ESRs & Workpackages
ESR1 will work on the extraction and exploitation of the hidden information existing in heterogeneous data. Different types of signals, including vibrations, acoustics, wear debris, oil quality and operating conditions will be fused/integrated in order to reveal the relationship between the measured signals and the true degradation of the system, focusing on smart, self-monitored mechatronic systems operating under time-varying conditions. Knowledge and experience from signal processing, tribology, machine dynamics and design, vibrations, acoustics and oil analysis will be combined. ESR1 will work on the underlying physics of the degradation and how this information is present in the captured signals. Wear modelling and tribology tests will be used here. A multi-sensor approach will be employed to capture and analyse the heterogeneous data. Advanced signal-processing tools based on cyclo-stationarity will be developed and applied to the signals, extracting important monitoring indicators. The correlation and quantification of the system degradation with the monitoring indicators will lead to a novel methodology for the prediction of the remaining useful life of mechatronic components. The methodology will be tested and validated on dedicated test rigs, specially designed and constructed at KU Leuven as well as on real industrial signals.
Innovative aspects: Exploitation of heterogeneous signals, Correlation of wear and monitoring features
ESR2 will work on machine-learning and artificial-intelligence methods for self-learning and self-monitoring cyber-physical systems. A framework based on novelty detection will be established using dynamically updated clustering methodologies, capturing the transition of the system between multiple steady and time-varying operating modes. Moreover, self-organizing model methods will be investigated and further developed. ESR2 will establish a procedure in order to perform comparisons of identical or possibly similar cyber-physical systems (drivetrains, vehicles, machines) using similarity measures, in order to identify and monitor abnormal phenomena, such as wear, failures and noise. The developed methodologies and algorithms will first be tested and evaluated on simulated data and on two dedicated similar but not identical laboratory drivetrains, in order to quantify the rate of false alarms and missed detections. The final methodology evaluation will be performed on real industrial cases.
Innovative aspects: Unsupervised self-learning method based on streaming fleet data
ESR3 will develop model-based techniques for monitoring individual units in a fleet from a probabilistic approach. The principle is based on the comparison of a parametric model identified from measured data with a generic model provided by dynamic modelling, by taking into account the acceptable differences between the generic model of the fleet and individual units. This will be solved as an inverse problem in the presence of “modelling errors” that reflect the departure of individual units from the nominal model representing the fleet. After identification of the model parameters from experimental data – a non-trivial issue due to the fact that modelling errors are highly structured and do not comply with the simplifying assumption commonly advocated in system-identification theory – each unit will be assigned a different realization of the same model, seen as a random object. This will be formalized in the probabilistic space spanned by the parameter values, thus leading to the concept of a population of models embedded in a probabilistic manifold endowed with a topological information geometry. This formalism will be useful to view and manipulate models as probabilistic objects.
Innovative aspects: Probabilistic fleet monitoring, combination of dynamic and probabilistic models, information geometry on a manifold of models.
ESR4 will work on fleet monitoring, using a generic model and solving an inverse identification problem. The choice and the use of different models that map the physical quantities of interest to the measured signals will be investigated in order to improve the general performances of a “virtual sensor”, which is achieved through inverse modelling. The richness of the sensor equipment of a unit (machine, vehicle etc.) will be used in order to correlate the virtual signal estimations from different inverse model architectures. Methodologies will be proposed for the identification of the particular operating conditions of each unit, updating simultaneously their estimation by sharing reconstructed signals of the fleet. The methodologies will be applied to fleets of vehicles and wind turbines.
Innovative aspects: Novel approach in inverse modelling and model selection for fleet monitoring.
ESR5 will focus on probabilistic methods that can identify and, ideally, forecast the deterioration of patients’ health in hospital settings. Over 25,000 avoidable deaths occur each year in UK hospitals alone, which is similar to the number of deaths each year from road-traffic accidents, and which is a common problem across healthcare systems throughout Europe. Improving the ability to monitor and forecast the health of patients promises to reduce this risk of avoidable mortality. ESR5 will develop methods based on recent advances in deep learning to fuse the time-series data of patients’ vital signs along with data from the electronic health records to provide probabilistic assessments of patient health. This research will build on proof-of-principle studies in which we have developed methods based on Gaussian processes (GPs), which are a Bayesian nonparametric means of performing inference by defining probability distributions over (potentially infinitedimensional) objects such as time-series functions.
Innovative aspects: Deep learning, information fusion, probabilistic assessments.
ESR6 will develop a novel hybrid prognostic methodology focusing on the assessment of the state of the health of a system. The methodology will integrate physics-based and data-driven prognostics models in order to enhance the prognostic accuracy, robustness and applicability. Dedicated experimental test rigs such as a clogged filter, a machinery fault simulator and a linear actuator failure simulator will be used to obtain reproducible datasets under different operating conditions. Finally, the performance of the developed technique will be evaluated based on the most recent prognostic evaluation metrics.
Innovative aspects: Development of hybrid prognostic technologies.
ESR7 will develop a methodology describing how autonomous machines can learn from human experts in a maintenance context. The aim of this work is to critically assess the maintenance actions for automated production lines in order to identify factors that contribute to performance and technical efficiency. ESR7 will focus on the identification of the elements that are necessary for machines to conduct maintenance actions. More specifically, implicit situation awareness measurement techniques and state-of-the-art techniques will be used in the exploration of cognitive tasks as a platform to develop extended methods and to explore intuitive cognition. An empirical study to explore the intuitive cognition phenomena is necessary; to explore and answer the what and why questions. Such studies can also contribute to refining the theoretical foundation as well as to improving the human cognition taxonomy at different levels of maintenance automation. This is a pertinent issue, to obtain the acquisition and transfer of tacit maintenance know how, from legacy systems to intelligent systems. There is an untapped potential of simulation for digital training using virtual reality solutions, not only for humans to gain skills but also for intelligent systems to learn from intuitive human interactions. Therefore, ESR7 will explore the opportunities of integrating these methods as a strategy to digitalize human intuitive cognition.
Innovative aspects: Strategy to digitalize human intuitive cognition including psychological, philosophical, and engineering stances.
ESR8 will work on the condition monitoring of motioncontrol applications, such as independent cart conveyors. These conveyor systems are an emerging technology in industries seeking to replace induction motors and kinematic chains in several applications. The condition monitoring of these systems is a challenging industrial activity that will be developed in three steps. Firstly, a data-driven approach will be considered to perform anomaly detection on the main parts of motion-control applications and ESR8 will work on clustering algorithms, data-reduction and classification. Afterwards, a model-based approach using a few subsets of critical components will be developed, ensuring accuracy in the diagnosis and optimization of computational resources. ESR8 will focus on the modelling of physical systems and advanced signal processing. In the third step the two previous approaches will be fused, developing a hybrid condition-monitoring system that automatically combines the information from the data-driven and model-based algorithms, giving a different level of details on the root cause of the fault and making a fast and reliable monitoring of the motion-control applications. The developed methodologies will be evaluated over simulations and real data captured initially on experimental facilities. The methodology will be further applied and evaluated on a real industrial case during a secondment.
Innovative aspects: The development of a hybrid condition monitoring system combining model-based and data-driven methodologies.
ESR9 will focus on the analysis of long-term, heterogeneous, historical big-data from time-varying systems operating under impulsive, non-Gaussian noise. Mining machines are very specific: unique, complex and designed for heavy duty, they operate under non-stationary conditions, they can experience multiple faults and they are frequently overloaded (especially during start-up or emergency stops). The condition monitoring of these machines is complicated. Often, diagnostics measurements are affected by the environmental conditions (temperature) and impulsive disturbances (impulsive noise). The monitoring of such machines requires the processing of a huge number of different variables and the nearly online validation and processing of data. The ESR will develop methodologies in order to extract and exploit the knowledge about the evolution of the monitoring process, exploiting information from different types of sensors and signals.
Innovative aspects: Development of novel modelling/analysis/separation approaches for long-term, big Data, heterogeneous processes for condition monitoring with a focus on non-stationary operation and heavy-tailed, non-Gaussian noise.
ESR10 will develop a methodology, based on the Failure-Modes-and-Effects-Analysis (FMEA) and multi-class classification methods, to monitor large-scale infrastructures, such as bridges and wind turbines, as well as the sensor network that is used. In special use-cases in rough environments, e.g., if the submerged structures of offshore wind turbines should be monitored, it is commonly accepted that the sensors will only have a useful lifetime of less than one year (so called ‘lost sensors’). A novel method will be developed to replace the information coming from the lost sensors by information from the remaining sensors in the network, replacing the “lost sensors” by “virtual” ones. The methodologies will be validated using real data from a full-scale wind turbine (2 MW, 100 m hub-height) equipped with a network of smart sensors.
Innovative aspects: Simultaneous monitoring of infrastructure and sensors network degradation.
ESR11 will define a methodology to evaluate the remaining fatigue life of large uncertain industrial systems. The methodology will focus on obtaining reliable stress time series at fatigue-critical locations of industrial systems under operational conditions. The estimated stresses will then be used with advanced fatigue-and-damage tolerance techniques in order to assess the remaining useful life of the system. In this way the evaluation of the current fatigue life will shift from a “worst-case scenario” to a more realistic “current-case scenario”. Therefore, with this methodology, the evaluation of the remaining useful life of large industrial machinery will be improved, allowing a shift from predictive to condition-based maintenance. ESR11 will study the cases in which the installation of a dedicated physical sensor is not feasible. To address this problem, model-based monitoring approaches will be used. These approaches aim at combining measurements from multiple sensors with a model of the system, thus combining the best of both worlds: (1) current information on the condition of the system, and (2) a physical insight into the system’s behaviour. Therefore, these techniques will allow studying the behaviour of the system under operational conditions. Furthermore, given the complexity of modern industrial systems, special attention will be given to coping with the high modelling uncertainty.
Innovative aspects: Development of reliable stress estimation algorithms, application to fatigue life prediction, validation in test rigs of industrially relevant systems.
ESR12 will contact research on two transfer learning strategies. In the first strategy, a physical simulation model will provide a source dataset which can be used for training of an initial machine learning model. Such an approach was already successful in previous work on bearing fault detection, where simulation models were used for training a support vector machine or a deep neural network. However, it is expected that other cases will require more advanced transfer learning methods (in particular, domain adaptation), for example if the simulations cannot capture all features which will be present in real-life data and which are relevant for the specific industrial context. The second strategy assumes that the machine under-study is part of a larger fleet of similar (but not necessarily identical) machines. For example, if older machines have been in the field already for some years, the data and knowledge gained on these machines might be transferrable to predict quantities on a newly deployed machine.
Innovative aspects: Transfer learning applied to simulation data and fleet data for time series prediction.
ESR13 will investigate machine-learning methods that are trained to recognize incorrect sensor data. A systematic approach will be followed: in the first stage, a supervised learning approach will be adopted, whereby it is assumed that an historical dataset with fully labelled examples is available. As this assumption might not prove to be practically realizable in many cases, an unsupervised anomaly-detection approach will be investigated in the second stage. Such an approach does not require labelled data, but is typically more difficult to implement successfully compared to a supervised approach. An interesting third alternative that will be investigated is a semi-supervised approach, where a small labelled dataset (e.g., obtained from expert user feedback is available in addition to the larger unlabelled dataset. Besides the systematic investigation outlined above (supervised–unsupervised–semi-supervised), a particular focus point will be to leverage the fact that there are multiple sensors, i.e., there is a certain redundancy in the measurement setup so that some sensors will be measuring related quantities. While measurement anomalies are non-physical events that occur at random times (so that they will likely not be observed in multiple sensor channels), real physical events likely affect multiple (closely located) sensors. A comparison between sensor pairs (e.g., linear or nonlinear correlation analysis) could thus be exploited in order to better detect the measurement anomalies (for example, to distinguish an incorrect measurement spike from a true physical shock event in the data).
Innovative aspects: Machine learning methods for automatic validation of multiple sensors, whereby in particular semi-supervised learning and multi-sensor strategies beyond the current state-of-the-art.
ESR14 will develop methodologies to fuse and exploit information extracted from heterogeneous data, including vibration, temperature, images, sound and local environmental parameters, captured by inspection robots, for monitoring and diagnosing machines and infrastructure (electric cables, pipelines, structures) used in deep mines. Due to the harsh environment and natural hazards (gas and rock outbursts, seismic risks), the future of mining is related to introducing fully automated machinery and robotic inspection systems in order to minimize the risks for humans. The development of mining-inspection robots equipped with various types of sensors, including microphones, accelerometers, temperature/humidity/gas sensors and cameras, will provide a massive and heterogeneous amount of data gathered under harsh mining conditions.
Innovative aspects : Development of data-fusion techniques for the exploitation of heterogeneous data captured by robotic devices for monitoring of mines.
ESR15 will work on the analysis, identification and modelling of vibratory and acoustic sources of aircraft engines. Vibratory and acoustic sources include random phenomena related to fuel combustion, cyclic phenomena related to the rotations of shafts/rotors and nonlinear phenomena related to the fluidic link between the high-pressure and the low-pressure turbines. The digital industrial evolution is characterized by an interconnection of machines and systems (objects) that is mainly based on the use of many sensors that collect information, but with a degraded quality related to the nature of the “general public” sensors they are equipped with. ESR15 will study how the health condition of aircraft engines can be analysed in real time using this type of generally low-quality sensor and establish a diagnosis that will anticipate and optimize the maintenance. The work will rely on baseline data collected on board during a flight from the specialized monitoring equipment and on data captured by common smartphones/tablets during actual flights. The captured data will be firstly structured and ESR15 will work towards the correlation of the different sources and the development and application of novel high-level signal processing tools in order to extract the diagnostic information hidden in the sound-and-vibration signals collected by common smartphones/tablets (low-quality sensors). The statistical properties of the captured signals will be exploited to characterize the sources and provide indicators of the motor’s health. Moreover, the ESR will develop a model for predictive maintenance exploiting additional information collected on the ground and during the planned maintenance.
Innovative aspects : Condition monitoring of aircraft engines using smartphones/tablets