14:30
S6: Prognostics for CBM Strategies at Amphi Bézier (AB)
Chair: Marinus Bos and Nick Eleftheroglou
14:30
20 mins
|
Exploring Prognostics Requirements: Reliability, Robustness & Feasibility
Nick Eleftheroglou
Abstract: Condition-based maintenance (CBM) represents a cornerstone strategy in modern engineering, shifting maintenance paradigms from fixed schedules to dynamic approaches based on condition monitoring (CM) data collected in real-time from assets under study. Essentially, CBM aims to optimize maintenance activities by harnessing CM data to assess the current condition of assets and forecast their degradation trajectory until failure. This approach minimizes downtime and maximizes operational efficiency. Prognostics play a pivotal role in CBM by predicting the future health and performance of assets based on their present condition and operational context.
To effectively utilize prognostics for decision-making regarding CBM strategies, prognostic methodologies must meet specific criteria: feasibility, robustness, and reliability. Feasibility entails the ability of the prognostic methodology to be trained with limited degraded data, as acquiring extensive degradation data is often cost-prohibitive. For instance, in aviation structures, only a single full-scale fatigue experiment occurs during the certification process, providing just one degradation history for training prognostic methodologies. Secondly, prognostics must demonstrate robustness, ensuring reliable performance across a broad spectrum of operational conditions not seen during the training process. Thirdly, prognostics must be reliable, given their inherent uncertainties. These uncertainties stem from various factors, including past, present, and future conditions, as well as modeling uncertainties, which collectively contribute to the overall uncertainty within the field of prognostics. For Remaining Useful Life (RUL) prognostics, this means considering RUL as a random variable. To ensure the reliability of RUL prognostics, the mean RUL prediction should closely align with actual RUL values, and the estimated uncertainty of RUL should provide decision-makers with actionable information.
This work will explore the aforementioned prognostic requirements using state-of-the-art prognostic methodologies, such as the Adaptive Non-homogeneous Hidden Semi-Markov Model (ANHHSMM) [1], and the Similarity Learning Hidden Semi-Markov Model (SLHSMM) [2]. In addition, a novel time-dependent prognostic measure will be introduced for the first time. These methodologies will be validated to composite aerospace specimens [1].
The training dataset consists of strain data exclusively from specimens subjected to fatigue loading conditions. Conversely, the validation-testing dataset includes strain data from a specimen subjected to both fatigue and in-situ impact loading, along with one specimen exhibiting a manufacturing defect. These variations represent unexpected events, unseen in the training data. Figure 1 illustrates the axial degradation histories of the training and testing dataset. These variations will enable an assessment of the robustness of the proposed state-of-the-art prognostic models. Regarding feasibility, the provided training dataset will be dynamically reduced to explore how varying data amounts influence prognostic reliability. This process will aid in identifying the most feasible prognostic methodology in terms of required training data. Finally, reliability will be assessed by comparing the accuracy of estimated mean RUL values and evaluating RUL uncertainty.
Figure 1. Axial strain degradation histories.
In Figure 2, the RUL predictions of the ANHHSMM are presented and compared with those of the NHHSMM for the specimen with the manufacturing defect, showcasing the increased robustness and reliability of the ANHHSMM. The results pertaining to the SLHSMM, the fatigue and impact specimens, feasibility, and the new time-dependent prognostic measure will be presented during the conference.
Figure 2. RUL estimations for the manufacturing-defected specimen.
In conclusion, this study underscores the importance of developing prognostic methodologies that are robust, reliable, and feasible. To achieve this goal, state-of-the-art prognostic methodologies were compared and validated using a dataset derived from composite specimens. By introducing domain adaptation along with similarity learning aspects, it becomes feasible to enhance the reliability, robustness, and feasibility of prognostics.
References
[1] Eleftheroglou, N., Zarouchas, D., and Benedictus, R. "An Adaptive Probabilistic Data-Driven Methodology for Prognosis of the Fatigue Life of Composite Structures." Composite Structures, no. 245, 2020, 112386, doi:10.1016/j.compstruct.2020.112386.
[2] Eleftheroglou, N., Galanopoulos, G., and Loutas, T. "Similarity learning hidden semi-Markov model for adaptive prognostics of composite structures." Reliability Engineering & System Safety 243 (2024): 109808.
|
14:50
20 mins
|
Enhancing Predictive Maintenance in Aviation: A Quantile Regression Approach for Health Index-Based Remaining Useful Life Estimation
Dario Goglio, Kristupas Bajarunas, Pierre Dersin, Bruno Santos, Manuel Arias Chao
Abstract: Airline delays, accounting for approximately 25% of the delays reported in 2023 [1], are
often attributed to unforeseen technical issues, placing additional strain on the already tight
airline operations. To mitigate these technical disruptions and improve operational efficiency,
the aviation industry has shown a growing interest in predictive maintenance. In particular,
data-driven solutions for predictive maintenance have gained popularity due to the increased
availability of condition-monitoring data, technological advancements, and improved algorithms.
However, developing operational predictive models in aviation forecasting the temporal evolution
of the health conditions of aircraft components (e.g., prognostic models) is notably
challenging due to stringent safety requirements, legislative regulations, and the need for
high predictive accuracy amid the scarcity of labeled data.
In this context, despite the advancements in purely data-driven models that directly predict
remaining useful life (RUL) without providing information about the current health condition
and its temporal evolution, these models are often seen as black-box models and face limited
acceptance due to their lack of transparency and interpretability [2]. To overcome this issue
and achieve practical usability in aircraft systems, recent works such as [3]–[5] have proposed
resorting to the inference of health indicators (HI) to offer a more interpretable alternative to
RUL models based on direct mappings. However, an important discussion in this direction
is whether inferred HI from truncated degradation trajectories can be efficiently extrapolated
until complete system failure, thereby enabling effective RUL modeling.
Building on prior research, [6] & [7], in this work, unit-specific HI’s are inferred through
a fusion of deep learning techniques and general knowledge on degradation dynamics. The
approach involves defining HIs through a power-law function with random coefficient, which
facilitates the modeling of RUL and intermediate degradation levels across a fleet. To address
unit-specific variability, quantile regression is used to extrapolate inferred unit-specific health index from truncated degradation trajectories until complete system failure (HI = 0) along
with confidence interval for uncertainty estimation.
To illustrate our methodology, we present a case study using the N-CMAPSS dataset of turbofan
engines [8], extrapolating the RUL after a few samples available from the estimated
HI. This case study validates our assumptions and clearly demonstrates the steps involved
in our methodology, providing a practical example of how extrapolation of estimated health
index curves enhances RUL prediction for predictive maintenance.
This research bridges a significant gap in predictive maintenance modeling by increasing the
transparency and accuracy of RUL predictions. By doing so, it holds the potential to reduce
operational disruptions, increase safety and improve efficiency within the aviation industry,
marking a step forward in the integration of advanced analytics into real-world applications.
|
15:10
20 mins
|
A viable opportunity for fielding an aircraft structural health monitoring system
Marcel Bos, Frank Grooteman
Abstract: Aircraft operators worldwide are looking for ways to maintain the availability and airworthiness of their fleets of aircraft while decreasing the cost of ownership. The sustainment costs of aircraft generally constitute a substantial part of the total life cycle costs, which implies that the application of innovative methods and technologies in the sustainment process may lead to large cost savings. An important trend in this respect is the transition from preventative maintenance to condition based maintenance (CBM). For CBM it is essential that the actual system condition can be measured and that the measured condition can be reliably extrapolated to a convenient moment in the future in order to facilitate the planning process while maintaining flight safety. Much research effort is currently being put in the development of CBM enabling technologies, among which Structural Health Monitoring (SHM) systems. Good progress has already been made when it comes to sensors, sensor networks, data acquisition, models and algorithms, data fusion/mining techniques, etc. However, the transition of these technologies into service is very slow. There are two reasons for this:
• Business Cases are difficult to define since CBM represents a disruptive technology that produces a paradigm shift for maintenance support;
• Certification is difficult as the validation of the SHM system’s capability to reliably and accurately detect impending in-service failures is extremely challenging; in addition, the procedures for obtaining maintenance credits are still being developed.
One option to validate the performance of a particular SHM system is to use a seeded fault test. This requires a high-fidelity and expensive test bench and a good a priori knowledge of the location and the nature of the failure modes that are to be detected. An alternative is to field the SHM system in one or more aircraft and evaluate its performance after a sufficient number of flight cycles. ‘Sufficient’ in this respect is indeterminate and may cover a significant part of the service life in order to be able to collect relevant data. Fortunately there are some special cases where certification of an SHM system for use in aircraft is much easier. This paper describes a possibility for fielding a SHM system on a fleet of aircraft with a minimum amount of certification issues and with a good prospect of a positive return on investment. For appropriate airframe components the application of SHM will reconcile the fail-safety and slow crack growth damage tolerance approaches that can be used for safeguarding the continuing airworthiness of these components, combining the benefits of both approaches and removing the drawbacks.
|
15:30
20 mins
|
The DICONDE information model for inspection of aircraft structures
Geo Jacob, Florian Raddatz, Gerko Wende
Abstract: Introduction
Non-destructive testing (NDT) ensures the reliability and integrity of aircraft structures by effective damage detection, localization and damage characterisation. NDT of aircraft structures involves a large amount of heterogeneous data associated with the inspection such as structure characteristics, flaw characterisation variables, NDT equipment parameters, settings, post-processing data and results. Acquired NDT data is a crucial asset. The heterogeneous and complex data needs proper integration and management to derive insights about the inspected structure. The DICONDE (Digital Imaging and Communication in Non-Destructive Evaluation) open standard is an important enabler for the storage, management and exchange of NDT data [1, 2, 3]. DICONDE is based on the Digital Imaging and Communications in Medicine (DICOM) standard [4]. However, the use of proprietary data formats, less documentation, software and isolated NDT systems introduce intricacies and challenges for data integration [5, 6]. This requires a standardized and structured approach to manage NDT data using DICONDE.
Research Question
This research focuses on adopting the DICONDE information model for the inspection of components. The DICONDE information model defines a model to map the entities from the real world. With the structured representation of different types of NDT data generated by diverse inspection methods for a component under different states, the focus is on representing entity-relationships and properties of NDT data.
Research Method
In this research, immersion ultrasonic testing (UT) data produced by a Hillger NDT system is used. Twelve composite samples at various states pre and post impact damage are inspected using UT through-transmission and pulse-echo methods. The UT system produces data in *.hgy format with the header data containing metadata such as inspection date and time, hardware configuration, probe information, coordinate information and resolution.
A pydicom-based framework is developed to convert the proprietary data format into the DICONDE based *.dcm data format [6]. The DICONDE data model is defined using the DICONDE Information Object Definitions (IODs). IODs encapsulate information about real-world data. They are grouped together based on information entities (IE) about the NDT data which contain mandatory, conditional and optional metadata. Unique Identifiers (UIDs) used in the DICONDE standard ensure global uniqueness and thus information entities can be distinguished from one another. The condition of the composite samples at two different sessions are correlated using the entity-relationship diagram adopted from the DICOM standard (Figure 1). A composite sample "CAI 18" at the component level could include information such as an identifier (ID) or a name. All other data is organized based on this entity. Each component could have one or more Studies. In this case, the condition of the sample before impact and after impact are identified by Study Instance UIDs. Each Study has Series that represent inspections using the UT through-transmission method and the pulse-echo method. The Series UID is used to identify the series. Each Series might contain one or more Instances such as raw data, processed data (C-scans), equipment, equipment settings and annotations. Each instance is identified by a unique Service-Object Pair (SOP) Instance UID.
Results and Conclusion
The DICONDE information model is suitable to handle NDT inspections from different time points and to store and manage the associated information. The relation between entities that are part of this inspection can be digitally mapped and stored using the DICONDE standard. The DICONDE open data standard is an important solution for the transition into NDE 4.0. Adopting the DICONDE open standard for NDT data formats ensures a structural approach for a robust database representation, seamless integration of data from heterogeneous NDT methods and data management across different inspection timelines for an aircraft structural component. Similarly, the completeness of NDT data enables the application of AI-based data analysis, which requires consistent data structures. The DICONDE information model addresses the critical needs for standardization, interoperability, data sharing, and integration of the NDT data for improving operational efficiency and decision-making in maintenance processes. A key challenge to overcome is the correct spatial alignment of data from different inspection systems. This is part of ongoing research and development.
References
[1] ASTM E2339 E07.11 Committee, Standard Practice for Digital Imaging and Communication in Nondestructive Evaluation (DICONDE), West Conshohocken, PA: ASTM International, 2016.
[2] ASTM E2663 E07.11 Committee, Standard Practice for Digital Imaging and Communication in Nondestructive Evaluation ({DICONDE}) for Ultrasonic Test Methods, West Conshohocken, PA: ASTM International, 2018.
[3] N. Brierley, R. Casperson, D. Engert, S. Heilmann, F. Herold, D. Hofmann, H. Küchler, F. Leinenbach, S.-J. Lorenz, J. Martin, J. Rehbein, B. Sprau and A. Suppes, Specification ZfP 4.0 - 01E: DICONDE in Industrial Inspection, German Society for Non-Destructive Testing (DGZfP), 2023.
[4] National Electrical Manufacturers Association (NEMA), Digital imaging and communications in medicine (DICOM), NEMA Standards Publications.
[5] J. Aigner, H. Meyer, F. Raddatz and G. Wende, “Digitalization of Repair Processes in Aviation: Process Mapping, Modelling and Analysis for Composite Structures.,” in Deutscher Luft- und Raumfahrtkongress 2023, Stuttgart, Germany, 2023.
[6] G. Jacob and F. Raddatz, “Data fusion for the efficient NDT of challenging aerospace structures: a review,” in Proceedings Volume 12049, NDE 4.0, Predictive Maintenance, and Communication and Energy Systems in a Globally Networked World, Long Beach, California, United States, 2022.
[7] “Pydicom,” Pydicom, [Online]. Available: https://pydicom.github.io/. [Accessed 12 January 2023].
|
15:50
20 mins
|
Evaluation Metrics for Condition based Swarm Operations
Lorenz Dingeldein
Abstract: The organization of robot swarms gains momentum with enhanced operational capabilities and heightened reliability, propelling us forward in current research. The increasing interest in the development of UAVs, which are compared to traditional aircraft and helicopters cheaper in production as well as in operation, emphasise this trend (Zhen et al. 2020; Zhou et al. 2018). In the aviation industry, swarm operations prove high value for reconnaissance missions, particularly when time and coverage performance are critical factors. Meanwhile, the UAV is subject to different loads during the execution of the mission due to environmental conditions such as wind (Chu et al. 2021; Khaghani und Skaloud 2018). Digitalization and data-driven approaches enable smart deployment strategies of swarms where machine learning and comparable techniques make a significant contribution (Theile et al. 2020). To harness the advantages of swarm operations, ensuring reliability is crucial. This can be achieved through the integration on condition-based monitoring approaches and mission suggestions considering a condition specific system usage to keep risks for system failure as low as possible (Darrah et al. 2022).
As the variety of approaches for integrating usage parameters in swarm strategy deployment increases, so does the need for evaluation metrics. While conventional metrics for evaluating CBM and PHM approaches focus on algorithm performance (Saxena et al. 2008), metrics for condition-based swarm strategies must acquire other parameters. Beside prognostic performance and algorithm accuracy the change of system condition as well as their usage needs to be monitored for the period of the mission. Mission objectives must be taken into consideration to evaluate the benefits of integrated CBM approaches. Taking the above aspects into account, an approach for evaluating the condition-based swarm organisation is derived as follows.
Based on a complete-area path-coverage task that is performed by a swarm of UAVs this paper identifies various parameters that are useful to integrate into a metric that evaluates the increase in reliability for swarm missions. Integrating these parameters into a metric and applying it to the use-case it can be verified that it is possible to delimit the mission reliability with respect to their usage and degradation. The technical evaluation of swarm performance is crucial for conducting a profitability analysis and determining the extent to which integrating CBM approaches into operations can be financially beneficial.
|
16:10
20 mins
|
Optimizing a maintenance strategy by combining age-based maintenance and an imperfect prognostic model
Ingeborg de Pater, Marta Ribeiro
Abstract: In aviation, age-based maintenance is employed for many components. With age-based maintenance, a preventive maintenance strategy, components are always maintained after a fixed amount of usage time or upon failure. For some components, historically observed failure times are available. With these historically observed failure times, the optimal age to maintain a component, minimizing the maintenance and failure costs per time unit the component is used, can easily be found with the classical renewal reward theory.
Even with age-based maintenance, however, the costs of aircraft maintenance are still high, representing around 10% of the total operational costs of the airlines [1]. Airlines are therefore interested in using a predictive (also called condition-based) maintenance strategy instead. In predictive maintenance, the measurements of the sensors installed around a component are used to predict when a component will fail. Predictive maintenance is often divided into two steps. First, the measurements are used to generate an alarm when a component becomes unhealthy, and might fail in the near future, with a fault detection model. Once such an alarm is generated, the Remaining Useful Life (RUL, the time left until failure) can optionally be predicted as well.
In literature, prognostic models are often developed with simulated or lab data, which gives very good results. When optimizing the maintenance planning with these prognostic models, it is assumed that these very good results will also be obtained in practice (see, for instance, 2,3). Planning maintenance with such very good results is straightforward, as components are simply maintained just before the predicted failure time, with a small safety margin.
In real life, however, this is completely unrealistic for most components. Unfortunately, with real-life data, it is very hard to get such good prognostic results, due to, among others, noisy, missing or unlabelled data. For instance, in 4, the authors obtain an F1-score of roughly 60% in an anomaly detection model with real data from satellite reaction wheels. Using the outcomes of such an imperfect prognostic model in the maintenance planning is complicated. By solely planning maintenance based on the outcomes of an imperfect prognostic model, the number of failures and maintenance tasks might actually increase, compared to age-based maintenance.
However, also an imperfect prognostic model might still provide useful information on the potential failures of the components. In this presentation, we will therefore combine an age-based maintenance strategy with the results of an imperfect prognostic model. As a first exploration of this topic, we exclusively focus on components for which there is a prognostic fault detection model, which aims to raise an alarm when a component becomes unhealthy (but before this component fails). We assume that this model is imperfect, with a known false positive rate (the number of false positives relative to the total number of positives) and a known false negative rate (the number of false negative relative to the total number of negatives). Here, with a false positive, the model identifies a component as healthy, while it is unhealthy. With a false negative, the model identifies a component as unhealthy, while it is healthy.
We will mathematically derive some formulas that will allow airlines to combine their imperfect fault detection model with age-based maintenance. We then analyse these formulas with a small case study, where we assume a certain lifetime distribution for the component. We show the influence of different false negative and false positive rates on the final results. Specifically, we will analyse the following topics:
• If we only maintain a component after an alarm, then the number of failures might actually increase (compared to pure age-based maintenance) due to the false negatives. Preventive maintenance, based on the age of the component, is therefore still necessary. However, the optimal age to preventively maintain a component might be later with the alarms, than without the alarms, especially if the false negative rate is low. By adjusting the classical renewal reward optimization of age-based maintenance with the alarms, we find a formula for the optimal moment to preventively replace a component. Here, we minimize the costs per time unit the component is used. In the case study, we show how the optimal maintenance moment becomes later (i.e., how we preventively maintain the component after a longer period of usage) if the false negative rate of our model becomes lower (i.e., if we detect more failures before they occur). However, we also show how even models with a very high false negative rate can still contribute to lowering the maintenance costs.
• If we maintain a component after each alarm, then the number of maintenance tasks might actually increase (compared to pure age-based maintenance) due to the false positives. With Bayes theorem, we therefore derive a formula for the probability of a false positive based on the age of the component. In the case study, we show how this will allow airlines to reduce the number of maintenance tasks by only acting upon alarms from a certain age of the component onwards.
• Last, we combine the two points above and optimize a preventive maintenance strategy taking into account the age of the component and the false positive and false negative rate of the fault detection model.
The derived formulas can be used with any type of lifetime distribution, and any type of fault detection model. Airlines can therefore simply use these formulas with their own models, and also use imperfect fault detection models to lower their maintenance costs. Since we show that prognostic models do not need to be perfect to be employed in practice, the proposed idea may also contribute to the certification of prognostic models by EASA.
References
1. Markou, C., & Cros, G. (2020). Airline maintenance cost executive commentary FY2019 data Public Version. Maintenance Cost Technical Group. IATA.
2. de Pater, I., Reijns, A., & Mitici, M. (2022). Alarm-based predictive maintenance scheduling for aircraft engines with imperfect Remaining Useful Life prognostics. Reliability Engineering & System Safety, 221, 108341.
3. Wang, L., Zhu, Z., & Zhao, X. (2024). Dynamic predictive maintenance strategy for system remaining useful life prediction via deep learning ensemble method. Reliability Engineering & System Safety, 245, 110012.
4. Bieber, M., Verhagen, W. J., Cosson, F., & Santos, B. F. (2023). Generic Diagnostic Framework for Anomaly Detection—Application in Satellite and Spacecraft Systems. Aerospace, 10(8), 673.
|
|