and MRO such as flights details, pilot reports, aircraft configuration… but also maintenance reports and inspection reports. It describes a set of flights by detailing the airports from which the aircraft took-off and landed, flight duration, percentage of aircraft loading and fuel consumption. It also includes maintenance tasks performed during the life of the aircraft: each such task is linked to the associated equipment’s part number (PN) and serial number, to the counter values of the equipment and to the whole counter values of aircraft.
  • Referential data includes OEM specifications such as the maintenance planning document (MPD) and aircraft maintenance manual (AMM). The AMM describes the frequency of and conditions to apply precautionary maintenance tasks during the aircraft’s life. These tasks represent a hierarchical tree describing the different parts of the aircraft and this structure is available through the manufacturer’s technical data book, called Illustrated Parts Catalogue (IPC), also produced by manufacturers and compliant with ATA Spec standard.

Figure 1: In-service and referential data alignment 

The pre-process module of our framework aligns and merges both kinds of available data according to time and to the aircraft’s operations (number of flight hours, take-offs and landings, etc.). The connection with heterogeneous data is performed chronologically and describes events that occurred during the equipment/aircraft life cycle. Figure 1 shows an example of merging in-service and referential data. The chronology aligns aircraft/equipment status, mission conditions, expert reports and maintenance tasks. Scheduled tasks are represented twice: when they were planned (represented by green triangles) and when they were actually performed (represented by red triangles). 

In order to identify critical behaviors, users of the framework have to determine ‘critical points’ (failure, default, malfunction, maintenance task…). Critical points are reference flags that users can apply to analyze trends and associated usage behaviors. For each corresponding critical point, aligned data is associated to a usage sequence. A usage sequence is a succession of life cycle events ending at a specific critical point. Sequences ending with the same critical event are grouped in order to identify typical usage/ condition behaviors contributing to the corresponding ‘critical point’.

According to the representation described above (referential and in-service data), the framework builds a pattern of typical wear behavior for aircraft /aircraft parts/ equipment. In practice, for aircraft maintenance, repair and overhaul (MRO) purposes, each maintenance action is defined for a specific default/ malfunction/ fault/ failure caused on a particular equipment identifier (PN). According to this principal, we can infer that possible correlation between the type of use of the equipment and failure/ wear models can be considered as a correlation between type of usage for equipment and maintenance task associated to the failure/wear. Analyzing and comparing critical behaviors leading to several ‘critical points’ allows us to identify some usage or trend influences on other equipment or groups of equipment.

The following section describes the data mining task that identifies critical behaviors and allows the prediction of maintenance tasks.

Framework Mining Task

The data-mining task used to identify critical behaviors and how the resulting patterns are used to predict aircraft maintenance events and analyze equipment performance trends.

Current thinking informs us that there are plenty of forecasting tools on the market. However, the optimization of scheduled and unscheduled maintenance tasks dates taking into account several constraints set by the aircraft operators and/or the maintenance centres (as qualified resources availability based on the aircraft configuration) is done manually.

We propose a data mining tool that compares the maintenance schedule set by manufacturers with available and actual historical in-service data (such as flight hours) in order to forecast future dates for aircraft maintenance. This forecast represents the current aircraft maintenance plan. On this proposed framework, next steps forecasting identification is configured in the context of the aircraft/equipment life cycle. They are keyed out from the sequences of databases described above.

Remember that sequences depict combined aircraft lifecycle data associated to a maintenance task application or to a specific critical point. We apply a descriptive and predictive maintenance task that creates a case-based model that identifies typical critical behaviors. This model relies upon data to identify the most frequent (and then typical) sequence of usage and environment conditions leading to the specific critical point.

Extracting frequent behaviors is interesting and very useful when building wear models for each critical point sequence. If a group of behaviors occurs several times on several aircraft (of the same kind) before the execution of maintenance task(s), a logical correlation between corresponding equipment wear and those behaviors can be established. Those models are built by extracting frequent historic patterns from in-service data. Pattern extraction is used in several domains to identify several kinds of machine behaviors as in ‘Rabatel 09’ for trains and in ‘Hirate 06’ for astronomic robots.

Our data-mining application extracts straightforward sequences from a set of built data. It provides a set of temporal interval sequential patterns from a set of discrete temporal sequences. The extraction process takes several parameters (temporal, frequency, accuracy, interested data dimension) for adapting the provided patterns to users’ specifications. It also applies a growth pattern approach by performing a vertical extraction based on database projection. Recursively the algorithm first extracts frequent patterns and then projects sequences data base.

The extraction implements specific parameter (time) constraints and applies a sliding window in order to consider different data merging combinations.

For instance, for aircraft operational data…

… let Vi refer to the flight i…

Mj refer to a maintenance task (a critical point) 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 a frequency constraint be equal to 2.

Our method returns the sequence: < ([0;0]V1)([2;3]V2 V3)([5;6]M1) >.

These sequence intervals express an uncertainty for the exact moment when data occur. The extracted sequences mean that : “If flight V1 occurs, followed by both flights V2 and V3 in any order but in a time interval [2;3] after V1 then, maintenance task M1 is performed in a time lying in the interval [5;6] after V1”.

Such a pattern allows us to group V2 and V3 in the same relevant behavior. As far as we know, there is no pattern extraction algorithm that provides such structured patterns.

The extracted result consists of temporal interval sequence patterns representing frequent behaviors associated to a specific critical point or to a maintenance task. Interval timestamps express an uncertainty as to the exact moment when specific events/observations appeared relative to the whole patterns events/observations. The degree of uncertainty is managed by the size of the sliding window which is fixed by users. Each extracted pattern is associated with a frequency and to an accuracy value respectively greater than a frequency and accuracy thresholds fixed by users. These extracted patterns help with anticipating maintenance tasks execution according to their usage history: mission conditions, flights, reports and aircraft configuration; to their usages conditions: flight configuration, aircraft configuration, weather conditions, etc.

In addition to frequency and accuracy extraction criteria other relevant measures can be used. We can combine the criteria for the life-service part of the pattern with those of patterns associated to other critical points. Such a correlation allows us to evaluate the accuracy of a prediction decision with regards to a specific part or to the whole aircraft wear model status

The aim of our method is to provide usage based models that can have a dual purpose: first a contextualized and precise management of aircraft maintenance tasks, second a fleshed out model of wear and technical events (fault/malfunction/ failure/ error) of aircraft equipment.

For aircraft fleet management, the aim of frequent sequence extraction is to understand and predict critical technical events and maintenance task execution. Typically, for a specific type of aircraft, when a flight plan is defined, verifying its match with an appropriate wear model allows us to have a good idea about the aircraft interruption scheduling. A matching process using the flight plan is applied to several patterns extracted on critical point/maintenance task behaviors to predict technical events/ maintenance tasks execution with an accurate time interval. In addition there are several other uses. For instance, figure 2 and figure 3 show Graphical User Interfaces that use the extracted results to visualize trends (figure 2) and the scheduling of the maintenance tasks (figure3).

Figure 2: BFly® (2MoRO software) web interface using extracted results to visualize data trends.

Figure 3: BFly® (2MoRO software) web interface using extracted results to visualize maintenance scheduling

The extraction model enables the selection of the kind of time parameter, which can either be real time or equipment life cycle counter with different stamp granularity level (e.g. for time stamp different time granularity can be considered – minutes for sensor reading values and hours for flights information). So the estimated maintenance management is more accurate than the theoretical maintenance manual. These estimations will enhance aircraft availability by accruing a good knowledge of future downtime and then reducing maintenance and exploitation costs.

Wear and technical events models are also a way to correlate in-service data and environmental conditions with the criticality of execution for maintenance tasks. This correlation provides relevant equipment behavior profiles for specific environmental conditions. Hence based on forecast aircraft orders, equipment profiles can be exploited by aircraft manufacturers during pre-design analysis to adapt maintenance plans with the type of flight missions and environmental conditions foreseen by airlines and aircraft operators.

Conclusion

The next improvement in and ideas for industrial applications of the proposed framework

We present in this paper a general description of a framework prototype. The prototype provides a data-mining service for contextualized maintenance forecast and contextualized equipment trend analyses.

The proposed solution uses heterogeneous (in-service and referential) aircraft data. It provides a straightforward added-value used to optimize the planning of maintenance tasks and to analyze equipment behavior trends. The services provided by 2MoRO solutions framework is an adaptable data-mining task where users can configure the mining parameters to comply with application needs and data configuration.

The proposed framework is an independent tool which can be integrated within a more general framework or platform services like BFly®, an architecture proposed by 2MoRO Solutions. It can also be integrated within a collaboration platform which distributes and secures data (e.g. SIMID[1] project to which several companies contribute, including 2MoRO Solutions Airbus, Dassault Aviation, Eurocopter)

The framework presented is central to the topical issue about reducing the costs of aircraft use by taking advantage of the growing amount of data available through the aircraft lifecycle. It can also be considerably improved by implementing several other pre-processing tasks that will take advantage of continuous or textual data.

Bibliography

[Rabatel 09] Rabatel, J., S. Bringay, and P. Poncelet (2009). So_mad : «Sensor mining for anomaly detection in railway data», in Proceedings of the 9th Industrial Conference on Advances in Data Mining. Applications and Theoretical Aspects, Berlin, Heidelberg, pp. 191–205. Springer Verlag.

[Hirate 06] Hirate, Y. And Yamana H. (2006). « Generalized sequential pattern mining with item intervals” Journal of Computers 1.


[1] http://www.simid-project.eu

 

Comments (0)

There are currently no comments about this article.

To post a comment, please login or subscribe.