Predictive maintenance analytics and implementation for aircraft: Challenges and opportunities

The increase in available data from sensors embedded in industrial equipment has led to a recent rise in the use of industrial predictive maintenance. In the aircraft industry, predictive maintenance has become an essential tool for optimizing maintenance schedules, reducing aircraft downtime, and identifying unexpected faults. Despite this, there is currently no comprehensive survey of predictive maintenance applications and techniques solely devoted to the aircraft manufacturing industry. This article is an in‐depth state‐of‐the‐art systematic literature review of the different data types, applications, projects, and opportunities for predictive maintenance in this industry. The goal of this review is to identify, and highlight the challenges and opportunities for future research in this field. This review found that the current focus of research is too biased towards aircraft engines due to a lack of publicly available data sets, and that greater automation is an important step to optimize aircraft maintenance to its full potential.

intervals to reduce the probability of failure in the future. Interval periods for PM are generated by following maintenance programs, such as the Maintenance Review Board Report 6 , where engineers use their experience to perform experiments and collect data to determine the most appropriate length of maintenance intervals.
There are many definitions for predictive maintenance (PdM), but the common thread is that PdM is to review the data from the mechanical conditions, operating efficiency and similar indicators of the condition of a mechanical device, to make appropriate maintenance decisions as to maximize the interval between repairs 5 . PdM is where the system is regularly monitored, and maintenance action is only triggered by a predefined condition of the system. PdM can exploit networks of sensors to gather data that can be analyzed to identify the health and degradation of a given system. By analyzing a systems physical parameters such as temperature, pressures, or vibration using either trend analysis, pattern recognition, or statistical analysis, it is possible to predict the condition of the system at which failure is imminent. Therefore, before the degradation level reaches this threshold, the system that is about to fail can be replaced. PdM is not a perfect strategy. Performing a combination of the different maintenance strategies is still the most reliable approach for maintaining aircraft effectively.
Aircrafts are more capable than ever of recording vast amounts of sensor data across almost all of their components in flight, with an Airbus A380 having up to 25,000 sensors 7 . This increase in data has driven greater use of data-driven PdM, that is to build and train PdM algorithms using data rather than domain experience. The data collected from an aircraft can be analyzed using statistical models to determine relationships and generate predictions of measured parameters. There are three main use cases for PdM in the aerospace industry; real-time diagnostics, real-time flight assistance, and prognostics 8 . Real-time diagnostics allow for faults detected in flight to be recorded for immediate repair on landing, and real-time fight assistance can provide guidance for the pilot. Prognostics is responsible for predicting the degradation of a system by interpreting the operational and environmental condition to estimate the system's remaining useful lifetime (RUL) 9 or its end-of-life (EOL). These metrics can be used to help determine the optimal maintenance schedules for replacing and repairing aircraft components to maximize their lifespan. Without effectively utilizing this data for PdM, terabytes of available data are effectively wasted where it could be used to save money, time, and manpower.

Contributions
There are several state-of-the-art reviews for PdM; however, to our knowledge, there does not exist an exhaustive evaluation of the current state-of-the-art focused on PdM for all available aircraft systems.
This paper compiles and compares the current demographic of publications in the field of aircraft maintenance, to support readers in future research. The documents collected can be used to identify areas where predictive maintenance has and could be applied, which datasets and predictive models have been used to compare results against, what tools the industry has been developing to aid in these problems for customers, and the challenges and new opportunities the field contains.

Paper organization
This paper follows the review structure outlined in Figure 1 to provide a thorough literature review and provide a detailed discussion of future opportunities for new researchers in this field. It starts with an extensive review of available academic literature regarding which data types can be used for prognostics in Section 2, and what benchmark datasets are used for replicating results. This is expanded by identifying which models and tools have been applied to these datasets and others in different PdM applications in Section 3. Section 4 outlines different projects and industrial services for PdM to highlight the growth within academia and industry. Section 5 reviews the challenges researchers in this field will encounter, as well as opportunities afforded by new technologies. Section 6 concludes the main points, summarizing the trends from the most impactful papers from the literature review, and identifying key research areas in the future.

Research methodology
The purpose of a literature review is to gather the available sources on the topic being researched and perform a thorough evaluation to identify research gaps, trends, and so forth. To conduct an effective literature review then, a methodology is required to structure the research toward the goal of conducting a thorough review. This research methodology consists of three main parts as follows, establishing the research questions this review intends to answer, conducting a bibliometric analysis, and a thorough review of the material to identify trends and research gaps. A diagram for the methodology can be seen in Figure 2. The scope of this review will only extend to 2015, as only the state-of-the-art techniques are being evaluated by this review.

Research questions
This paper will attempt to answer the following research questions to provide an effective review of this field. Two reviews were found to have a more direct focus on aircraft.
Wen et al. 13 conducted a review of data-driven prognostic algorithms, for conventional and deep learning models. Aircrafts were highlighted as a major application of PdM, and they dedicated a subsection to reviewing recent papers applying different models to a group of aircraft equipment. Finally, 14 reviewed the trends and challenges for PdM of aircraft engines and hydraulics specifically. It reviews different prognostic methods that have been used for aircraft engines and hydraulics in recent years, followed by a case study using data from an aircraft's hydraulics system using a support vector machine (SVM).
Of the available reviews, there was only one directly focused on just two aircraft systems. There are far more aircraft systems than those covered in these reviews, and no paper found delves too deeply into the aircraft industry, and what tools are used within it. Therefore, this state-of-the-art review will be the first exhaustive review solely focused on aircraft in both academia and the aircraft industry, and identify the challenges

Analysis of countries
The countries that these journals originate from were also analyzed to identify which countries are producing the most papers in this field.
The top eight countries and their respective paper counts can be seen  Natural language Pilot complaints, equipment failure logs, 21 and post flight reports Graphical data Imaging of aircraft fuselage and wing 22 were selected for their aerospace focus and consistent use within 10 or more state-of-art-papers in the past 5 years.

Data types
There are three main data types for aircraft maintenance data, time series, natural language and graphical data. The source, use, and papers where this data has been used for aircraft are displayed in Table 3.
The number of time series datasets greatly outnumbers the others due to the ease in collecting and processing the data compared to natural language processing (NLP) and computer vision required for language and graphical data, respectively. NLP could provide a suitable redundancy for identifying indicators of problems with aircraft; however, widespread application of NLP is doubtful as a major challenge of NLP is language barriers and maintaining the meaning of sentences 16 . Logically, there will be inconsistencies between airline reporting protocols and the written language that pilots use around the world, producing inconsistent data that will be more difficult to accurately process.
Graphical data have rarely been used for aircraft PdM so far, but its greatest use is for technicians inspecting aircraft bodies, and since 1998, it has been proposed that much of this work could be offloaded to robots 17 . This can be performed by gathering graphical data consisting of photos using robotics systems. One such approach recorded aircraft fuselage images taken by drone Aircraft by Airbus for automated fuselage inspection, reducing inspection times from 2 h to 10-15 min 18 .

Benchmark datasets
There are datasets that have been released to encourage research in the field and enable greater cross-comparison between work. Many of these datasets have been available online in a Data Repository operated by NASA 15 , but only one has been used in the field of aircraft maintenance.
The Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) is a transient simulation of a large commercial turbofan jet engine, with a realistic engine control system developed by NASA 23 . It has been used frequently in publications to generate multivariate time series engine datasets for developing novel prognostics and health management (PHM) models. The most commonly used example is a set of run-to-failure datasets, 24

PREDICTIVE MAINTENANCE APPLICATIONS AND CONTRIBUTIONS
Strategies for performing PdM are being applied to a wide range of different industrial fields and applications, with many novel methods developed in recent years. Many authors have applied different methods to applications, using a mix of data analytics and machine learning. A number of papers have summarized and compared different machine-learning algorithms for PdM in the general industry already 10,32 . This section identifies the key state-of-the-art methods published in journals in recent years for specific applications. This section highlights the paper's key features, the highest-performing models for each appellation, and the future work proposed by each paper to encourage future innovations. In doing so, this section works to answer our second research question.
Of the papers highlighted in this section, different traditional and ML models were applied. These are shown in Table 4, with their respective strengths and weaknesses. What follows is a review of the different PdM applications that have been addressed within the aircraft industry specifically and the publications that represent the current state-of-the-art. Figure 5 shows a table of all the papers that were highlighted by this review, both for aircraft specifically, and transferable industries.

Aircraft engine
Aircraft engines are complex and require regular maintenance, making up 35-40% of the total aircraft maintenance expenses from an operator 48 . Turbofan engines can contain large suites of sensors that record values such as fan inlet temperature and pressure, and physical fan speed 49 . C-MAPSS generated datasets have been found to be used most frequently in publications, particularly the datasets released for the PHM 2008 data challenge, 24 which has cemented itself as an established benchmark for new approaches.
State-of-the-art reviews have already been conducted investigating aircraft engines. Due to the time series nature of most engine data, it was suggested that machine learning models will be used more frequently, specifically Long Short-Term Memory Networks (LSTMs) 14 .
However, this paper only highlights LSTM examples that are hydraulics focussed. Another paper also supports a move towards LSTMs; however, also highlighting Random Forests as a powerful traditional model 50 . For this section, we have looked at the paper that both fit within these trends and those that defy them. Table 5

Aircraft bearings
Bearings are components that reduce friction between moving parts moving relative to one desired axis. In aircraft, they are commonly found in engines, landing gear, hydraulic fuel pumps, doors, and cockpit controls. The reliability of a bearing is paramount, as a single bearing failure can potentially jeopardize hundreds of lives 53

Hydraulics and pneumatics
Hydraulics is a mechanical function that operates through the force of liquid pressure. In hydraulics-based systems, mechanical movement A comparison of many state-of-the-art machine learning algorithms was performed by testing against hydraulic system sensor data 32 . They found that the traditional methods with feature engineering outperformed deep learning models likely due to the small dataset size, which deep models struggle more with. Table 7 contains a list of the papers covering PdM for hydraulics and pneumatic's that were investigated as part of this review.

Fuselage
The fuselage and frame of an aircraft are just as vital a component as the engine and are liable to damage from bird strikes, lightning strikes, and degradation over time. In recent years, particle 36 and Kalman filters 37 have been used to estimate and predict the size of flaws and cracks in the frame and wing of the aircraft leading to significant cost reduction. For a more thorough monitoring, structural health monitoring has been used to assess the condition of engineered systems.
It is conducted by observing and analyzing the sensor measurements of a system to assess the health of the structure. An overview of piezoelectric transducer-based SHM system technology for aircraft addresses some of the challenges of applying SHM to aircraft but suggests that the field is expanding from diagnostics to prognostics, using data-driven methods to predict the life and performance of the Investigate two problems: (1) The difficulties simulating random mutation of degradation process.
(2) How the degradation process is split into stages by time.
Wu et al. 56 LSTM Predict health of a manufacturing system. Superior classification of critical states than SVM.
Increase the accuracy on early stages by employing parameter tuning within the architecture of the RNN.
aircraft structure 59 . It has been suggested that the aviation industry is unable to exploit SHM-based inspections as it is not cost-effective, and the weight of the sensor's systems must first be reduced 60 . SHM has been used in other industries already, some elements of which could be reapplied to future SHM for aircraft when these challenges have been addressed. In recent years, it has been used to identify defects in wind turbine blades 61 and railway tunnel structures 62 . Table 8 contains a list of the papers covering PdM for the aircraft body, and transferable papers covering SHM that were investigated as part of this review.

Methodologies
The applications of PdM in aircraft are not the only innovations in recent years, as several publications have focused on the methodologies implemented alongside them. A methodology to estimate overall systems-level RUL, with the goal of interpreting component-level RUL to make replacements that will benefit the system RUL, was proposed 63 . Despite some existing state-of-the-art methodologies, one major drawback is the lack of a rigorous process for defining requirements and proposed a systematic derivation for system requirements for the further development of PHM systems 64 . Table 9 contains a list of the papers covering maintenance methodologies that were investigated as part of this review.

Additional aircraft systems
There are other specific aircraft systems that have been optimized using PdM. The Auxiliary power unit is an essential piece of equipment for an aircraft; however, it has a nonlinear degradation process.

TA B L E 9
Publications proposing state-of-the-art methodologies for PdM Incorporate troubleshooting tasks to the planning optimization process.
Data-driven and physics models alone make poor predictions on these, so a hybrid of the two was proposed, feeding exhaust gas temperature data into an LSTM to generate the RUL 19 . Random Forest has been used to assess the performance and predict the RUL of an aircraft auxiliary power unit 34 . Using Random Forest and Bayesian dynamic models to quantify degradation, achieving a prediction error rate of less than 4%. It was tested against a multivariate ACMS report from a commercial aircraft fleet covering values such as pressures and temperatures for air, bleed, and oil. Low-pressure environments are more prone to corona and arc tracking, and three methods were proposed to monitor them 66 . This includes an example of graphical data used by UV imaging sensors to detect arcs. These methods allow for online monitoring of this activity and are compatible with PdM approaches. Table 10 contains a list of the papers covering PdM for the additional aircraft systems that were investigated as part of this review.  As well as the grants afforded to academic research and universities, companies are developing services to handle and process the growing available data to enable more optimized maintenance. The most popular of these services are shown in Table 12, organized by

PREDICTIVE MAINTENANCE PROJECTS
year of release. Most of the tools provide the benefits of reducing aircraft downtime and return to service time, which highlights these as the key needs for airlines. The features each of these services provide are displayed in Table 13

DISCUSSION: CHALLENGES AND OPPORTUNITIES
There are a number of challenges that researchers will face, summarized in Figure 7. This section outlines the challenges that researchers TA B L E 1 2 Identified PdM services and tools provided by key members of the industry Application of intelligent diagnostics related to spare parts optimization for 20 aircraft, saving 1400 h of downtime in 3 months (unnamed C-130J operator 84 ).
Ascentia 85 Collins Aerospace 2020 Decrease in potential delays and unscheduled maintenance.
Applying advanced data management and analytics services to reduce potential delays and cancellations related to components and systems monitored on the Boeing 787 fleet by 30% 85 .
Smart Link Plus 86 Bombardier 2020 Streamline customer service relationships and efficiently dispatch, troubleshoot and track aircraft service needs.
None found yet Insight Accelerator 87 Boeing 2021 Improved efficiency for predictive analytics, faster insights into issues, and simplified data analytics.
None found yet in this field will need to overcome to enable the widespread and effective use of PdM for aircraft. New technologies that can be exploited to provide valuable opportunities to expand research in this field are also highlighted. The investigation of these works answers our third research question.

Challenges
Few papers that we found suggested that their solution addressed the exact challenges faced by the field, rather than providing novel models to increase the performance of PdM for a given application. Given the

Data cleaning and preprocessing
Data cleaning is seldom mentioned in the journal articles discussed as part of this review. There is a common assumption in these papers that the data being used in these examples are already clean, and in cases of using publicly available datasets such as the C-MAPSS dataset, this is the case. However, from a practical standing, the data will not always be clean and across fleets of aircraft, anomalies, noise, and mistakes will make analyzing the data more difficult. Aircraft data is noisy as it is being recorded and must be filtered before being stored. Clean data allow for more efficient, reliable, and accurate analysis of the data, and given the volume, traditional data cleansing is not an effective option. to lead to the development of an "overall solution with several interacting components" but questions both the costs of the development of deep learning tools against the benefits they propose and the lack of consistent high-quality data in the field. As computational power and data collection capacity increase these concerns will be mitigated, and the use of a single automated system appears to be a common goal for those in the industry. Automated machine learning (Auto-ML) could also be applied to build complex DL systems with minimal human assistance required. Tools like Auto-Keras can be used to build DL models for regression, classification, and time forecasting problems, which have applications for predicting aircraft system deterioration.

Data fusion
While system deterioration can be predicted from single data sources, data fusion can integrate data from multiple sources. This improves the accuracy of the prediction of deterioration and better utilizes the abundance of recorded data. A data-level data fusion method for early detection of incipient faults and achieved a lower variance before the occurrence of incipient faults when tested of a C-MAPSS generated dataset 108 . It can also be used at a decision level, such as for predicting the RUL of an aircraft by interpreting it as a convex optimization problem instead of the traditional linear regression problem and outperforming preliminary decisions using individual sensors. Datasets generated using C-MAPSS have been used as they provide up to 21 parameters. There is room to improve on this work, either by integrating a greater number of parameters or applying the method to real-time prognostics.

CONCLUSION
This paper serves as a state-of-the-art review to identify the novel solutions that are being applied to PdM problems and plot the current landscape of the field. PdM can be more optimized than alternative maintenance strategies for maximizing the RUL of aircraft components.
By applying prognostic methods to the growing number of available benchmark datasets, it is more possible than ever to develop novel PdM methods. Further development of PdM is inevitable, given the rising number of novel methods and potential applications in the field. The enhancements afforded by new technologies such as robotics and AI will further optimize and automate these procedures. Greater use of it has the potential to greatly reduce maintenance costs for aircraft manufacturers and operators.
In the current landscape, PdM is performed by data engineers in the industry and researchers in academia, but it is inaccessible to in-experienced users who could benefit from it most. Even easily accessible tools such as Microsoft Azure, which possess PdM guides using C-MAPSS data 109 , require some level of domain knowledge and programming experience to understand and use effectively. Dedicated PdM tools that utilize new technologies such as AI and Auto-ML to provide greater automation would enable a wider user base. Automated tools will enable a greater number of people to build PdM models on aircraft data. Greater research into automated tools in this field will encourage both more development and use in the industry, leading to greater savings and safety afforded to in-service aircraft.
There is an abundant new technology that will provide opportunities to optimize and automate this work in the future. Many of these will directly mitigate the challenges highlighted in this review, but will these require the integration of physical new technology into the industry, which will be a slow process to normalize amongst large fleets. Using AI and Auto-ML to provide greater automation could mitigate many of these challenges and enable a wider user base. Automated tools will enable a greater number of people to build PdM models on aircraft data. Greater research into the integration of AI in this field will encourage both more development and greater use in the industry, leading to greater savings and safety afforded to in-service aircraft.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available from the corresponding author upon reasonable request.