Predictive modeling of non-routine maintenance workload in aircraft operations: a task card-level approach using real mro data and machine learning

Abstract

Aircraft maintenance planning has increasingly become a critical area of research due to the high operational costs and complexity associated with unexpected maintenance demands. Since aircraft maintenance and repair is a complex and human factor-intensive field of activity, unexpected failures are frequently encountered during maintenance. Non-routine damage findings encountered during aircraft maintenance and repair activities cause great difficulties in maintenance planning. The unpredictability of the encountered findings makes it difficult to prepare resources such as man/hour supply, materials, tools and components before maintenance. Although civil aviation maintenance and repair procedures are standardized in a certain systematic manner by civil aviation authorities and aircraft manufacturers, damages encountered during maintenance can only be accurately identified through inspection or testing. Therefore, data-driven prediction models have gained importance in improving operational efficiency and planning by predicting the workload of unplanned findings. Studies in the field of machine learning in aircraft maintenance and repair mostly focus on anomaly detection, component fault detection, component remaining useful life (RUL) prediction from sensor signal records on the aircraft. Some studies have focused on unplanned material requirements using maintenance historical data. Most of the studies in the literature focus on classification problems rather than workload prediction and use simplified and synthetic data sets without considering parameters such as hierarchical structures, flight history, time between two maintenance. In this study, a modeling framework is presented using real historical data of an aircraft maintenance and repair organization (MRO) to predict the total man/hour values of non-routine tasks encountered in scheduled maintenance cards using machine learning methods in order to fill the mentioned literature gap. In addition to time-based variables, variables such as maintenance type, aircraft type and age, content of maintenance cards (task description), cumulative flight hours and number of flights performed by the aircraft until the relevant maintenance card is applied are used in the dataset. However, the structural and statistical characteristics of the target value, such as its right-skewed distribution structure in the dataset, its sparsity in the data, its inclusion in the time-based supergroups (Maintenance Package) and its density at the zero value, necessitated more careful data cleaning and preparation processes before modeling and led to more appropriate strategies for estimation. In the data cleaning and preparation processes, variables such as low category maintenance card task type, aircraft type and maintenance type etc. were made ready for modeling with one-hot encoding, while variables such as high category maintenance card content and description, aircraft tail code etc. were made ready for modeling with label encoding method. In addition, new variables that were not included in the raw data set, such as the number of flight hours and the number of flights performed by the aircraft between the two maintenance, and the number of times the maintenance card was previously applied in the maintenance organization, were generated from the data set and used in the modeling. In order to prevent data leakage during the modeling, cross-validation method with GroupKFold was used for maintenance package (Work Order) information. For this reason, it was ensured that the operationally interdependent maintenance cards were only on one side of the train/test separation before modeling, since they were in the same maintenance package. Thus, each fold was created to include maintenance packages not previously seen by the model in the test data set. In the study, a two-stage modeling was adopted and the number of unplanned openings was first estimated with the Random Forest algorithm. The output of the first estimation is included in the dataset as a new variable in the second stage of the model. In the second stage of the model, various regression algorithms such as Linear Regression, Random Forest, XGBoost, LightGBM and CatBoost were used and compared to estimate the man/hour value of unplanned workload. As a result, it was observed that especially the decision tree-based algorithms such as LightGBM and CatBoost performed better than other algorithms. Thus, it is concluded that aircraft maintenance and repair processes, which are complex in nature and where the human factor is intense, can provide significant predictive power when supported by the right structural strategies.