10:00
S3: Physics & Data-driven PHM at Amphi Fournel (AF)
Chair: Manuel Arias Chao and Catarina Silva
10:00
20 mins
|
Concept of a multifunctional UAV powertrain test bench to develop methods for data-based condition monitoring
Florian Coors, Immo Schmidt
Abstract: For the successful commercial deployment of unmanned aerial vehicles (UAVs), it is essential to conduct beyond visual line of sight (BVLOS) flights. To achieve this, a specific operations risk assessment (SORA) must be carried out to provide information about the mission risk [1]. Reliable detection of faults in the electrical powertrain is crucial to reduce potential risks for people and property on the ground.
There are various methods to detect faults and damage. According to [2], there are three main categories of methods: knowledge-based, process model-based, and data and signal models. Knowledge-based methods involve expert systems that use rule-based approaches to identify faults. However, these methods have limitations, such as difficulty adapting to new situations, inability to detect time-dependent phenomena, and inability to learn from their own mistakes. Simultaneously, the development and maintenance of such systems is considered to be costly [2]. Nevertheless, as computers continue to improve in the area of machine learning, data-driven methods for fault detection are becoming more prevalent. To make use of these methods, several parameters of the UAV are recorded over time, and a diagnosis is carried out using e.g. machine learning methods. Such approaches are appropriate for analyzing and recognizing complex correlations among a large amount of data, and allow the consideration of numerous parameters in the diagnosis of a UAV [2].
The development of data-based fault detection algorithms requires a large and reliable database. This should include data from various operating conditions and fault types. Collecting fault detection data from real UAV flights is difficult due to the possible consequences of introduced faults. However, some public datasets for UAV fault detection are available. In [3] and [4], flight data were collected for multirotor UAVs with propeller faults, while in [5], flight tests were conducted for fixed-wing UAVs with different types of control surface faults.
To obtain a larger database comprising multiple fault mechanisms of UAV powertrains, it is necessary to develop a powertrain test rig. Part of the test rig is an electric powertrain unit consisting of a BLDC motor, which is controlled by an ESC, and a propeller. A modular design shall enable the testing of motors and propellers of different sizes on the test bench. While the majority of the powertrains tested will be of smaller size, the design and measurement ranges of the sensors are selected so that rotors with up to 12.5 kgf thrust can be tested on the test bench. This means that, in theory, all quadcopter configurations within the “open” category of the EU regulations for UAV [1] with a maximum weight of 25 kg and a thrust-to-weight ratio of 2:1 can be tested.
Two wind generators are installed to simulate a variety of operating conditions that are as realistic as possible for a real-world application. The first wind generator can blow air into the test chamber from above, thus simulating the vertical movement of the UAV during a climb. To simulate forward flight, a second wind generator is used to generate a lateral wind. The propeller must be tiltable to realize different angles of attack to the incoming airflow.
By simulating the airflow, different sections of a real flight can be simulated and the effects on the measurement data and thus the fault detection can be assessed. This also makes it possible to simulate real flight missions.
In order to develop a reliable diagnostic algorithm, it is necessary to have a wide range of defects throughout the entire powertrain in the database. The powertrain components that are prone to faults include the BLDC motor, all bearings along the shaft, and the propeller itself. When testing on the test bench, faults in the windings of the BLDC motor will be considered as they can cause malfunctions. Furthermore, several damage types to the inner and outer rings of the bearings will be considered, as well as damage to the propeller such as a deformed propeller blade or a cut-off propeller tip.
To detect faults, operational parameters such as vibration data, thrust and torque data, rotational speed data, rotor position and wind speed, as well as motor current and voltage data will be recorded. To develop a diagnostic algorithm for real UAVs, it is essential to use only data that is already measured in common UAVs or is measurable with little financial and technical effort.
In conclusion, the proposed test rig concept is a key enabler to create a database with a vast variety of operating conditions, hardware configurations and possible insertion of damages while being closer to the real-world application compared to static test rigs. Additionally, this test rig is planned to be embedded in a real-time manual flight simulation to test and validate the developed algorithms.
References
[1]
European Union Aviation Safety Agency, „EASA,“ 09 2022. [Online]. Available: https://www.easa.europa.eu/en/document-library/easy-access-rules/online-publications/easy-access-rules-unmanned-aircraft-systems. [Zugriff am 27 03 2024].
[2]
D. Miljkovi, „Fault Detection Methods: A Literature Survey,“ 2011.
[3]
R. Puchalski und W. Giernacki, „UAV Fault Detection Methods, State-of-the-Art,“ Bd. 6, Nr. 330, 2022.
[4]
A. Baldini, L. D’Alleva, R. Felicetti, F. Ferracuti, A. Freddi und A. Monteriù, „UAV-FD: a dataset for actuator fault detection in multirotor drones,“ 2023.
[5]
A. Keipour, M. Mousaei und S. Scherer, „ALFA: A Dataset for UAV Fault and Anomaly Detection,“ Bd. 40, Nr. 2-3, 2021.
|
10:20
20 mins
|
Health Monitoring for a Helicopter Hydraulic Pump
Pieter-Jan Dewitte, Pauline Vos, Ivo Diepeveen, Gerben van de Vrie, Wim Lammen
Abstract: The pump of the hydraulic system provides hydraulic power to the flight control system. Degradation and failure of the pump affects the functioning of the hydraulic system and therefore flight safety. The hydraulic pump examined is an axial piston machine with an internal swash plate for pressure/flow control. The pump does not meet the prescribed reliability and is maintained “on-condition”. The poor performance in combination with sufficient in-service flight data make it a good candidate for developing a PHM approach.
Historical failures formed the starting point for the development of diagnostics for the hydraulic pump. Time series data of run-to-failure serial numbers have been analysed (catastrophically failed pumps). Physical models were developed and tuned based on in-flight data and test bench measurements. The combination of simulation, test bench and in-service flight data has contributed to a greater understanding of common degradation mechanisms and is a step in the development of fault identification algorithms. The hybrid approach (combining a physical model plus data-based fault identification) is a promising way forward. The developed algorithms, methods and tools provide valuable decision information for an operator and contributes to improved flight safety.
Figure 1: Example of pump behaviour during pre-flight sessions, before (red) and after (green) pump failure and replacement. In the run-up to a catastrophic failure, the hydraulic pressure is high and fluctuates frequently.
Flow
|
10:40
20 mins
|
Statistical Analysis and Simulation-Based Prognostic Approach of Aircraft Fuel System Failure
Shuai Fu, Nicolas P. Avdelidis
Abstract: As a result of the rapidly decreasing product lifecycles and the increasing complexity of customer requirements, the aviation sector is currently confronted with considerable hurdles in terms of maintenance procedures. An increased interest in prognostics, particularly in predicting the remaining usable life (RUL) of aviation components, has been ignited as a result of the fact that conventional maintenance methods are becoming unsustainable. The purpose of this work is to propose a robust prognostic methodology for predictive maintenance. This methodology makes use of the physics of failure (PoF) and statistical methodologies in order to enhance the dependability and safety of aircraft operations [1].
The solution that has been suggested makes use of both sensor data and algorithms in order to monitor and anticipate the system's potential for deterioration. Through the utilisation of the PoF methodology, it performs an exhaustive analysis of the failure mechanisms that are present in aircraft fuel systems. A precise and anticipatory maintenance schedule can be generated through the utilisation of the approach, which makes use of datasets derived from laboratory tests and simulation models. It is necessary to have a full grasp of the physical features of the materials, the stresses that are experienced during operation, and the environmental conditions that have an effect on the components in order to do this [2].
The research was conducted at the Integrated Vehicle Health Management (IVHM) Centre at Cranfield University, where an aircraft fuel system test rig was utilised for the purpose of conducting the research. The test rig is designed to simulate a variety of failure modes, including pump leakage from both the outside and the inside, blockage and leakage from the fuel oil heat exchanger (FOHE), and obstruction from the nozzle. For the purpose of data collection, the LabVIEW programme was utilised. The types of parameters that were collected included pressure, flow rate, and pump speed.
Simulation models that accurately recreate the structure and processes of an aviation fuel system were incorporated into the experimental setups, which in turn led to improvements. For each of the several operational situations, the simulation takes into account a number of different elements, including drops in fuel supply pressure, bank angles, and pump flow rates. When it comes to designing a strategy for predictive maintenance, having a solid understanding of the various failure mechanisms and degradation patterns is absolutely necessary.
The prognostic model takes into consideration three different aspects: the starting degradation, the normal degradation, and the atypical degradation elements. For the purpose of predicting the health index of fuel system components throughout the course of time, the model makes use of statistical methods such as Bootstrap Forest, Boosted Tree, and Neural Boosted algorithms. There was a confirmation that the model was accurate. The findings were encouraging because they showed a decrease in the amount of time that the system was offline and an improvement in the optimisation of the maintenance schedules.
Significant improvements in the precision of predictive maintenance were observed as a result of the use of the PoF-based prognostic technique. With an R-square value of 0.9995 and a Root Average Square Error (RASE) of 5.79, the Boosted Tree technique displayed remarkable performance, showing a high level of accuracy in forecasting system failures. This indicated that the method was able to accurately anticipate system failures. The findings shed light on the potential for the use of proof-of-fee in conjunction with data analytics to bring about a change in the maintenance procedures utilised in the aviation industry.
Incorporating advanced sensor technology, data analytics, and proof-of-failure procedures into aircraft maintenance presents a potentially game-changing opportunity for the aviation industry. Not only does this technology enhance the reliability and security of aeroplanes, but it also reduces the costs of operations and the amount of time that aircraft are out of service. In the subsequent investigations, the refinement of these models and their integration with live monitoring systems will be given priority in order to improve the efficiency of predictive maintenance practices.
Keywords: predictive maintenance, prognostic, aeronautic failure, remaining useful life, physics of failure
References
1. Fu S., Avdelidis NP. Aeronautics failure: a prognostic methodology based on the physics of failure and statistical approaches for predictive maintenance. Proc.SPIE. 2024. p. 1295205. Available at: DOI:10.1117/12.3018035
2. Gharib H., Kovács G. A Review of Prognostic and Health Management (PHM) Methods and Limitations for Marine Diesel Engines: New Research Directions. Machines. 2023; 11(7). Available at: DOI:10.3390/machines11070695
|
11:00
20 mins
|
Estimating the System Reliability of Unmanned Aircraft Vehicles with a State Based Approach
Max Weigert
Abstract: For many technical systems, monitoring their reliabilty requires the consideration of several components and their interaction. Especially for safety critical and dynamic systems like unmanned aerial vehicles (UAVs), the correct assessment of the expected reliability over the course of a mission is an important prerequisite for their autonomous operation. Such an assessment determines the risk of injuries caused by a crash, as demonstrated by (Breunig und Sayed 2018), where it is determined in respect of the population density and possible crash trajectory. As a result, this risk must be factored into decision-making processes during contingency management in autonomous operations, as discussed by (Eduardo und Schmidt 2021; Baculi und Ippolito 2019; Airbus 2018). The literature discusses various approaches to incorporating real-time data on component states into mission management. These include a Bayesian network approach in (Ancel et al. 2017), model-based fault detection in (Schumann et al. 2019), analysis of the influence of safety-critical components on system behavior in (Darrah et al. 2020), and a performance indicator for the overall UAV system in (Rodrigues 2018).
In this work, a flight simulation model is utilized to analyze the influence of a performance reduction of several UAV actuators due to degradation on the flight stability. The model represents a hybrid UAV configuration capable of both fixed-wing flight and multicopter fligh. A comprehensive description of the UAV is provided in (Prochazka et al. 2019) and its aerodynamic properties are discussed in (Prochazka et al. 2020). The flight model in this works is based on the formulation of UAV flight equations developed during the creation of combined physical-based and data-driven models, which are described in (Enkelmann et al. 2022). With the flight mechanics and actuation force model, the physical response of the UAV to the flight controller, a locally executed ArduPilot instance as described by (ArduPilot Dev Team), is represented.
Degradation of subsystems is simulated for the Brushless Direct Current (BLDC) motors, which propel four airscrews in upward direction along with one pusher airsscrew for forward flight, the servo motors actuating the aileron, rudder and elevator surfaces and the battery providing electric energy. For the BLDC motors a decrease in efficiency during their actuation time is assumed as in (Darrah et al. 2020). The servo motors are affected by a random deviation between control input and resulting control surface angle, with an increasing variance with each actuation as observed in (ElSaid et al. 2019). The battery is assumed to decrease in efficiency with every load cycle due to the degradation mechanisms described in (Birkl et al. 2017). Multiple UAV time series are simulated to create a database, with varying degradation rates for the individual components. For each UAV, a series of missions is simulated until it experiences a loss of control during the flight. The simulated mission profiles are randomly generated and represent back and forth flights between a fixed home waypoint and random target waypoints with intermediate waypoints as a result of a path planning to avoid restricted airspaces. For each mission, a random wind profile is generated as well. During the missions, a distinction between the flight modes of multicopter flight during takeoff and landing, fixed wing flight on the way to the target and two transition phases from multicopter to fixed wing flight and vice versa can be made.
The developed risk estimation method incorporates preprocessing of the current flight and subsystem health data, classification of the currently observed state and an estimation of the true system state with a hidden semi-Markov Model. A similar approach to apply a Markov state estimation for the system reliability determination is described in (Anger 2018). The applied preprocessing aims to enable clustering of similar flight segments and reduce redundant information. To do so, the flight data is divided into one second flight segments, with a sampling rate of 100 samples per second. Only segments with a constant flight mode are considered and time segments entered less then one second before a changing flight mode are discarded. The sample values in each segment, including actuation inputs, resulting movements, wind speed, and mission parameters such as current flight mode, distance to the mission target, and remaining turn maneuvers, are reduced to their mean and standard deviation. Similar flight segments are clustered using both an unlabeled approach via a K-Means algorithm and a rule-based approach using a Decision Tree. The Decision Tree is trained to differentiate labeled classes based on parameters such as the remaining time until mission failure, as well as the wind and actuator health conditions. Multiple hidden semi-Markov Models are trained, each using one observation sequence up to the point of UAV mission failure. These models facilitate the identification of similar UAV degradation behavior in unknown test data compared to the trained Markov Models. Advantages of the hidden semi-Markov Models lie within their ability to account for degradation within each state of the model in terms of a time-dependent residence probability and the consideration of uncertainty between observed behavior and the true system state, facilitated by the observation probability. A further introduction to hidden Markov Models is given in (Rabiner 1989).
With the developed model, the current states of unseen UAVs are evaluated, demonstrating its ability to quantify the risk of a mission failure not only in the current mission but also in those occurring in the near future. A further analysis dives into the models ability to handle uncertain observations with a particular focus on the health estimation of the UAV subsytems. The robustness of the approach is examined in light of noisy health data, enabling a discussion of the required quality of the subsystem health estimation for an accurate system reliability assessment. With the developed approach dependencies within a technical system can be recognized and the probablity of meeting system performance requirements in the future can be linked to current operational data.
|
11:20
20 mins
|
Optimal sensor placement for structural health monitoring using machine learning algorithms
Minyoung Yun, Mikhael Tannous, Chady Ghnatios, Francisco Chinesta
Abstract: The first step in structural health monitoring, creating a hybrid modelling of a structure, or building a smart structure, is the instrumentation of the part in consideration. Optimal sensor placement is currently an open issue of high relevance in the scientific community. Multiple works have tackled optimal sensor placement, using different techniques, to minimise the prediction error, improve defect detection as well as the predictive maintenance operations. In this framework, our work proposes a novel approach for optimal sensor placement, based on advanced lightly intrusive Model Order Reduction techniques (MOR), combined with machine learning algorithms.
In this work, we propose, first, to build a generic reduced order model based on lightly intrusive modelling techniques, for the structure. The proposed lightly intrusive model order reduction method combines the accuracy of the intrusive methods, along with the ease of implementation and use of the non-intrusive ones. The proposed lightly intrusive model order reduction is later leveraged to identify the optimal sensor placement on the structure at hand.
In a first step, the proposed method builds a reduced order basis of the solution, like classical MOR methods, but also reduced the problem operator. The operator can be a linear parametric operator, a non-linear operator, or a non-linear parametric one. In any case, the algorithm builds a reduced order basis of the physical operator, representing the underlying physics represented by this operator. The basis is used to reconstruct the operators using a machine learning algorithm, which can identify in real-time the best weights to attribute to each of the vectors of the basis. This machine learning algorithm is trained on a subset of the available dataset and tested on the remaining ones. In this work, we tested both random forest and deep neural network techniques, which have led to similar results when trained, for the selected use case.
Once the reduced order model is available, and the machine learning algorithm is trained, Random Permutation Features Importance Method (RPFIM) is used to select the most relevant mesh nodes, used by the machine learning algorithm, as optimal sensor placement zones. Once the sensors are placed, the selected sensors are used to identify multiple defects, one at a time, varying in size and location all over the part. The defect detection is performed by solving inverse identification problems, using the genetic algorithm. The inverse identification runs within less than five minutes, thanks to the use of the lightly intrusive reduced order model, where each evaluation of the solution is performed within a fraction of a second.
The defect detection algorithm is used over several defects, without changing the sensor placement, to evaluate the performance of the sensor at that location. Therefore, a criterion to assess the sensor placement quality is proposed. The selected points using RPFIM are compared to state-of-the-art techniques, based on Discrete empirical interpolation method (DEIM), a popular technique for nonlinear model order reduction. This comparison helped us deduced the strong and weak points of both methods.
Later, the aforementioned sensor placement techniques are combined together, to select the optimal sensor placement, from both methods. The combination uses the best and most highly performant sensors from both methods. The proposed sensor combination is again validated against classical sensor placement methods and its superiority is proven.
With the recurrent requirement of the minimisation of the number of sensors, the optimal sensor placement remains an open question. The proposed method can, in real-time, reconstruct the deformation solution, identify the possible presence of a defect in the part, and locate the defect and its size, automatically and without experts’ intervention.
|
11:40
20 mins
|
On the use of Topological Data Analysis for structural and process health monitoring
Arthur Lejeune, Nicolas Hascoët, Marc Rebillat, Eric Monteiro, Nazih Mechbal
Abstract: The Topological Data analysis (TDA) has proved its relevance to extract information from any dimension data in several application domains [1]–[3]. This approach is based on the extraction of features representing the topology of multi-dimensional data. These features are homology classes which enable to compare and regroup data by topological nature. TDA is first design for high dimension data description but several applications for 1D-timeseries show interesting results. Several examples of the use of TDA for Structural Health Monitoring and Process Health Monitoring will be presented to highlight the TDA capabilities for damage detection and classification and for monitoring in general.
[1] M. Gidea, « Topological Data Analysis of Critical Transitions in Financial Networks | SpringerLink », in Springer Proceedings in Complexity, Tel-Aviv, Israel, 2017, p. 47‑59.
[2] Y. Skaf et R. Laubenbacher, « Topological data analysis in biomedicine: A review », J. Biomed. Inform., vol. 130, p. 104082, juin 2022, doi: 10.1016/j.jbi.2022.104082.
[3] R. Iniesta, E. Carr, M. Carrière, N. Yerolemou, B. Michel, et F. Chazal, « Topological Data Analysis and its usefulness for precision medicine studies », SORT-Stat. Oper. Res. Trans., p. 115‑136, juin 2022, doi: 10.2436/20.8080.02.120.
|
|