Here is an overview of the conference program:
Abstracts in alphabetical order of author. Maintenance presentation abstracts behind operations.
Operations Presentations
Airline Disruption Management with Delay Ledgers
Chris Chin, MIT
The impact of disruptions may result in reduced capacities at airports, forcing airlines to revise schedules and delay flights. However, due to myriad factors (e.g., passengers who may miss their connections, remaining flights to be performed by an aircraft, high-valued passengers with elite statuses), a delayed flight may be more or less costly to an airline, even when compared to another similarly delayed flight. Currently, identifying optimal slot swaps between airlines requires sharing the airline-specific delay cost of each flight. However, this is not amenable as sharing these private delay costs could reveal sensitive business practices. We propose the use of a procedure called the Delay Ledger (DELED) which enables airlines to identify a set of beneficial slot swaps across a network of airports which guarantees improvements in terms of private delay costs while ensuring that no private flight-specific valuations are shared. DELED is guaranteed to lower airline delay costs, incentivizes truthful airline participation, and supports flexible airline privacy preferences. We evaluate DELED across 30 days with 8 major US airlines, resulting in average reductions in private delay costs of 8-22% per day compared to current approaches.
Schedule recovery under uncertainty with focus on resource-constrained airline ground operations
Jan Evler, Institute of Logistics and Aviation, TU Dresden
This presentation summarizes a dissertation project that has developed an integrated schedule recovery model that enables airlines to define their optimal flight priorities for schedule disturbances arising from Air Traffic Flow Management (ATFM) capacity constraints around their hub airport. These flight priorities consider constrained airport resources and different methods are studied how to communicate them confidentially to external stakeholders for the usage in collaborative solutions, such as the assignment of reserve resources or ATFM slot swapping.
The integrated schedule recovery model is an extension of the Resource-Constrained Project Scheduling Problem and integrates aircraft turnaround operations with existing approaches for aircraft, crew and passenger recovery. The model is supposed to provide tactical decision support for airline operations controllers at look-ahead times of more than two hours prior to a scheduled hub bank. System-inherent uncertainties about process deviations and potential future disruptions are incorporated into the optimization via stochastic turnaround process times and the novel concept of stochastic delay cost functions. These functions estimate the costs of delay propagation and derive flight-specific downstream recovery capacities from historical operations data, such that scarce resources at the hub airport can be allocated to the most critical turnarounds. The model is analysed within a case study of a network carrier that aims at minimizing its tactical costs from several ATFM disturbance scenarios.
Optimizing Tail Assignment through Times of Change
Tomas Gustafsson, Boeing
Assigning aircraft to flights is usually a manual or semi-automatic process following standard procedures to comply with maintenance requirements and operational constraints. If done holistically, crew turns, fuel costs and more are taken into account. This is challenging in itself but when conditions change, manual processes are hard to adapt.
By applying optimization to the problem, we can quickly adapt trade-offs among the KPIs. Here we provide a case study for a short-haul operation of ~30 aircraft, and a planning horizon of 8 days. We use fuel price as the main change indicator, and show impact on the tail allocation for various scenarios. Trade-offs are controlled between fuel cost and other KPIs, such as crew connections and operational buffers.
Disruption Management + Benefits of adding sustainability aspects to holistic optimization approach
Sebastian Heger, M2P Consulting GmbH
It is expected that almost the entire travel demand is going to be quickly returned after the COVID-19 peak. This results in an urgent need for all airlines to be well prepared also on the operations side for the upcoming seasons. Short-term initiatives and measures to secure operational actionability while increasing necessary cost efficiencies are in focus now. So, it is worthful investigating for hidden potential capabilities at the OCC to contribute to these goals as quickly as possible.
Typical disruption management solutions are focusing on the recovery of irregular operations. But these tools could be augmented to be used for proactive steering to avoid emerging disruptions as well as to cost-optimize upcoming operations. This talk will present our approach on how fuel and emission costs could be considered by a holistic optimization tool. We will also show how OCC of a major carrier is realizing significant cost benefits by establishing an extended usage of their disruption management solution.
IndiGo’s Approach to the New Normal
Jason Herter, Indigo Airlines
An airline update of how one airline has successfully navigated the COVID pandemic
Crew Recovery Using Machine Learning and Optimization
Ahmet Esat Hizir, MIT
Due to the irregular nature of flight operations, airlines need to take a range of actions to recover their aircraft and crew schedules. The limited time frames prevent airlines from using a full-scale optimization approach. Consequently, airlines typically apply recovery solutions that can be far from optimal. This study proposes a practical method that combines machine learning and optimization to find improved recovery solutions. Our procedure is based on the idea that the most effective constraints to add to the recovery models without sacrificing the solution quality, can be determined in advance by leveraging the similarities between disruptions. Our experiments show that, this approach can reduce solution time significantly while still achieving high-quality solutions
Monitoring and Predicting Safety Margins in Terminal Airspace
Peter Kostiuk, Robust Analytics
Airlines must protect aircraft in the air and on the ground from numerous risks and hazards. There are many sources of data to identify potential hazards and flight risks, but they typically come from different sources, at various times, and without any assessment of interactions, dependencies, and how combinations of conditions increase risk. Safety-related events commonly are caused by a combination of hazards, but airlines have no systems to identify combinations of conditions that increase risk. Forecasts of evolving safety conditions do not exist, forcing operators into reactive modes that prevent effective planning and mitigation. Preventing safety-related events requires identifying multiple hazards, understanding how combinations of conditions affect risk, and predicting when unsafe combinations of conditions may occur. In partnership with the NASA System Wide Safety Project over the past few years, Robust Analytics has developed a suite of applications to monitor and predict safety margins in terminal airspace.
The Flight and Airport-Airspace Monitor (FAAM) offers airlines a single source for integrating diverse hazard and risk data into a comprehensive assessment of airspace risk status and flight risk status. The tool predicts airspace risk status four hours ahead, in 15-minute intervals. Hazards that affect an aircraft during the final 60 minutes before landing are monitored continuously. FAAM integrates the airspace safety margin forecast with a crew fatigue model to generate a comprehensive flight risk profile. FAAM currently predicts safety margins at 13 US airports, and will increase to 27 airports by June 2022.
The System Wide Analysis Network for Safety (SWANS) will deliver up-to-the-minute measurement of hazards and risk events in terminal areas throughout the NAS. Updated with current data every five minutes, SWANS provides timely insight into safety margin trends and detects increases in risk events. These events may indicate changes in airspace risk status and identify possible procedural and other deficiencies. SWANS offers a data processing, distribution, and display infrastructure for adding a wide array of data sources and algorithms to track hazards and measure and predict risk.
Pre-tactical advice using machine learning for Air Traffic Flow Management delay estimation
Sergi Mas-Pujol & Luis Delgado, Technical University of Catalonia
Airlines prepare and update their operation plan pre-tactically (D-1) to identify which flights might require intervention during the day (e.g. potential aircraft swaps or cancellations). Deviations between their plan and the execution could be related to many factors, and in particular to Air Traffic Flow Management (ATFM) regulations. A collection of machine learning models, developed within Dispatcher3, a CleanSky2 innovation action, can be used to estimate which flights are likely to be affected by ATFM regulations and the potential impact of these delays. The outcome of these individual models is integrated into higher-level interpretable predictions in an advice generator to be used pre-tactically
Weather Analytics and disruption management - a data science approach
Navid Mirmohammadsadeghi, Tim Niznik, Yuxi Xiao, American Airlines
Inclement weather elements (e.g. convective, lightning, wind, snow, etc.) can have severe disruptive impacts on airline operation, resulting in Irregular Operations (IROPS). Fast recovery from IROPS and improving resiliency are top priorities for many airlines post-pandemic. Understanding weather patterns and uncertainty is key to planning and executing a reliable operation. Although past studies have explored weather impacts on various aspects of the aviation industry, the complex interactions between air transportation systems and weather make quantifying this relationship an active area of research and a critical need for airlines. For this study we collected multiple years of historical weather data – ASOS (Automated Surface Observing System) ]– and applied text mining techniques on the actual weather conditions to identify similar patterns via clustering. The insights from this study have been helping the operation center at American Airlines learn from past IROPS events and form the foundation for a data-driven approach to operational decision making.
Fair and Risk-Averse Demand Capacity Balancing in Structured Airspace
Luying Sun, Weijun Xie, Peng Wei, Robert Hoffman, Bert Hackney; Virginia Tech
Demand capacity balancing (DCB) has been explored in the unmanned aircraft system traffic management (UTM) and urban air mobility (UAM) contexts to mitigate potential traffic congestion. The traffic management initiative (TMI) in conventional aviation is one type of DCB for structured airspace, i.e., an air traffic route network. As one of the TMI or DCB techniques, the Collaborative Trajectory Options Program (CTOP) manages the imbalance between demand and capacity and provides aircraft operators flexibility with different route options. In our recent works, the stochastic CTOP rate planning models aimed to minimize the total delay and assign the best options to aircraft considering the airspace capacity uncertainty. However, the existing stochastic models in the literature (including our previous models) neglect the disparity among different aircraft, sometimes causing significant delays on certain flights. Besides, inaccurate prediction or insufficient data in stochastic models can cause misleading decisions. In this paper, besides minimizing the total route and delay cost, we consider the fairness among all the aircraft using the min-max fairness measure and use risk-aversion to achieve better decision-making against capacity uncertainty and inaccurate predictions. Notably, we formulate three fair and risk-averse centralized disaggregate models: a two-stage static model fairly assigning ground delays at the beginning of planning horizon, a semi-dynamic model fairly assigning ground delays to the aircraft scheduled departure time, and a dynamic model allowing the ground delay to be fairly assigned before the actual departure time. Finally, we perform numerical studies based on historical data on the U.S. east coast to demonstrate the effectiveness of the proposed models.
Machine Learning in Aircraft Recovery Problem with Dependent Connection Time
Keji Wei, CAE
A multi-commodity flow model has been successfully used for modeling various transportation problems such as the tail assignment problem, fleet assignment problem in airline industry. However, as one of most used multi-commodity flow models, the classical time-space network is difficult to generate connections between two events when the connecting time is dependent, which in turn, lead to a huge amount of additional spent in the total running time. Those level of time spent is particularly unacceptable in airline recovery field. Thus, we are trying to introduce a machine learning based approach to address this dependent ground time issues. We take reinforcement learning with graph attention model to capture the operational details needed to produce a realistic model of transportation operations.
Taking airline recovery problem as the case study: we present a detailed comparison of our integrated comprehensive approach with the various other approaches found in the literature and in practice, using a representative case from multiple airlines in the world. We validate that, across a variety of scenarios, our modeling and algorithmic framework can yield significant time reduction.
Maintenance Presentations
Optimizing Aircraft Engines’ Maintenance Schedule
Mohammad Rahim Akhavan & Keith Dugas, Air Canada
Engines are the most valuable and complex assets of an airline. Approximately half the maintenance cost is related to engine maintenance events. Regulatory agencies mandate that an airline upholds a maintenance program based on three core constraints, the number of hours/cycles or calendar days that have accumulated since the last refurbishment of a component installed on an engine. The problem complexity for an engine’s life span is in millions of variables. In this presentation, we explore how Air Canada built a strategic planning tool that determines the optimal schedule of engines’ shop visits by minimizing its total remaining life while considering extensive list of operational constraints
Data Driven Analytics for Maintenance Operations Performance
Dimitry Gorinevsky, Vitali Volovoi, Mark Tedone, Mitek Analytics, M2P Consulting, Stanford University
Advanced decision support analytics is needed for operation of maintenance supply chain in ever changing conditions. The two discussed AI applications are centered around maintenance and logistics process models built from existing data by custom ML tools based on convex optimization. The AI provides monitoring and observability of maintenance operations, reliability, and supply chain. From system and controls perspective, supply chain management processes can be viewed as three cascaded loops operating at different timescales. The slowest, strategy loop, supports investments (e.g., allocation planning) and has been extensively modeled. The fastest loops support tactical management of individual events, e.g., AOG response or maintenance repair recommendations; there are several software tools for that. We focus on a third, intermediate, operational loop of process analytics, above the tactical and below the strategic. The innovative comprehensive apps were initially developed for USAF. They are deployed for effectiveness monitoring and performance observability of DoD supply chain processes. We will discuss the applicability of these AI tools to the commercial airline industry. The predictive analytics can be used to control deviations from strategic planning in actual operation of the supply chain and maintenance. Such analytics are particularly relevant as the commercial airline industry re-emerges from pandemic shutdown with schedules that are in continuous flux and defy traditional forecasts of next year operation based on previous year data. Some of specific topics will include: adjusting inventory by identifying stockout conditions and recommending redistribution in advance of an AOG, identifying and remedying supply issues in the reverse logistics loop, and reliability-centric management of maintenance demand.
Designing predictive maintenance using Gaussian process learning: A case study of the landing gear brakes
Juseong Lee, Mihaela Mitici, Delft University of Technology, The Netherlands
With the increasing use of on-board sensors that continuously monitor the health of components, aircraft maintenance is shifting to predictive maintenance (PdM). Traditionally, aircraft maintenance is designed such that inspections and/or component replacements are performed at fixed periods of time (e.g., every month, every 2 years). For PdM, the health of the components is evaluated continuously, and inspections/replacements are triggered as soon as degradation thresholds are exceeded. The challenge is to specify these thresholds. One way to evaluate these thresholds is by means of extensive simulations. This is particularly challenging since the evaluation should consider multiple objectives: maintenance costs, system reliability, aircraft downtime, etc. Choosing a low degradation threshold to trigger replacements will ensure system reliability, but will lead to frequent, costly replacements. Choosing a high degradation threshold will lead to less frequent replacements, will decrease the wasted life of components, but may lead to a component failure. This presentation introduces an efficient algorithm to design PdM strategies using Gaussian process (GP) learning. We train GP models that estimate multiple objectives of a maintenance strategy, i.e., GP models provide a mapping from the design space (e.g., degradation threshold) to the objective space (e.g., costs, reliability metrics). We rapidly estimate the objectives using these GP models without computationally expensive simulations. At each iteration, we sample design parameters whose objectives are expected to be Pareto optimal based on the estimation of the GP models. Then, only these sampled design parameters are used for simulations to assess the objectives. Finally, the new simulation results are included in the training data to update the GP models. Repeating these steps, we generate a set of Pareto optimal PdM strategies. The proposed algorithm is applied to design PdM strategies of landing gear brakes. As a result, we identify multiple Pareto optimal maintenance strategies and analyse the trade-off between cost efficiency and reliability. Also, for this case study, the proposed algorithm generates Pareto optimal maintenance strategies with a much smaller number of simulations (less computational cost), than traditional approaches such as genetic algorithms. In general, the proposed algorithm is readily applicable to design other PdM strategies for different components of aircraft efficiently.
Online model-based remaining-useful-life prognostics for aircraft cooling units using time-warping degradation clustering
Mihaela Mitici, Ingeborg de Pater, Delft University of Technology, The Netherlands
Remaining-useful-life prognostics for aircraft components are central for efficient and robust aircraft maintenance. In this presentation we discuss an end-to-end approach to obtain online, model-based remaining-useful-life prognostics by learning from clusters of components with similar degradation trends.
First, time-series degradation measurements are clustered using dynamic time-warping. For each cluster, a degradation model and a corresponding failure threshold are proposed. Using a health indicator, a component is diagnosed as healthy or unhealthy. Once a component reaches the unhealthy stage, a degradation model is selected based on the similarity between the degradation trend of this component and clusters of a library of health indicators. This approach exploits the potential learning from degradation trends of other components.
Second, we use the obtained cluster-specific degradation models, together with a particle filtering algorithm, to obtain online remaining-useful-life prognostics.
As a case study, we consider the operational data of several cooling units originating from a fleet of aircraft. The cooling units are clustered based on their degradation trends and remaining-useful-life prognostics are obtained in an online manner. The results show that our proposed methodology is able to identify the degradation models of components and estimate their RUL. From a practical point of view, our RUL estimation results have the potential to support aircraft maintenance stakeholders with maintenance task scheduling.
Alarm-based predictive maintenance scheduling for aircraft engines with imperfect Remaining Useful Life prognostics
Ingeborg de Pater, Arthur Reijns, Mihaela Mitici, Delft University of Technology, The Netherlands
The increasing availability of condition monitoring data for aircraft components has incentivized the development of Remaining Useful Life (RUL) prognostics in the past years. However, only few studies focus on the integration of such RUL prognostics in the maintenance planning. In this presentation, we first develop RUL prognostics using a Convolutional Neural Network. Next, we introduce a dynamic, predictive maintenance scheduling framework for a fleet of aircraft integrating imperfect RUL prognostics. We generate RUL prognostics for aircraft turbofan engines using a Convolutional Neural Network (CNN) and multi-variate sensor measurements. We obtain RUL prognostics for which the errors decrease towards the moment of failure. Yet, since prediction errors are non-zero, we refer to these prognostics as imperfect. Next, we integrate these imperfect RUL prognostics in the maintenance planning of aircraft. For this, we periodically update the prognostics. Based on the evolution of the prognostics over time, alarms are triggered. The scheduling of maintenance tasks is initiated only after these alarms are triggered. On one hand, alarms ensure that maintenance tasks are not rescheduled multiple times. On the other hand, the alarms also ensure that there still is enough time to plan and prepare a maintenance task before failure. Once an alarm is raised, a maintenance task is scheduled using a safety factor. This safety factor accounts for potential errors in the imperfect RUL prognostics and thus helps to avoid component failures. We optimize the timing of the alarms using a genetic algorithm. Using our framework, the cost with engine failures account for only 7.4% of the total maintenance costs. In general, we provide in this presentation a roadmap to develop data-driven RUL prognostics and to integrate these (imperfect) RUL prognostics into the maintenance planning of a fleet of aircraft.
A Framework for Optimizing PHM Analytic Hyperparameters
Shashvat Prakesh, Collins
PHM analytics are, generally, continuous degradation models which drive a discrete maintenance action. Recently, more full flight data sets have become available from a wider array of commercial airline operators. As cloud computing capabilities grow, so does the potential for large scale model deployment. Thus, there is a need to formalize methods which deliver maximum impact to the customer fleet. Increasingly, the models must be explainable and this requires an amalgam of physically explainable features with data-driven heuristics and decision boundaries. Current approaches at evaluating prognostic models tend to rely heavily on the binary classifier confusion matrix, pitting reality and prediction in a true-false comparison. This simplistic approach and requires too many arbitrary boundaries around time horizon and thresholds. A more apt approach will first, consider the holistic signal behavior near the event, and then drive a decision based on operating costs. We propose a framework that will meaningfully evaluate the performance of a prognostic model and thereby offer a means to optimize it. The first framework element is model sensitivity to the intended failure mode. The targeted event is detectable when the model output shows sufficient sustained deviation from its nominal, historical value range near the event. An appropriate objective cost function will drive a model towards greater sensitivity to detectable wear-out modes. The second framework element sets the decision boundary or the alerting threshold based on operating costs. Alerts should be optimally placed in a manner that balances the cost of acting on a PHM alert to the cost of inaction. With this framework, multiple features can be evaluated individually or as a weighted model with the sensitivity and operational costs informing the overall loss function. By modelling the design tradeoffs, the analytic concept can achieve the desired detectability with sufficient lead time without sacrificing too much remaining useful life.
Cost-efficient aircraft maintenance planning integrating task and workforce scheduling over multiple bases
Jose Ignacio Sanhueza, Felipe Delgado, Mathias Klapp, Pontificia Universidad Catolica de Chile, Santiago, Chile
Aircraft maintenance centers offer technician’s person-hours to airlines on different bases to ensure aircraft serviceability. These centers receive a weekly list of tasks for each aircraft in the airline fleet, demanding skilled jobs, that must be completed. These centers have a workforce composed of full-time technicians who can be multi-skilled. If demand exceeds supply, these centers can also hire on-call technicians at a higher cost and outsource the operation in extreme cases. This paper goal is to minimize the cost of completing maintenance tasks with the available resources. We present a mathematical model that helps in efficiently planning when and where to execute each aircraft maintenance task, integrated with a work shift schedule for each available full-time and on-call technician. We evaluate the potential impact of additional sources of flexibility in the work- force planning operation, such as multi-skilled technicians or the dynamic re- location of technicians between different maintenance bases using the available flight resources. Several studies within the OR community have studied air- craft maintenance and workforce scheduling, which assign tasks to workforce considering a fixed itinerary in a mid to short period of time, but most of them don’t take both decisions independently, don’t include multiple maintenance bases, and discard the above sources of flexibility. This research integrates both decisions and all those factors.
Condition-Based Maintenance of an Aircraft Fleet Under Partial Observability: A Reinforcement Learning Approach
Iordanis Tseremoglou, Bruno F. Santos, Delft University of Technology, The Netherlands
With the increased number of sensors installed in modern aircraft and the rapid development of machine learning prognostics algorithms, Condition-Based Maintenance (CBM) has attracted much attention during the recent years. However, combining the CBM data with the routine/non-routine tasks already included in the Maintenance Planning Document (MPD) of the airline to generate an efficient maintenance schedule remains a challenging problem, due to the uncertainty included in the output of the prognostics. To the best of the author's knowledge, there is no work addressing the scheduling problem of routine/non-routine tasks combined with RUL prognostics for a wide fleet of aircraft. In this paper, we therefore present a two-stage framework for scheduling the maintenance of a commercial aircraft fleet, each aircraft having multiple monitored systems. In addition, it explicitly considers uncertainty in the output of the prognostics algorithms, while it also takes into accounts the multiple routine/non routine tasks included in the MPD. An overview of the framework is presented in the attached picture. In the first stage, we formulate the decision-making process for the maintenance of each monitored component individually as a Partially Observable Markov Decision Process. This is tackled by the task scheduling block. Specifically, the prognostics-driven tasks, along with the available maintenance slots and the updated RUL predictions will be passed into the task scheduling block. The task scheduling block using a POMCP-based algorithm, will solve a relaxed version of the scheduling problem, i.e., the hangar capacity constraints and the current maintenance schedule are not taken into account. The solution would consist of a planning output, including the optimal maintenance actions for each individual component for the considered planning horizon. In case a component is scheduled for maintenance within the planning horizon, this planning output will be treated as providing also the optimal maintenance date for the component. In the second stage, the Reinforcement Learning agent, taking into account the previously defined optimal maintenance date of the prognostics-driven tasks, as well as the due date of the routine/non-routine tasks, determines which aircraft to schedule and in which maintenance opportunity. The chosen actions are constrained by the hangar capacity and the current maintenance schedule, since last-minute schedule changes should be prevented. The model’s objective is to match the available resources and maintenance opportunities with the aircraft demand (aircraft maintenance tasks) at each point in time. To illustrate our approach, a case study is conducted using aircraft data from multiple systems of an aircraft fleet. A preliminary experiment was performed for a fleet of 20 aircraft with 200 CBM tasks, using a rolling horizon approach. For every day of the planning horizon, the prognostic models predict a RUL that follows the normal distribution. The preliminary results indicate that 98.5% of the considered components were maintained on time. Also, the average RUL exploitation of the components was found to be approximately 70%. The next steps will focus on extending the model to account for more aircraft and for the routine/non routine tasks.